deepscholar-bench/DeepScholarBench · Datasets at Fast360

parent_paper_title stringclasses 63 values	parent_paper_arxiv_id stringclasses 63 values	citation_shorthand stringlengths 2 56	raw_citation_text stringlengths 9 63	cited_paper_title stringlengths 5 161	cited_paper_arxiv_link stringlengths 32 37 ⌀	cited_paper_abstract stringlengths 406 1.92k ⌀	has_metadata bool 1 class	is_arxiv_paper bool 2 classes	bib_paper_authors stringlengths 2 2.44k ⌀	bib_paper_year float64 1.97k 2.03k ⌀	bib_paper_month stringclasses 16 values	bib_paper_url stringlengths 20 116 ⌀	bib_paper_doi stringclasses 269 values	bib_paper_journal stringlengths 3 148 ⌀	original_title stringlengths 5 161	search_res_title stringlengths 4 122	search_res_url stringlengths 22 267	search_res_content stringlengths 19 1.92k
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	NBERw21340	\cite{NBERw21340}	Effective Policy for Reducing Inequality? The Earned Income Tax Credit and the Distribution of Income	null	null	true	false	Hoynes, Hilary W and Patel, Ankur J	2,015	July	http://www.nber.org/papers/w21340	10.3386/w21340	null	Effective Policy for Reducing Inequality? The Earned Income Tax Credit and the Distribution of Income	Effective Policy for Reducing Inequality? The Earned Income	https://ideas.repec.org/p/nbr/nberwo/21340.html	Our results show that a policy-induced $1000 increase in the EITC leads to a 7.3 percentage point increase in employment and a 9.4 percentage point reduction
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	NBERw21211	\cite{NBERw21211}	The Earned Income Tax Credit (EITC)	null	null	true	false	Nichols, Austin and Rothstein, Jesse	2,015	May	http://www.nber.org/papers/w21211	10.3386/w21211	null	The Earned Income Tax Credit (EITC)	What is the earned income tax credit? - Tax Policy Center	https://taxpolicycenter.org/briefing-book/what-earned-income-tax-credit	The earned income tax credit (EITC) provides substantial support to low- and moderate-income working parents who claim a qualifying child.
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	Foo2019ProcessAC	\cite{Foo2019ProcessAC}	Process and Critical Approaches to Solving the Systemic Climate Change Governance Problem	null	null	true	false	Check Woo Foo	2,019	null	https://api.semanticscholar.org/CorpusID:235319207	null	Politics \& Energy eJournal	Process and Critical Approaches to Solving the Systemic Climate Change Governance Problem	Process and Critical Approaches to Solving the Systemic Climate ...	https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3608501	The most important and urgent task, besides avoiding nuclear war, is abatement of the existential threat of systemic climate change,
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	Patjoshi2015DesignAD	\cite{Patjoshi2015DesignAD}	Design and Development of Advanced Control strategies for Power Quality Enhancement at Distribution Level	null	null	true	false	Rajesh Kumar Patjoshi	2,015	null	https://api.semanticscholar.org/CorpusID:112918597	null	null	Design and Development of Advanced Control strategies for Power Quality Enhancement at Distribution Level	(PDF) Advanced Control Strategies for UPQC to Improve ...	https://www.researchgate.net/publication/279289697_Advanced_Control_Strategies_for_UPQC_to_Improve_Power_Quality_of_Power_Distribution_Systems	PDF \| On Jul 2, 2014, Quoc Nam Trinh published Advanced Control Strategies for UPQC to Improve Power Quality of Power Distribution Systems
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	10.1257/jep.25.4.165	\cite{10.1257/jep.25.4.165}	The Case for a Progressive Tax: From Basic Research to Policy Recommendations	null	null	true	false	Diamond, Peter and Saez, Emmanuel	2,011	December	https://www.aeaweb.org/articles?id=10.1257/jep.25.4.165	10.1257/jep.25.4.165	Journal of Economic Perspectives	The Case for a Progressive Tax: From Basic Research to Policy Recommendations	The Case for a Progressive Tax	https://economics.mit.edu/sites/default/files/2022-09/jep.25.4.165.pdf	Therefore, optimal income tax theory is fi rst a normative theory that shows how a social welfare objective combines with constraints arising from theory that shows how a social welfare objective combines with constraints arising from limits on resources and behavioral responses to taxation in order to derive specifi c limits on resources and behavioral responses to taxation in order to derive specifi c The Case for a Progressive Tax: From Basic Research to Policy Recommendations † ■ ■ Peter Diamond is Professor Emeritus of Economics, Massachusetts Institute of Tech-nology, Cambridge Massachusetts. doi=10.1257/jep.25.4.165 Peter Diamond and Emmanuel Saez 166 Journal of Economic Perspectives tax policy recommendations. In addition, optimal income tax theory can be used to tax policy recommendations.
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	10.2307/2296779	\cite{10.2307/2296779}	An Exploration in the Theory of Optimum Income Taxation12	null	null	true	false	Mirrlees, J. A.	1,971	04	https://doi.org/10.2307/2296779	10.2307/2296779	The Review of Economic Studies	An Exploration in the Theory of Optimum Income Taxation12	Exploration in the Theory of Optimum Income Taxation12	https://academic.oup.com/restud/article-abstract/38/2/175/1527903	by JA Mirrlees · 1971 · Cited by 7415 — J. A. Mirrlees; An Exploration in the Theory of Optimum Income Taxation12, The Review of Economic Studies, Volume 38, Issue 2, 1 April 1971, Pages 175–208,
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	RePEc:aea:aecrev:v:61:y:1971:i:1:p:8-27	\cite{RePEc:aea:aecrev:v:61:y:1971:i:1:p:8-27}	Optimal Taxation and Public Production: I--Production Efficiency	null	null	true	false	Diamond, Peter and Mirrlees, James	1,971	null	https://EconPapers.repec.org/RePEc:aea:aecrev:v:61:y:1971:i:1:p:8-27	null	American Economic Review	Optimal Taxation and Public Production: I--Production Efficiency	[PDF] Optimal Taxation and Public Production I: Production Efficiency	http://hassler-j.iies.su.se/Courses/DynPubFin/Papers/DiamondMirrlees.pdf	Theories of optimal production in a planned economy have usually assumed that the tax system can allow the govern- ment to achieve any desired redistribution of
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	10.1111/1467-937X.00166	\cite{10.1111/1467-937X.00166}	Using Elasticities to Derive Optimal Income Tax Rates	null	null	true	false	Saez, Emmanuel	2,001	01	https://doi.org/10.1111/1467-937X.00166	10.1111/1467-937X.00166	The Review of Economic Studies	Using Elasticities to Derive Optimal Income Tax Rates	Using Elasticities to Derive Optimal Income Tax Rates	https://academic.oup.com/restud/article/68/1/205/1568609	by E Saez · 2001 · Cited by 1885 — This paper derives optimal income tax formulas using compensated and uncompensated elasticities of earnings with respect to tax rates.
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	10.1257/pol.6.1.230	\cite{10.1257/pol.6.1.230}	Optimal Taxation of Top Labor Incomes: A Tale of Three Elasticities	null	null	true	false	Piketty, Thomas and Saez, Emmanuel and Stantcheva, Stefanie	2,014	February	https://www.aeaweb.org/articles?id=10.1257/pol.6.1.230	10.1257/pol.6.1.230	American Economic Journal: Economic Policy	Optimal Taxation of Top Labor Incomes: A Tale of Three Elasticities	Optimal Taxation of Top Labor Incomes: A Tale of Three Elasticities	https://www.nber.org/papers/w17616	This paper presents a model of optimal labor income taxation where top incomes respond to marginal tax rates through three channels.
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	10.1257/pol.20180033	\cite{10.1257/pol.20180033}	Optimal Income Taxation with Unemployment and Wage Responses: A Sufficient Statistics Approach	null	null	true	false	Kroft, Kory and Kucko, Kavan and Lehmann, Etienne and Schmieder, Johannes	2,020	February	https://www.aeaweb.org/articles?id=10.1257/pol.20180033	10.1257/pol.20180033	American Economic Journal: Economic Policy	Optimal Income Taxation with Unemployment and Wage Responses: A Sufficient Statistics Approach	Optimal Income Taxation with Unemployment and Wage Responses	https://www.aeaweb.org/articles?id=10.1257/pol.20180033	We derive a sufficient statistics tax formula in a model that incorporates unemployment and endogenous wages to study the shape of the optimal income tax. Key
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	zheng2020aieconomistimprovingequality	\cite{zheng2020aieconomistimprovingequality}	The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies	http://arxiv.org/abs/2004.13332v1	Tackling real-world socio-economic challenges requires designing and testing economic policies. However, this is hard in practice, due to a lack of appropriate (micro-level) economic data and limited opportunity to experiment. In this work, we train social planners that discover tax policies in dynamic economies that can effectively trade-off economic equality and productivity. We propose a two-level deep reinforcement learning approach to learn dynamic tax policies, based on economic simulations in which both agents and a government learn and adapt. Our data-driven approach does not make use of economic modeling assumptions, and learns from observational data alone. We make four main contributions. First, we present an economic simulation environment that features competitive pressures and market dynamics. We validate the simulation by showing that baseline tax systems perform in a way that is consistent with economic theory, including in regard to learned agent behaviors and specializations. Second, we show that AI-driven tax policies improve the trade-off between equality and productivity by 16% over baseline policies, including the prominent Saez tax framework. Third, we showcase several emergent features: AI-driven tax policies are qualitatively different from baselines, setting a higher top tax rate and higher net subsidies for low incomes. Moreover, AI-driven tax policies perform strongly in the face of emergent tax-gaming strategies learned by AI agents. Lastly, AI-driven tax policies are also effective when used in experiments with human participants. In experiments conducted on MTurk, an AI tax policy provides an equality-productivity trade-off that is similar to that provided by the Saez framework along with higher inverse-income weighted social welfare.	true	true	Stephan Zheng and Alexander Trott and Sunil Srinivasa and Nikhil Naik and Melvin Gruesbeck and David C. Parkes and Richard Socher	2,020	null	https://arxiv.org/abs/2004.13332	null	null	The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies	[PDF] Improving Equality and Productivity with AI-Driven Tax Policies - arXiv	http://arxiv.org/pdf/2004.13332	The AI Economist uses AI to discover tax policies that improve the trade-off between equality and productivity, achieving a 16% improvement
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	NBERc14009	\cite{NBERc14009}	The Impact of Machine Learning on Economics	null	null	true	false	Susan Athey	2,018	January	http://www.nber.org/chapters/c14009	null	null	The Impact of Machine Learning on Economics	The Impact of Machine Learning on Economics	https://www.gsb.stanford.edu/faculty-research/publications/impact-machine-learning-economics	# The Impact of Machine Learning on Economics This paper provides an assessment of the early contributions of machine learning to economics, as well as predictions about its future contributions. It begins by briefly overviewing some themes from the literature on machine learning, and then draws some contrasts with traditional approaches to estimating the impact of counterfactual policies in economics. Next, we review some of the initial “off-the-shelf” applications of machine learning to economics, including applications in analyzing text and images. Finally, we overview a set of broader predictions about the future impact of machine learning on economics, including its impacts on the nature of collaboration, funding, research tools, and research questions. ## Footer contact links ## Footer 1 ## Footer 2 ## Footer legal links
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	AxtellFarmer2022	\cite{AxtellFarmer2022}	Agent Based Modeling in Economics and Finance: Past, Present, and Future	null	null	true	false	Axtell, R. and Farmer, J.	2,022	null	null	null	Journal of Economic Literature	Agent Based Modeling in Economics and Finance: Past, Present, and Future	[PDF] Agent-Based Modeling in Economics and Finance: Past, Present ...	https://complexityhandbook.uni-hohenheim.de/fileadmin/einrichtungen/complexityhandbook/AXTELL_Robert.pdf	Abstract. Agent-based modeling is a novel computational methodology for representing the behavior of individuals in order to study social phenomena.
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	DelliGatti2018	\cite{DelliGatti2018}	Contents	null	null	true	false	Delli Gatti, Domenico and Fagiolo, Giorgio and Gallegati, Mauro and Richiardi, Matteo and Russo, Alberto	2,018	null	null	null	null	Contents	CONTENTS \| definition in the Cambridge English Dictionary	https://dictionary.cambridge.org/us/dictionary/english/contents	everything that is contained within something: contents of The contents of his bag spilled all over the floor. He didn't need to open the box because
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	shen2025phyxdoesmodelwits	\cite{shen2025phyxdoesmodelwits}	PhyX: Does Your Model Have the "Wits" for Physical Reasoning?	http://arxiv.org/abs/2505.15929v2	Existing benchmarks fail to capture a crucial aspect of intelligence: physical reasoning, the integrated ability to combine domain knowledge, symbolic reasoning, and understanding of real-world constraints. To address this gap, we introduce PhyX: the first large-scale benchmark designed to assess models capacity for physics-grounded reasoning in visual scenarios. PhyX includes 3K meticulously curated multimodal questions spanning 6 reasoning types across 25 sub-domains and 6 core physics domains: thermodynamics, electromagnetism, mechanics, modern physics, optics, and wave\&acoustics. In our comprehensive evaluation, even state-of-the-art models struggle significantly with physical reasoning. GPT-4o, Claude3.7-Sonnet, and GPT-o4-mini achieve only 32.5%, 42.2%, and 45.8% accuracy respectively-performance gaps exceeding 29% compared to human experts. Our analysis exposes critical limitations in current models: over-reliance on memorized disciplinary knowledge, excessive dependence on mathematical formulations, and surface-level visual pattern matching rather than genuine physical understanding. We provide in-depth analysis through fine-grained statistics, detailed case studies, and multiple evaluation paradigms to thoroughly examine physical reasoning capabilities. To ensure reproducibility, we implement a compatible evaluation protocol based on widely-used toolkits such as VLMEvalKit, enabling one-click evaluation. More details are available on our project page: https://phyx-bench.github.io/.	true	true	Hui Shen and Taiqiang Wu and Qi Han and Yunta Hsieh and Jizhou Wang and Yuyue Zhang and Yuxin Cheng and Zijian Hao and Yuansheng Ni and Xin Wang and Zhongwei Wan and Kai Zhang and Wendong Xu and Jing Xiong and Ping Luo and Wenhu Chen and Chaofan Tao and Zhuoqing Mao and Ngai Wong	2,025	null	https://arxiv.org/abs/2505.15929	null	null	PhyX: Does Your Model Have the "Wits" for Physical Reasoning?	PhyX: Does Your Model Have the "Wits" for Physical Reasoning?	http://arxiv.org/pdf/2505.15929v2	Existing benchmarks fail to capture a crucial aspect of intelligence: physical reasoning, the integrated ability to combine domain knowledge, symbolic reasoning, and understanding of real-world constraints. To address this gap, we introduce PhyX: the first large-scale benchmark designed to assess models capacity for physics-grounded reasoning in visual scenarios. PhyX includes 3K meticulously curated multimodal questions spanning 6 reasoning types across 25 sub-domains and 6 core physics domains: thermodynamics, electromagnetism, mechanics, modern physics, optics, and wave\&acoustics. In our comprehensive evaluation, even state-of-the-art models struggle significantly with physical reasoning. GPT-4o, Claude3.7-Sonnet, and GPT-o4-mini achieve only 32.5%, 42.2%, and 45.8% accuracy respectively-performance gaps exceeding 29% compared to human experts. Our analysis exposes critical limitations in current models: over-reliance on memorized disciplinary knowledge, excessive dependence on mathematical formulations, and surface-level visual pattern matching rather than genuine physical understanding. We provide in-depth analysis through fine-grained statistics, detailed case studies, and multiple evaluation paradigms to thoroughly examine physical reasoning capabilities. To ensure reproducibility, we implement a compatible evaluation protocol based on widely-used toolkits such as VLMEvalKit, enabling one-click evaluation. More details are available on our project page: https://phyx-bench.github.io/.
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	zhao2024competeaiunderstandingcompetitiondynamics	\cite{zhao2024competeaiunderstandingcompetitiondynamics}	CompeteAI: Understanding the Competition Dynamics in Large Language Model-based Agents	http://arxiv.org/abs/2310.17512v2	Large language models (LLMs) have been widely used as agents to complete different tasks, such as personal assistance or event planning. While most of the work has focused on cooperation and collaboration between agents, little work explores competition, another important mechanism that promotes the development of society and economy. In this paper, we seek to examine the competition dynamics in LLM-based agents. We first propose a general framework for studying the competition between agents. Then, we implement a practical competitive environment using GPT-4 to simulate a virtual town with two types of agents, restaurant agents and customer agents. Specifically, the restaurant agents compete with each other to attract more customers, where competition encourages them to transform, such as cultivating new operating strategies. Simulation experiments reveal several interesting findings at the micro and macro levels, which align well with existing market and sociological theories. We hope that the framework and environment can be a promising testbed to study competition that fosters understanding of society. Code is available at: https://github.com/microsoft/competeai.	true	true	Qinlin Zhao and Jindong Wang and Yixuan Zhang and Yiqiao Jin and Kaijie Zhu and Hao Chen and Xing Xie	2,024	null	https://arxiv.org/abs/2310.17512	null	null	CompeteAI: Understanding the Competition Dynamics in Large Language Model-based Agents	CompeteAI: Understanding the Competition Dynamics in Large ...	https://arxiv.org/abs/2310.17512	In this paper, we seek to examine the competition dynamics in LLM-based agents. We first propose a general framework for studying the competition between
TaxAgent: How Large Language Model Designs Fiscal Policy	2506.02838v1	nie2024surveylargelanguagemodels	\cite{nie2024surveylargelanguagemodels}	A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges	null	null	true	false	Yuqi Nie and Yaxuan Kong and Xiaowen Dong and John M. Mulvey and H. Vincent Poor and Qingsong Wen and Stefan Zohren	2,024	null	https://arxiv.org/abs/2406.11903	null	null	A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges	A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges	http://arxiv.org/pdf/2406.11903v1	Recent advances in large language models (LLMs) have unlocked novel opportunities for machine learning applications in the financial domain. These models have demonstrated remarkable capabilities in understanding context, processing vast amounts of data, and generating human-preferred contents. In this survey, we explore the application of LLMs on various financial tasks, focusing on their potential to transform traditional practices and drive innovation. We provide a discussion of the progress and advantages of LLMs in financial contexts, analyzing their advanced technologies as well as prospective capabilities in contextual understanding, transfer learning flexibility, complex emotion detection, etc. We then highlight this survey for categorizing the existing literature into key application areas, including linguistic tasks, sentiment analysis, financial time series, financial reasoning, agent-based modeling, and other applications. For each application area, we delve into specific methodologies, such as textual analysis, knowledge-based analysis, forecasting, data augmentation, planning, decision support, and simulations. Furthermore, a comprehensive collection of datasets, model assets, and useful codes associated with mainstream applications are presented as resources for the researchers and practitioners. Finally, we outline the challenges and opportunities for future research, particularly emphasizing a number of distinctive aspects in this field. We hope our work can help facilitate the adoption and further development of LLMs in the financial sector.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	vllm	\cite{vllm}	Efficient Memory Management for Large Language Model Serving with PagedAttention	http://arxiv.org/abs/2309.06180v1	High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2-4$\times$ with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm	true	true	Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph Gonzalez and Hao Zhang and Ion Stoica	2,023	null	https://doi.org/10.1145/3600006.3613165	10.1145/3600006.3613165	null	Efficient Memory Management for Large Language Model Serving with PagedAttention	Efficient Memory Management for Large Language Model ...	https://arxiv.org/pdf/2309.06180	Efficient Memory Management for Large Language Model Serving with PagedAttention Woosuk Kwon 1,∗ Zhuohan Li 1,∗ Siyuan Zhuang 1 Ying Sheng 1,2 Lianmin Zheng 1 Cody Hao Yu 3 Joseph E. Gonzalez 1 Hao Zhang 4 Ion Stoica 1 1 UC Berkeley 2Stanford University 3Independent Researcher 4UC San Diego Abstract High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. To address this problem, we propose PagedAttention, an attention al-gorithm inspired by the classical virtual memory and pag-ing techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce mem-ory usage. To address the above limitations, we propose PagedAt-tention , an attention algorithm inspired by the operating system’s (OS) solution to memory fragmentation and shar-ing: virtual memory with paging . In this work, we build vLLM , a high-throughput distributed LLM serving engine on top of PagedAttention that achieves near-zero waste in KV cache memory.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	chunkattention	\cite{chunkattention}	ChunkAttention: Efficient Self-Attention with Prefix-Aware KV Cache and Two-Phase Partition	http://arxiv.org/abs/2402.15220v4	Self-attention is an essential component of large language models (LLM) but a significant source of inference latency for long sequences. In multi-tenant LLM serving scenarios, the compute and memory operation cost of self-attention can be optimized by using the probability that multiple LLM requests have shared system prompts in prefixes. In this paper, we introduce ChunkAttention, a prefix-aware self-attention module that can detect matching prompt prefixes across multiple requests and share their key/value tensors in memory at runtime to improve the memory utilization of KV cache. This is achieved by breaking monolithic key/value tensors into smaller chunks and structuring them into the auxiliary prefix tree. Consequently, on top of the prefix-tree based KV cache, we design an efficient self-attention kernel, where a two-phase partition algorithm is implemented to improve the data locality during self-attention computation in the presence of shared system prompts. Experiments show that ChunkAttention can speed up the self-attention kernel by 3.2-4.8$\times$ compared to the state-of-the-art implementation, with the length of the system prompt ranging from 1024 to 4096.	true	true	Lu Ye and Ze Tao and Yong Huang and Yang Li	2,024	null	https://aclanthology.org/2024.acl-long.623	null	null	ChunkAttention: Efficient Self-Attention with Prefix-Aware KV Cache and Two-Phase Partition	[PDF] Efficient Self-Attention with Prefix-Aware KV Cache and Two-Phase ...	https://aclanthology.org/2024.acl-long.623.pdf	ChunkAttention is a prefix-aware self-attention module that uses a prefix-aware KV cache and two-phase partition to improve memory utilization
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	cachedattention	\cite{cachedattention}	Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttention	null	null	true	false	Bin Gao and Zhuomin He and Puru Sharma and Qingxuan Kang and Djordje Jevdjic and Junbo Deng and Xingkun Yang and Zhou Yu and Pengfei Zuo	2,024	null	https://www.usenix.org/conference/atc24/presentation/gao-bin-cost	null	null	Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttention	Cost-Efficient Large Language Model Serving for Multi-turn ... - arXiv	https://arxiv.org/abs/2403.19708	This paper proposes CachedAttention, a new attention mechanism that enables reuse of KV caches across multi-turn conversations, significantly reducing the
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	promptcache	\cite{promptcache}	Prompt Cache: Modular Attention Reuse for Low-Latency Inference	http://arxiv.org/abs/2311.04934v2	We present Prompt Cache, an approach for accelerating inference for large language models (LLM) by reusing attention states across different LLM prompts. Many input prompts have overlapping text segments, such as system messages, prompt templates, and documents provided for context. Our key insight is that by precomputing and storing the attention states of these frequently occurring text segments on the inference server, we can efficiently reuse them when these segments appear in user prompts. Prompt Cache employs a schema to explicitly define such reusable text segments, called prompt modules. The schema ensures positional accuracy during attention state reuse and provides users with an interface to access cached states in their prompt. Using a prototype implementation, we evaluate Prompt Cache across several LLMs. We show that Prompt Cache significantly reduce latency in time-to-first-token, especially for longer prompts such as document-based question answering and recommendations. The improvements range from 8x for GPU-based inference to 60x for CPU-based inference, all while maintaining output accuracy and without the need for model parameter modifications.	true	true	In Gim and Guojun Chen and Seung{-}Seob Lee and Nikhil Sarda and Anurag Khandelwal and Lin Zhong	2,024	null	https://proceedings.mlsys.org/paper\_files/paper/2024/hash/a66caa1703fe34705a4368c3014c1966-Abstract-Conference.html	null	null	Prompt Cache: Modular Attention Reuse for Low-Latency Inference	[PDF] Prompt Cache: Modular Attention Reuse for Low-Latency Inference	https://proceedings.mlsys.org/paper_files/paper/2024/file/a66caa1703fe34705a4368c3014c1966-Paper-Conference.pdf	Prompt Cache accelerates LLM inference by reusing attention states of frequently occurring text segments, precomputed and stored in memory.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	sglang	\cite{sglang}	Efficiently Programming Large Language Models using SGLang	null	null	true	false	Lianmin Zheng and Liangsheng Yin and Zhiqiang Xie and Jeff Huang and Chuyue Sun and Cody Hao Yu and Shiyi Cao and Christos Kozyrakis and Ion Stoica and Joseph E. Gonzalez and Clark W. Barrett and Ying Sheng	2,023	null	https://doi.org/10.48550/arXiv.2312.07104	10.48550/ARXIV.2312.07104	CoRR	Efficiently Programming Large Language Models using SGLang	Efficiently Programming Large Language Models using SGLang	https://arxiv.org/html/2312.07104v1	SGLang simplifies the writing of LLM programs and boosts execution efficiency. Our experiments demonstrate that SGLang can speed up common LLM tasks by up to 5
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	cacheblend	\cite{cacheblend}	CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion	http://arxiv.org/abs/2405.16444v3	Large language models (LLMs) often incorporate multiple text chunks in their inputs to provide the necessary contexts. To speed up the prefill of the long LLM inputs, one can pre-compute the KV cache of a text and re-use the KV cache when the context is reused as the prefix of another LLM input. However, the reused text chunks are not always the input prefix, which makes precomputed KV caches not directly usable since they ignore the text's cross-attention with the preceding texts. Thus, the benefits of reusing KV caches remain largely unrealized. This paper tackles just one challenge: when an LLM input contains multiple text chunks, how to quickly combine their precomputed KV caches in order to achieve the same generation quality as the expensive full prefill (i.e., without reusing KV cache)? This challenge naturally arises in retrieval-augmented generation (RAG) where the input is supplemented with multiple retrieved texts as the context. We present CacheBlend, a scheme that reuses the precomputed KV caches, regardless prefix or not, and selectively recomputes the KV values of a small subset of tokens to partially update each reused KV cache. In the meantime, the small extra delay for recomputing some tokens can be pipelined with the retrieval of KV caches within the same job, allowing CacheBlend to store KV caches in slower devices with more storage capacity while retrieving them without increasing the inference delay. By comparing CacheBlend with the state-of-the-art KV cache reusing schemes on three open-source LLMs of various sizes and four popular benchmark datasets of different tasks, we show that CacheBlend reduces time-to-first-token (TTFT) by 2.2-3.3x and increases the inference throughput by 2.8-5x from full KV recompute without compromising generation quality. The code is available at https://github.com/LMCache/LMCache.	true	true	Jiayi Yao and Hanchen Li and Yuhan Liu and Siddhant Ray and Yihua Cheng and Qizheng Zhang and Kuntai Du and Shan Lu and Junchen Jiang	2,024	null	https://doi.org/10.48550/arXiv.2405.16444	10.48550/ARXIV.2405.16444	CoRR	CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion	CacheBlend: Fast Large Language Model Serving for RAG ... - arXiv	https://arxiv.org/abs/2405.16444	Image 4: arxiv logo>cs> arXiv:2405.16444 View a PDF of the paper titled CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion, by Jiayi Yao and 8 other authors View a PDF of the paper titled CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion, by Jiayi Yao and 8 other authors - [x] Bibliographic Explorer Toggle - [x] Connected Papers Toggle - [x] Litmaps Toggle - [x] alphaXiv Toggle - [x] Links to Code Toggle - [x] DagsHub Toggle - [x] GotitPub Toggle - [x] Huggingface Toggle - [x] Links to Code Toggle - [x] ScienceCast Toggle - [x] Replicate Toggle - [x] Spaces Toggle - [x] Core recommender toggle - [x] IArxiv recommender toggle
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	openaiapi	\cite{openaiapi}	OpenAI developer platform	null	null	true	false	OpenAI	null	null	null	null	null	OpenAI developer platform	Introducing Verdi, an AI dev platform powered by GPT-4o - OpenAI	https://openai.com/index/mercado-libre/	Verdi, a development platform layer using GPT-4o, GPT-4o mini, and GPT-3.5 Turbo, which is transforming how Mercado Libre handles customer service and other
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	genimiapi	\cite{genimiapi}	Gemini API	null	null	true	false	Google	2,025	null	null	null	null	Gemini API	Gemini Developer API \| Gemma open models \| Google AI for ...	https://ai.google.dev/	Gemini Developer API \| Gemma open models \| Google AI for Developers - Gemini Showcase - Gemini Showcase ### Integrate Google AI models with an API key Build with cutting-edge AI models, like Gemini, Imagen, and Veo, from Google DeepMind Integrate Google AI models with an API key Unlock AI capabilities for your apps with a simple call to the Gemini API. Integrate AI models like Gemini Nano into web apps with Chrome's built-in web platform APIs. Build trusted and secure AI with guidance for responsible design, development, and deployment of models and applications. See how the Ruby-based AI agent framework empowers developer teams to be more productive with the power of Gemini models.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	claudeapi	\cite{claudeapi}	Claude API	null	null	true	false	Anthropic	2,025	null	null	null	null	Claude API	Anthropic API	https://docs.anthropic.com/en/home	Home - Anthropic Claude Documentation Learn how to get started with the Anthropic API, the Console, and Claude Code. Explore the advanced features and capabilities now available in Claude.## API reference Integrate and scale using our API and SDKs.## Anthropic Console Learn about changes and new features in Claude and the API.## Upgrade to Claude 4 Upgrade to the latest model to access new tools and features available in Claude 4. ## Claude Code ## Claude Code quickstart Get started with Claude Code.## Claude Code reference Consult the Claude Code reference documentation for details on feature implementation and configuration.## Claude Code release notes Learn about changes and new features in Claude Code. See replicable code samples and implementations.## Anthropic Quickstarts
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	mooncake	\cite{mooncake}	Mooncake Trace	null	null	true	false	null	2,025	null	null	null	null	Mooncake Trace	kvcache-ai/Mooncake - GitHub	https://github.com/kvcache-ai/Mooncake	Moonshot AI. Now both the Transfer Engine and Mooncake Store are open-sourced! This repository also hosts its technical report and the open sourced traces.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	hu2024epic	\cite{hu2024epic}	EPIC: Efficient Position-Independent Context Caching for Serving Large Language Models	null	null	true	false	Junhao Hu and Wenrui Huang and Haoyi Wang and Weidong Wang and Tiancheng Hu and Qin Zhang and Hao Feng and Xusheng Chen and Yizhou Shan and Tao Xie	2,024	null	https://arxiv.org/abs/2410.15332	null	null	EPIC: Efficient Position-Independent Context Caching for Serving Large Language Models	EPIC: Efficient Position-Independent Caching for Serving Large...	https://openreview.net/forum?id=qjd3ZUiHRT&referrer=%5Bthe%20profile%20of%20Yizhou%20Shan%5D(%2Fprofile%3Fid%3D~Yizhou_Shan2)	Summary: This paper introduces PICI, an efficient position-independent context caching system for serving large language models. The system pre-computes the KV
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	streamingllm	\cite{streamingllm}	Efficient Streaming Language Models with Attention Sinks	http://arxiv.org/abs/2309.17453v4	Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory. Secondly, popular LLMs cannot generalize to longer texts than the training sequence length. Window attention, where only the most recent KVs are cached, is a natural approach -- but we show that it fails when the text length surpasses the cache size. We observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a "sink" even if they are not semantically important. Based on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence lengths without any fine-tuning. We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more. In addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2x speedup. Code and datasets are provided at https://github.com/mit-han-lab/streaming-llm.	true	true	Guangxuan Xiao and Yuandong Tian and Beidi Chen and Song Han and Mike Lewis	2,024	null	https://openreview.net/forum?id=NG7sS51zVF	null	null	Efficient Streaming Language Models with Attention Sinks	Efficient Streaming Language Models with Attention Sinks	http://arxiv.org/pdf/2309.17453v4	Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory. Secondly, popular LLMs cannot generalize to longer texts than the training sequence length. Window attention, where only the most recent KVs are cached, is a natural approach -- but we show that it fails when the text length surpasses the cache size. We observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a "sink" even if they are not semantically important. Based on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence lengths without any fine-tuning. We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more. In addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2x speedup. Code and datasets are provided at https://github.com/mit-han-lab/streaming-llm.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	h2o	\cite{h2o}	{H2O:} Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models	null	null	true	false	Zhenyu Zhang and Ying Sheng and Tianyi Zhou and Tianlong Chen and Lianmin Zheng and Ruisi Cai and Zhao Song and Yuandong Tian and Christopher R{\'{e}} and Clark W. Barrett and Zhangyang Wang and Beidi Chen	2,023	null	http://papers.nips.cc/paper\_files/paper/2023/hash/6ceefa7b15572587b78ecfcebb2827f8-Abstract-Conference.html	null	null	{H2O:} Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models	Hogwild! Inference: Parallel LLM Generation via Concurrent Attention	https://arxiv.org/html/2504.06261v1	H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	infinigen	\cite{infinigen}	InfiniGen: Efficient Generative Inference of Large Language Models with Dynamic KV Cache Management	http://arxiv.org/abs/2406.19707v1	Transformer-based large language models (LLMs) demonstrate impressive performance across various natural language processing tasks. Serving LLM inference for generating long contents, however, poses a challenge due to the enormous memory footprint of the transient state, known as the key-value (KV) cache, which scales with the sequence length and batch size. In this paper, we present InfiniGen, a novel KV cache management framework tailored for long-text generation, which synergistically works with modern offloading-based inference systems. InfiniGen leverages the key insight that a few important tokens that are essential for computing the subsequent attention layer in the Transformer can be speculated by performing a minimal rehearsal with the inputs of the current layer and part of the query weight and key cache of the subsequent layer. This allows us to prefetch only the essential KV cache entries (without fetching them all), thereby mitigating the fetch overhead from the host memory in offloading-based LLM serving systems. Our evaluation on several representative LLMs shows that InfiniGen improves the overall performance of a modern offloading-based system by up to 3.00x compared to prior KV cache management methods while offering substantially better model accuracy.	true	true	Wonbeom Lee and Jungi Lee and Junghwan Seo and Jaewoong Sim	2,024	null	https://www.usenix.org/conference/osdi24/presentation/lee	null	null	InfiniGen: Efficient Generative Inference of Large Language Models with Dynamic KV Cache Management	InfiniGen: Efficient Generative Inference of Large Language Models ...	https://arxiv.org/abs/2406.19707	In this paper, we present InfiniGen, a novel KV cache management framework tailored for long-text generation, which synergistically works with modern
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	pyramidkv	\cite{pyramidkv}	PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling	http://arxiv.org/abs/2406.02069v4	In this study, we investigate whether attention-based information flow inside large language models (LLMs) is aggregated through noticeable patterns for long context processing. Our observations reveal that LLMs aggregate information through Pyramidal Information Funneling where attention is scattering widely in lower layers, progressively consolidating within specific contexts, and ultimately focusing on critical tokens (a.k.a massive activation or attention sink) in higher layers. Motivated by these insights, we developed PyramidKV, a novel and effective KV cache compression method. This approach dynamically adjusts the KV cache size across different layers, allocating more cache in lower layers and less in higher ones, diverging from traditional methods that maintain a uniform KV cache size. Our experimental evaluations, utilizing the LongBench benchmark, show that PyramidKV matches the performance of models with a full KV cache while retaining only 12% of the KV cache, thus significantly reducing memory usage. In scenarios emphasizing memory efficiency, where only 0.7% of the KV cache is maintained, PyramidKV surpasses other KV cache compression techniques, achieving up to a 20.5 absolute accuracy improvement on TREC dataset. In the Needle-in-a-Haystack experiment, PyramidKV outperforms competing methods in maintaining long-context comprehension in LLMs; notably, retaining just 128 KV cache entries enables the LLAMA-3-70B model to achieve 100.0 Acc. performance.	true	true	Zefan Cai and Yichi Zhang and Bofei Gao and Yuliang Liu and Tianyu Liu and Keming Lu and Wayne Xiong and Yue Dong and Baobao Chang and Junjie Hu and Wen Xiao	2,024	null	https://doi.org/10.48550/arXiv.2406.02069	10.48550/ARXIV.2406.02069	CoRR	PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling	PyramidKV: Dynamic KV Cache Compression based on Pyramidal...	https://openreview.net/forum?id=jZVNmDiU86	We developed PyramidKV, a novel and effective KV cache compression method. This approach dynamically adjusts the KV cache size across different layers.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	KVQuant	\cite{KVQuant}	KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization	http://arxiv.org/abs/2401.18079v6	LLMs are seeing growing use for applications which require large context windows, and with these large context windows KV cache activations surface as the dominant contributor to memory consumption during inference. Quantization is a promising approach for compressing KV cache activations; however, existing solutions fail to represent activations accurately in sub-4-bit precision. Our work, KVQuant, facilitates low precision KV cache quantization by incorporating several novel methods: (i) Per-Channel Key Quantization, where we adjust the dimension along which we quantize the Key activations to better match the distribution; (ii) Pre-RoPE Key Quantization, where we quantize Key activations before the rotary positional embedding to mitigate its impact on quantization; (iii) Non-Uniform KV Cache Quantization, where we derive per-layer sensitivity-weighted non-uniform datatypes that better represent the distributions; and (iv) Per-Vector Dense-and-Sparse Quantization, where we isolate outliers separately for each vector to minimize skews in quantization ranges. By applying our method to the LLaMA, Llama-2, Llama-3, and Mistral models, we achieve < 0.1 perplexity degradation with 3-bit quantization on both Wikitext-2 and C4, outperforming existing approaches. Our method enables serving LLaMA-7B with a context length of up to 1 million on a single A100-80GB GPU and up to 10 million on an 8-GPU system. We develop custom CUDA kernels for KVQuant, showing that we can achieve up to ~1.7x speedups, compared to baseline fp16 matrix-vector multiplications, for the LLaMA-7B model.	true	true	Coleman Hooper and Sehoon Kim and Hiva Mohammadzadeh and Michael W. Mahoney and Yakun Sophia Shao and Kurt Keutzer and Amir Gholami	2,024	null	http://papers.nips.cc/paper\_files/paper/2024/hash/028fcbcf85435d39a40c4d61b42c99a4-Abstract-Conference.html	null	null	KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization	KVQuant: Towards 10 Million Context Length LLM Inference with KV ...	https://github.com/SqueezeAILab/KVQuant	GitHub - SqueezeAILab/KVQuant: [NeurIPS 2024] KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization [NeurIPS 2024] KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization [Paper] KVQuant is a methodology for efficient KV cache quantization that incorporates several innovations to acheive accurate low-precision quantization, thereby enabling efficient long context length inference. TLDR: KVQuant addresses the memory bottleneck with long context length inference by quantizing the KV cache to low precision. title={KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization}, [NeurIPS 2024] KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	lruk	\cite{lruk}	The {LRU-K} Page Replacement Algorithm For Database Disk Buffering	null	null	true	false	Elizabeth J. O'Neil and Patrick E. O'Neil and Gerhard Weikum	1,993	null	https://doi.org/10.1145/170035.170081	10.1145/170035.170081	null	The {LRU-K} Page Replacement Algorithm For Database Disk Buffering	[PDF] The LRU-K Page Replacement Algorithm For Database Disk Buffering	https://www.cs.cmu.edu/~natassa/courses/15-721/papers/p297-o_neil.pdf	The basic idea of. LRU-K is to keep track of the times of the last K references to popular database pages, using this information to statis- tieall y estimate
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	slru	\cite{slru}	Caching Strategies to Improve Disk System Performance	null	null	true	false	Ramakrishna Karedla and J. Spencer Love and Bradley G. Wherry	1,994	null	https://doi.org/10.1109/2.268884	10.1109/2.268884	Computer	Caching Strategies to Improve Disk System Performance	Caching strategies to improve disk system performance - IEEE Xplore	http://ieeexplore.ieee.org/document/268884/	In this article, we examine the use of caching as a means to increase system response time and improve the data throughput of the disk subsystem.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	twoq	\cite{twoq}	2Q: {A} Low Overhead High Performance Buffer Management Replacement Algorithm	null	null	true	false	Theodore Johnson and Dennis E. Shasha	1,994	null	http://www.vldb.org/conf/1994/P439.PDF	null	null	2Q: {A} Low Overhead High Performance Buffer Management Replacement Algorithm	2Q: A Low Overhead High Performance Buffer Management ...	https://dl.acm.org/doi/10.5555/645920.672996	2Q: A Low Overhead High Performance Buffer Management Replacement Algorithm. Authors: Theodore Johnson.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	eelru	\cite{eelru}	{EELRU:} Simple and Effective Adaptive Page Replacement	null	null	true	false	Yannis Smaragdakis and Scott F. Kaplan and Paul R. Wilson	1,999	null	https://doi.org/10.1145/301453.301486	10.1145/301453.301486	null	{EELRU:} Simple and Effective Adaptive Page Replacement	EELRU: Simple and Effective Adaptive Page Replacement	https://www.researchgate.net/publication/2822757_EELRU_Simple_and_Effective_Adaptive_Page_Replacement	EELRU is a simple adaptive replacement algorithm, which uses only the kind of information needed by LRU---how recently each page has been touched relative to
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	lrfu	\cite{lrfu}	{LRFU:} {A} Spectrum of Policies that Subsumes the Least Recently Used and Least Frequently Used Policies	null	null	true	false	Donghee Lee and Jongmoo Choi and Jong{-}Hun Kim and Sam H. Noh and Sang Lyul Min and Yookun Cho and Chong{-}Sang Kim	2,001	null	https://doi.org/10.1109/TC.2001.970573	10.1109/TC.2001.970573	{IEEE} Trans. Computers	{LRFU:} {A} Spectrum of Policies that Subsumes the Least Recently Used and Least Frequently Used Policies	[PDF] LRFU: a spectrum of policies that subsumes the least recently used ...	https://www.openu.ac.il/home/wiseman/2os/lru/lrfu.pdf	Of these, the Least Recently Used (LRU) and the. Least Frequently Used (LFU) block replacement policies constitute the two main streams. The LRU policy and its.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	lirs	\cite{lirs}	{LIRS:} an efficient low inter-reference recency set replacement policy to improve buffer cache performance	null	null	true	false	Song Jiang and Xiaodong Zhang	2,002	null	https://doi.org/10.1145/511334.511340	10.1145/511334.511340	null	{LIRS:} an efficient low inter-reference recency set replacement policy to improve buffer cache performance	LIRS: an efficient low inter-reference recency set replacement policy ...	https://www.researchgate.net/publication/367088056_LIRS_an_efficient_low_inter-reference_recency_set_replacement_policy_to_improve_buffer_cache_performance	Many studies are focused on cache replacement algorithms, such as FIFO, LRU, LFU, and some advanced cache algorithms like ARC [19], LIRS [15] and 2Q [16].
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	arc	\cite{arc}	{ARC:} {A} Self-Tuning, Low Overhead Replacement Cache	null	null	true	false	Nimrod Megiddo and Dharmendra S. Modha	2,003	null	http://www.usenix.org/events/fast03/tech/megiddo.html	null	null	{ARC:} {A} Self-Tuning, Low Overhead Replacement Cache	[PDF] ARC: A Self-Tuning, Low Overhead Replacement Cache	https://www.cs.cmu.edu/~natassa/courses/15-721/papers/arcfast.pdf	We propose a new cache management policy, namely, Adaptive. Replacement Cache (ARC), that has several advantages. In response to evolving and changing access
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	mq	\cite{mq}	Second-Level Buffer Cache Management	null	null	true	false	Yuanyuan Zhou and Zhifeng Chen and Kai Li	2,004	null	https://doi.org/10.1109/TPDS.2004.13	10.1109/TPDS.2004.13	{IEEE} Trans. Parallel Distributed Syst.	Second-Level Buffer Cache Management	[PDF] Second-Level Buffer Cache Management	https://www.openu.ac.il/home/wiseman/2os/lru/mq.pdf	This is a local cache replacement algorithm because it manages an L2 buffer cache without any information from first-level.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	car	\cite{car}	{CAR:} Clock with Adaptive Replacement	null	null	true	false	Sorav Bansal and Dharmendra S. Modha	2,004	null	http://www.usenix.org/events/fast04/tech/bansal.html	null	null	{CAR:} Clock with Adaptive Replacement	CAR: Clock with Adaptive Replacement - Stanford CS Theory	http://theory.stanford.edu/~sbansal/pubs/fast04.pdf	by S Bansal · Cited by 412 — CAR is a new algorithm that improves upon CLOCK by being scan-resistant, self-tuning, and adaptively capturing recency and frequency features.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	clockpro	\cite{clockpro}	CLOCK-Pro: An Effective Improvement of the {CLOCK} Replacement	null	null	true	false	Song Jiang and Feng Chen and Xiaodong Zhang	2,005	null	http://www.usenix.org/events/usenix05/tech/general/jiang.html	null	null	CLOCK-Pro: An Effective Improvement of the {CLOCK} Replacement	CLOCK-Pro: An Effective Improvement of the CLOCK Replacement	https://www.usenix.org/conference/2005-usenix-annual-technical-conference/clock-pro-effective-improvement-clock-replacement	We propose an improved CLOCK replacement policy, called CLOCK-Pro. By additionally keeping track of a limited number of replaced pages, CLOCK-Pro works in a
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	DBLP:journals/tos/EinzigerEFM22	\cite{DBLP:journals/tos/EinzigerEFM22}	Lightweight Robust Size Aware Cache Management	http://arxiv.org/abs/2105.08770v2	Modern key-value stores, object stores, Internet proxy caches, as well as Content Delivery Networks (CDN) often manage objects of diverse sizes, e.g., blobs, video files of different lengths, images with varying resolution, and small documents. In such workloads, size-aware cache policies outperform size-oblivious algorithms. Unfortunately, existing size-aware algorithms tend to be overly complicated and computationally~expensive. Our work follows a more approachable pattern; we extend the prevalent (size-oblivious) TinyLFU cache admission policy to handle variable sized items. Implementing our approach inside two popular caching libraries only requires minor changes. We show that our algorithms yield competitive or better hit-ratios and byte hit-ratios compared to the state of the art size-aware algorithms such as AdaptSize, LHD, LRB, and GDSF. Further, a runtime comparison indicates that our implementation is faster by up to x3 compared to the best alternative, i.e., it imposes much lower CPU overhead.	true	true	Gil Einziger and Ohad Eytan and Roy Friedman and Benjamin Manes	2,022	null	https://doi.org/10.1145/3507920	10.1145/3507920	{ACM} Trans. Storage	Lightweight Robust Size Aware Cache Management	Lightweight Robust Size Aware Cache Management	http://arxiv.org/pdf/2105.08770v2	Modern key-value stores, object stores, Internet proxy caches, as well as Content Delivery Networks (CDN) often manage objects of diverse sizes, e.g., blobs, video files of different lengths, images with varying resolution, and small documents. In such workloads, size-aware cache policies outperform size-oblivious algorithms. Unfortunately, existing size-aware algorithms tend to be overly complicated and computationally~expensive. Our work follows a more approachable pattern; we extend the prevalent (size-oblivious) TinyLFU cache admission policy to handle variable sized items. Implementing our approach inside two popular caching libraries only requires minor changes. We show that our algorithms yield competitive or better hit-ratios and byte hit-ratios compared to the state of the art size-aware algorithms such as AdaptSize, LHD, LRB, and GDSF. Further, a runtime comparison indicates that our implementation is faster by up to x3 compared to the best alternative, i.e., it imposes much lower CPU overhead.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	lhd	\cite{lhd}	{LHD:} Improving Cache Hit Rate by Maximizing Hit Density	null	null	true	false	Nathan Beckmann and Haoxian Chen and Asaf Cidon	2,018	null	https://www.usenix.org/conference/nsdi18/presentation/beckmann	null	null	{LHD:} Improving Cache Hit Rate by Maximizing Hit Density	LHD: improving cache hit rate by maximizing hit density	https://dl.acm.org/doi/10.5555/3307441.3307475	We introduce least hit density (LHD), a novel eviction policy for key-value caches. LHD predicts each object's expected hits-per-space-consumed (hit density).
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	cacheus	\cite{cacheus}	Learning Cache Replacement with {CACHEUS}	null	null	true	false	Liana V. Rodriguez and Farzana Beente Yusuf and Steven Lyons and Eysler Paz and Raju Rangaswami and Jason Liu and Ming Zhao and Giri Narasimhan	2,021	null	https://www.usenix.org/conference/fast21/presentation/rodriguez	null	null	Learning Cache Replacement with {CACHEUS}	Learning Cache Replacement with Cacheus	https://www.usenix.org/system/files/fast21-rodriguez.pdf	by LV Rodriguez · 2021 · Cited by 125 — Furthermore, CACHEUS enables augmenting state-of-the-art algorithms (e.g., LIRS, ARC) by combining it with a complementary cache replacement
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	sieve	\cite{sieve}	{SIEVE} is Simpler than {LRU:} an Efficient Turn-Key Eviction Algorithm for Web Caches	null	null	true	false	Yazhuo Zhang and Juncheng Yang and Yao Yue and Ymir Vigfusson and K. V. Rashmi	2,024	null	https://www.usenix.org/conference/nsdi24/presentation/zhang-yazhuo	null	null	{SIEVE} is Simpler than {LRU:} an Efficient Turn-Key Eviction Algorithm for Web Caches	SIEVE - An Efficient Turn-Key Eviction Algorithm for Web Caches	https://www.classcentral.com/course/youtube-nsdi-24-sieve-is-simpler-than-lru-an-efficient-turn-key-eviction-algorithm-for-web-caches-294624	Discover how SIEVE outperforms traditional algorithms like LRU in simplicity, efficiency, and scalability for web cache workloads. Learn about the algorithm's
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	cherkasova1998improving	\cite{cherkasova1998improving}	Improving WWW proxies performance with greedy-dual-size-frequency caching policy	null	null	true	false	Cherkasova, Ludmila	1,998	null	null	null	null	Improving WWW proxies performance with greedy-dual-size-frequency caching policy	Improving WWW proxies performance with Greedy-Dual- ...	https://www.researchgate.net/publication/228542715_Improving_WWW_proxies_performance_with_Greedy-Dual-Size-Frequency_caching_policy	This paper introduces the Greedy-Dual-Size-Frequency caching policy to maximize hit and byte hit rates for WWW proxies. Proposed caching strategy incorporates
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	yang2020twemcache	\cite{yang2020twemcache}	A large scale analysis of hundreds of in-memory cache clusters at Twitter	null	null	true	false	Juncheng Yang and Yao Yue and K. V. Rashmi	2,020	null	https://www.usenix.org/conference/osdi20/presentation/yang	null	null	A large scale analysis of hundreds of in-memory cache clusters at Twitter	[PDF] A large scale analysis of hundreds of in-memory cache clusters at ...	https://www.usenix.org/system/files/osdi20-yang.pdf	This paper is included in the Proceedings of the 14th USENIX Symposium on Operating Systems Design and Implementation November 4–6, 2020 978-1-939133-19-9 Open access to the Proceedings of the 14th USENIX Symposium on Operating Systems Design and Implementation is sponsored by USENIX A large scale analysis of hundreds of in-memory cache clusters at Twitter Juncheng Yang, Carnegie Mellon University; Yao Yue, Twitter; K. When memory is full, 192 14th USENIX Symposium on Operating Systems Design and Implementation USENIX Association #cluster request rate cache size cpu cores 0.00 0.25 0.50 0.75 1.00 Fraction of use case storage computation transient Figure 2: Resources consumed for the three cache use cases.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	berg2020cachelib	\cite{berg2020cachelib}	The {CacheLib} Caching Engine: Design and Experiences at Scale	null	null	true	false	Benjamin Berg and Daniel S. Berger and Sara McAllister and Isaac Grosof and Sathya Gunasekar and Jimmy Lu and Michael Uhlar and Jim Carrig and Nathan Beckmann and Mor Harchol-Balter and Gregory R. Ganger	2,020	null	https://www.usenix.org/conference/osdi20/presentation/berg	null	null	The {CacheLib} Caching Engine: Design and Experiences at Scale	The CacheLib Caching Engine: Design and Experiences at Scale	https://www.usenix.org/conference/osdi20/presentation/berg	CacheLib is a general-purpose caching engine, designed based on experiences with a range of caching use cases at Facebook, that facilitates the easy
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	icebreaker	\cite{icebreaker}	IceBreaker: warming serverless functions better with heterogeneity	null	null	true	false	Rohan Basu Roy and Tirthak Patel and Devesh Tiwari	2,022	null	https://doi.org/10.1145/3503222.3507750	10.1145/3503222.3507750	null	IceBreaker: warming serverless functions better with heterogeneity	[PDF] IceBreaker: Warming Serverless Functions Better with Heterogeneity	http://www1.ece.neu.edu/~ningfang/SimPaper/icebreaker-ASPLOS22.pdf	IceBreaker is a novel function pre-warming and keep-alive scheme for serverless functions that exploit server-heterogeneity to lower the keep-alive cost and
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	fasscache	\cite{fasscache}	FaasCache: keeping serverless computing alive with greedy-dual caching	null	null	true	false	Alexander Fuerst and Prateek Sharma	2,021	null	https://doi.org/10.1145/3445814.3446757	10.1145/3445814.3446757	null	FaasCache: keeping serverless computing alive with greedy-dual caching	[PDF] FaasCache: Keeping Serverless Computing Alive with Greedy-Dual ...	https://afuerst.github.io/assets/FaasCache.pdf	Keep-alive policies must keep functions alive based on their resource and usage characteristics, which is challenging due to the diversity in FaaS workloads.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	DBLP:conf/osdi/ZhongLCHZL0024	\cite{DBLP:conf/osdi/ZhongLCHZL0024}	DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving	http://arxiv.org/abs/2401.09670v3	DistServe improves the performance of large language models (LLMs) serving by disaggregating the prefill and decoding computation. Existing LLM serving systems colocate the two phases and batch the computation of prefill and decoding across all users and requests. We find that this strategy not only leads to strong prefill-decoding interferences but also couples the resource allocation and parallelism plans for both phases. LLM applications often emphasize individual latency for each phase: time to first token (TTFT) for the prefill phase and time per output token (TPOT) of each request for the decoding phase. In the presence of stringent latency requirements, existing systems have to prioritize one latency over the other, or over-provision compute resources to meet both. DistServe assigns prefill and decoding computation to different GPUs, hence eliminating prefill-decoding interferences. Given the application's TTFT and TPOT requirements, DistServe co-optimizes the resource allocation and parallelism strategy tailored for each phase. DistServe also places the two phases according to the serving cluster's bandwidth to minimize the communication caused by disaggregation. As a result, DistServe significantly improves LLM serving performance in terms of the maximum rate that can be served within both TTFT and TPOT constraints on each GPU. Our evaluations show that on various popular LLMs, applications, and latency requirements, DistServe can serve 7.4x more requests or 12.6x tighter SLO, compared to state-of-the-art systems, while staying within latency constraints for > 90% of requests.	true	true	Yinmin Zhong and Shengyu Liu and Junda Chen and Jianbo Hu and Yibo Zhu and Xuanzhe Liu and Xin Jin and Hao Zhang	2,024	null	https://www.usenix.org/conference/osdi24/presentation/zhong-yinmin	null	null	DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving	[PDF] DistServe: Disaggregating Prefill and Decoding for Goodput ...	https://www.usenix.org/system/files/osdi24-zhong-yinmin.pdf	July 10–12, 2024 • Santa Clara, CA, USA 978-1-939133-40-3 Open access to the Proceedings of the 18th USENIX Symposium on Operating Systems Design and Implementation is sponsored by DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving Yinmin Zhong and Shengyu Liu, Peking University; Junda Chen, UC San Diego; Jianbo Hu, Peking University; Yibo Zhu, StepFun; Xuanzhe Liu and Xin Jin, Peking University; Hao Zhang, UC San Diego DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving Yinmin Zhong1 Shengyu Liu1 Junda Chen3 Jianbo Hu1 Yibo Zhu2 Xuanzhe Liu1 Xin Jin1 Hao Zhang3 1School of Computer Science, Peking University 2StepFun 3UC San Diego Abstract DistServe improves the performance of large language mod-els (LLMs) serving by disaggregating the prefill and decoding computation.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	DBLP:journals/corr/abs-2404-09526	\cite{DBLP:journals/corr/abs-2404-09526}	LoongServe: Efficiently Serving Long-Context Large Language Models with Elastic Sequence Parallelism	http://arxiv.org/abs/2404.09526v2	The context window of large language models (LLMs) is rapidly increasing, leading to a huge variance in resource usage between different requests as well as between different phases of the same request. Restricted by static parallelism strategies, existing LLM serving systems cannot efficiently utilize the underlying resources to serve variable-length requests in different phases. To address this problem, we propose a new parallelism paradigm, elastic sequence parallelism (ESP), to elastically adapt to the variance between different requests and phases. Based on ESP, we design and build LoongServe, an LLM serving system that (1) improves computation efficiency by elastically adjusting the degree of parallelism in real-time, (2) improves communication efficiency by reducing key-value cache migration overhead and overlapping partial decoding communication with computation, and (3) improves GPU memory efficiency by reducing key-value cache fragmentation across instances. Our evaluation under diverse real-world datasets shows that LoongServe improves the maximum throughput by up to 3.85$\times$ compared to the chunked prefill and 5.81$\times$ compared to the prefill-decoding disaggregation.	true	true	Bingyang Wu and Shengyu Liu and Yinmin Zhong and Peng Sun and Xuanzhe Liu and Xin Jin	2,024	null	https://doi.org/10.48550/arXiv.2404.09526	10.48550/ARXIV.2404.09526	CoRR	LoongServe: Efficiently Serving Long-Context Large Language Models with Elastic Sequence Parallelism	LoongServe: Efficiently Serving Long-Context Large Language ...	https://colab.ws/articles/10.1145%2F3694715.3695948	LoongServe: Efficiently Serving Long-Context Large Language Models with Elastic Sequence Parallelism. Bingyang Wu 1. ,. Shengyu Liu 1. ,. Yinmin
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	DBLP:conf/sosp/KwonLZ0ZY0ZS23	\cite{DBLP:conf/sosp/KwonLZ0ZY0ZS23}	Efficient Memory Management for Large Language Model Serving with PagedAttention	http://arxiv.org/abs/2309.06180v1	High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2-4$\times$ with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm	true	true	Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph Gonzalez and Hao Zhang and Ion Stoica	2,023	null	https://doi.org/10.1145/3600006.3613165	10.1145/3600006.3613165	null	Efficient Memory Management for Large Language Model Serving with PagedAttention	Efficient Memory Management for Large Language Model ...	https://arxiv.org/pdf/2309.06180	Efficient Memory Management for Large Language Model Serving with PagedAttention Woosuk Kwon 1,∗ Zhuohan Li 1,∗ Siyuan Zhuang 1 Ying Sheng 1,2 Lianmin Zheng 1 Cody Hao Yu 3 Joseph E. Gonzalez 1 Hao Zhang 4 Ion Stoica 1 1 UC Berkeley 2Stanford University 3Independent Researcher 4UC San Diego Abstract High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. To address this problem, we propose PagedAttention, an attention al-gorithm inspired by the classical virtual memory and pag-ing techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce mem-ory usage. To address the above limitations, we propose PagedAt-tention , an attention algorithm inspired by the operating system’s (OS) solution to memory fragmentation and shar-ing: virtual memory with paging . In this work, we build vLLM , a high-throughput distributed LLM serving engine on top of PagedAttention that achieves near-zero waste in KV cache memory.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	alpaserve	\cite{alpaserve}	AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving	http://arxiv.org/abs/2302.11665v2	Model parallelism is conventionally viewed as a method to scale a single large deep learning model beyond the memory limits of a single device. In this paper, we demonstrate that model parallelism can be additionally used for the statistical multiplexing of multiple devices when serving multiple models, even when a single model can fit into a single device. Our work reveals a fundamental trade-off between the overhead introduced by model parallelism and the opportunity to exploit statistical multiplexing to reduce serving latency in the presence of bursty workloads. We explore the new trade-off space and present a novel serving system, AlpaServe, that determines an efficient strategy for placing and parallelizing collections of large deep learning models across a distributed cluster. Evaluation results on production workloads show that AlpaServe can process requests at up to 10x higher rates or 6x more burstiness while staying within latency constraints for more than 99% of requests.	true	true	Zhuohan Li and Lianmin Zheng and Yinmin Zhong and Vincent Liu and Ying Sheng and Xin Jin and Yanping Huang and Zhifeng Chen and Hao Zhang and Joseph E. Gonzalez and Ion Stoica	2,023	null	https://www.usenix.org/conference/osdi23/presentation/li-zhouhan	null	null	AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving	alpa-projects/mms: AlpaServe - GitHub	https://github.com/alpa-projects/mms	This is the official implementation of our OSDI'23 paper: AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving. To reproduce
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	DBLP:conf/osdi/YuJKKC22	\cite{DBLP:conf/osdi/YuJKKC22}	Orca: {A} Distributed Serving System for Transformer-Based Generative Models	null	null	true	false	Gyeong{-}In Yu and Joo Seong Jeong and Geon{-}Woo Kim and Soojeong Kim and Byung{-}Gon Chun	2,022	null	https://www.usenix.org/conference/osdi22/presentation/yu	null	null	Orca: {A} Distributed Serving System for Transformer-Based Generative Models	Orca: A Distributed Serving System for Transformer-Based ... - USENIX	https://www.usenix.org/conference/osdi22/presentation/yu	We have implemented a distributed serving system called ORCA, with additional designs for scalability to models with hundreds of billions of parameters.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	DBLP:conf/isca/PatelCZSGMB24	\cite{DBLP:conf/isca/PatelCZSGMB24}	Splitwise: Efficient generative LLM inference using phase splitting	http://arxiv.org/abs/2311.18677v2	Recent innovations in generative large language models (LLMs) have made their applications and use-cases ubiquitous. This has led to large-scale deployments of these models, using complex, expensive, and power-hungry AI accelerators, most commonly GPUs. These developments make LLM inference efficiency an important challenge. Based on our extensive characterization, we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a memory-intensive token generation, each with distinct latency, throughput, memory, and power characteristics. Despite state-of-the-art batching and scheduling, the token generation phase underutilizes compute resources. Specifically, unlike compute-intensive prompt computation phases, token generation phases do not require the compute capability of the latest GPUs, and can be run with lower power and cost. With Splitwise, we propose splitting the two phases of a LLM inference request on to separate machines. This allows us to use hardware that is well-suited for each phase, and provision resources independently per phase. However, splitting an inference request across machines requires state transfer from the machine running prompt computation over to the machine generating tokens. We implement and optimize this state transfer using the fast back-plane interconnects available in today's GPU clusters. We use the Splitwise technique to design LLM inference clusters using the same or different types of machines for the prompt computation and token generation phases. Our clusters are optimized for three key objectives: throughput, cost, and power. In particular, we show that we can achieve 1.4x higher throughput at 20% lower cost than current designs. Alternatively, we can achieve 2.35x more throughput with the same cost and power budgets.	true	true	Pratyush Patel and Esha Choukse and Chaojie Zhang and Aashaka Shah and {\'{I}}{\~{n}}igo Goiri and Saeed Maleki and Ricardo Bianchini	2,024	null	https://doi.org/10.1109/ISCA59077.2024.00019	10.1109/ISCA59077.2024.00019	null	Splitwise: Efficient generative LLM inference using phase splitting	Splitwise: Efficient generative LLM inference using phase splitting	http://arxiv.org/pdf/2311.18677v2	Recent innovations in generative large language models (LLMs) have made their applications and use-cases ubiquitous. This has led to large-scale deployments of these models, using complex, expensive, and power-hungry AI accelerators, most commonly GPUs. These developments make LLM inference efficiency an important challenge. Based on our extensive characterization, we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a memory-intensive token generation, each with distinct latency, throughput, memory, and power characteristics. Despite state-of-the-art batching and scheduling, the token generation phase underutilizes compute resources. Specifically, unlike compute-intensive prompt computation phases, token generation phases do not require the compute capability of the latest GPUs, and can be run with lower power and cost. With Splitwise, we propose splitting the two phases of a LLM inference request on to separate machines. This allows us to use hardware that is well-suited for each phase, and provision resources independently per phase. However, splitting an inference request across machines requires state transfer from the machine running prompt computation over to the machine generating tokens. We implement and optimize this state transfer using the fast back-plane interconnects available in today's GPU clusters. We use the Splitwise technique to design LLM inference clusters using the same or different types of machines for the prompt computation and token generation phases. Our clusters are optimized for three key objectives: throughput, cost, and power. In particular, we show that we can achieve 1.4x higher throughput at 20% lower cost than current designs. Alternatively, we can achieve 2.35x more throughput with the same cost and power budgets.
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	298501	\cite{298501}	{Cost-Efficient} Large Language Model Serving for Multi-turn Conversations with {CachedAttention}	null	null	true	false	Bin Gao and Zhuomin He and Puru Sharma and Qingxuan Kang and Djordje Jevdjic and Junbo Deng and Xingkun Yang and Zhou Yu and Pengfei Zuo	2,024	null	https://www.usenix.org/conference/atc24/presentation/gao-bin-cost	null	null	{Cost-Efficient} Large Language Model Serving for Multi-turn Conversations with {CachedAttention}	Cost-Efficient Large Language Model Serving for Multi-turn ... - arXiv	https://arxiv.org/abs/2403.19708	View a PDF of the paper titled Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttention, by Bin Gao and 8 other authors To address the problem, this paper proposes CachedAttention, a new attention mechanism that enables reuse of KV caches across multi-turn conversations, significantly reducing the repetitive computation overheads. View a PDF of the paper titled Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttention, by Bin Gao and 8 other authors - [x] Bibliographic Explorer Toggle - [x] Connected Papers Toggle - [x] Litmaps Toggle - [x] alphaXiv Toggle - [x] Links to Code Toggle - [x] DagsHub Toggle - [x] GotitPub Toggle - [x] Links to Code Toggle - [x] ScienceCast Toggle - [x] Replicate Toggle
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	DBLP:journals/corr/abs-2412-17246	\cite{DBLP:journals/corr/abs-2412-17246}	Fast and Live Model Auto Scaling with {O(1)} Host Caching	null	null	true	false	Dingyan Zhang and Haotian Wang and Yang Liu and Xingda Wei and Yizhou Shan and Rong Chen and Haibo Chen	2,024	null	https://doi.org/10.48550/arXiv.2412.17246	10.48550/ARXIV.2412.17246	CoRR	Fast and Live Model Auto Scaling with {O(1)} Host Caching	Fast and Live Model Auto Scaling with 𝑂⁢(1) Host Caching	https://arxiv.org/html/2412.17246v1	Model autoscaling is the key mechanism to achieve serverless model-as-a-service, but it faces a fundamental trade-off between scaling speed and storage/memory
KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider	2506.02634v1	shahrad2020serverless	\cite{shahrad2020serverless}	Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider	http://arxiv.org/abs/2003.03423v3	Function as a Service (FaaS) has been gaining popularity as a way to deploy computations to serverless backends in the cloud. This paradigm shifts the complexity of allocating and provisioning resources to the cloud provider, which has to provide the illusion of always-available resources (i.e., fast function invocations without cold starts) at the lowest possible resource cost. Doing so requires the provider to deeply understand the characteristics of the FaaS workload. Unfortunately, there has been little to no public information on these characteristics. Thus, in this paper, we first characterize the entire production FaaS workload of Azure Functions. We show for example that most functions are invoked very infrequently, but there is an 8-order-of-magnitude range of invocation frequencies. Using observations from our characterization, we then propose a practical resource management policy that significantly reduces the number of function coldstarts,while spending fewerresources than state-of-the-practice policies.	true	true	Mohammad Shahrad and Rodrigo Fonseca and Inigo Goiri and Gohar Chaudhry and Paul Batum and Jason Cooke and Eduardo Laureano and Colby Tresness and Mark Russinovich and Ricardo Bianchini	2,020	null	https://www.usenix.org/conference/atc20/presentation/shahrad	null	null	Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider	Characterizing and Optimizing the Serverless Workload at ...	https://www.usenix.org/system/files/atc20-shahrad.pdf	by M Shahrad · 2020 · Cited by 879 — This paper characterizes Azure Functions' serverless workload, showing most functions are invoked infrequently, and proposes a resource
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	liu2024:visual	\cite{liu2024:visual}	Visual Instruction Tuning	http://arxiv.org/abs/2304.08485v2	Instruction tuning large language models (LLMs) using machine-generated instruction-following data has improved zero-shot capabilities on new tasks, but the idea is less explored in the multimodal field. In this paper, we present the first attempt to use language-only GPT-4 to generate multimodal language-image instruction-following data. By instruction tuning on such generated data, we introduce LLaVA: Large Language and Vision Assistant, an end-to-end trained large multimodal model that connects a vision encoder and LLM for general-purpose visual and language understanding.Our early experiments show that LLaVA demonstrates impressive multimodel chat abilities, sometimes exhibiting the behaviors of multimodal GPT-4 on unseen images/instructions, and yields a 85.1% relative score compared with GPT-4 on a synthetic multimodal instruction-following dataset. When fine-tuned on Science QA, the synergy of LLaVA and GPT-4 achieves a new state-of-the-art accuracy of 92.53%. We make GPT-4 generated visual instruction tuning data, our model and code base publicly available.	true	true	Liu, Haotian and Li, Chunyuan and Wu, Qingyang and Lee, Yong Jae	2,024	null	null	null	Advances in neural information processing systems	Visual Instruction Tuning	Visual Instruction Tuning	http://arxiv.org/pdf/2304.08485v2	Instruction tuning large language models (LLMs) using machine-generated instruction-following data has improved zero-shot capabilities on new tasks, but the idea is less explored in the multimodal field. In this paper, we present the first attempt to use language-only GPT-4 to generate multimodal language-image instruction-following data. By instruction tuning on such generated data, we introduce LLaVA: Large Language and Vision Assistant, an end-to-end trained large multimodal model that connects a vision encoder and LLM for general-purpose visual and language understanding.Our early experiments show that LLaVA demonstrates impressive multimodel chat abilities, sometimes exhibiting the behaviors of multimodal GPT-4 on unseen images/instructions, and yields a 85.1% relative score compared with GPT-4 on a synthetic multimodal instruction-following dataset. When fine-tuned on Science QA, the synergy of LLaVA and GPT-4 achieves a new state-of-the-art accuracy of 92.53%. We make GPT-4 generated visual instruction tuning data, our model and code base publicly available.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	bai2023:qwen	\cite{bai2023:qwen}	Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond	null	null	true	false	Bai, Jinze and Bai, Shuai and Yang, Shusheng and Wang, Shijie and Tan, Sinan and Wang, Peng and Lin, Junyang and Zhou, Chang and Zhou, Jingren	2,023	null	null	null	null	Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond	Qwen-VL: A Versatile Vision-Language Model for Understanding...	https://openreview.net/forum?id=qrGjFJVl3m	Despite the effort in open-sourcing the model and its weights, the reviewers find QWEN-VL lacking in significant research contributions and technical novelty. * _Open-source:_ Qwen-VL is an open-sourced large vision-language model that excels in (i) achieving leading performance across a wide range of vision-language understanding and generation tasks, (ii) offering multi-lingual support, particularly in English and Chinese, (iii) accommodating multi-image and high-resolution inputs, and (iv) demonstrating fine-grained visual perception abilities, particularly in scene text-oriented visual question-answering and visual grounding. Unlike previous representative vision-language models like PaLI-X, which leverages proprietary in-house data and utilize publicly inaccessible model weights (_e.g._, ViT-22B), along with significantly high training costs, our Qwen-VL's training process is more practical and holds considerable referential significance for future research.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	chen2023:sharegpt4v	\cite{chen2023:sharegpt4v}	ShareGPT4V: Improving Large Multi-Modal Models with Better Captions	http://arxiv.org/abs/2311.12793v2	In the realm of large multi-modal models (LMMs), efficient modality alignment is crucial yet often constrained by the scarcity of high-quality image-text data. To address this bottleneck, we introduce the ShareGPT4V dataset, a pioneering large-scale resource featuring 1.2 million highly descriptive captions, which surpasses existing datasets in diversity and information content, covering world knowledge, object properties, spatial relationships, and aesthetic evaluations. Specifically, ShareGPT4V originates from a curated 100K high-quality captions collected from advanced GPT4-Vision and has been expanded to 1.2M with a superb caption model trained on this subset. ShareGPT4V first demonstrates its effectiveness for the Supervised Fine-Tuning (SFT) phase, by substituting an equivalent quantity of detailed captions in existing SFT datasets with a subset of our high-quality captions, significantly enhancing the LMMs like LLaVA-7B, LLaVA-1.5-13B, and Qwen-VL-Chat-7B on the MME and MMBench benchmarks, with respective gains of 222.8/22.0/22.3 and 2.7/1.3/1.5. We further incorporate ShareGPT4V data into both the pre-training and SFT phases, obtaining ShareGPT4V-7B, a superior LMM based on a simple architecture that has remarkable performance across a majority of the multi-modal benchmarks. This project is available at https://ShareGPT4V.github.io to serve as a pivotal resource for advancing the LMMs community.	true	true	Chen, Lin and Li, Jisong and Dong, Xiaoyi and Zhang, Pan and He, Conghui and Wang, Jiaqi and Zhao, Feng and Lin, Dahua	2,023	null	null	null	arXiv preprint arXiv:2311.12793	ShareGPT4V: Improving Large Multi-Modal Models with Better Captions	Improving Large Multi-Modal Models with Better Captions - arXiv	https://arxiv.org/abs/2311.12793	Image 4: arxiv logo>cs> arXiv:2311.12793 arXiv:2311.12793 (cs) View a PDF of the paper titled ShareGPT4V: Improving Large Multi-Modal Models with Better Captions, by Lin Chen and 7 other authors View a PDF of the paper titled ShareGPT4V: Improving Large Multi-Modal Models with Better Captions, by Lin Chen and 7 other authors - [x] Bibliographic Explorer Toggle - [x] Connected Papers Toggle - [x] Litmaps Toggle - [x] scite.ai Toggle - [x] alphaXiv Toggle - [x] Links to Code Toggle - [x] DagsHub Toggle - [x] GotitPub Toggle - [x] Huggingface Toggle - [x] Links to Code Toggle - [x] ScienceCast Toggle - [x] Replicate Toggle - [x] Spaces Toggle - [x] Spaces Toggle - [x] Core recommender toggle
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	li2023:videochat	\cite{li2023:videochat}	VideoChat: Chat-Centric Video Understanding	http://arxiv.org/abs/2305.06355v2	In this paper, we initiate an attempt of developing an end-to-end chat-centric video understanding system, coined as VideoChat. It integrates video foundation models and large language models via a learnable neural interface, excelling in spatiotemporal reasoning, event localization, and causal relationship inference. To instructively tune this system, we build a video-centric instruction dataset, composed of thousands of videos associated with detailed descriptions and conversations. This dataset emphasizes spatiotemporal reasoning and captures causal relationships, providing a valuable asset for training our chat-centric video understanding system. Preliminary qualitative experiments demonstrate the potential of our system across a broad spectrum of video applications, which could serve as a simple prototype system for future research on chat-centric video understanding. Access our code and data at https://github.com/OpenGVLab/Ask-Anything	true	true	Li, KunChang and He, Yinan and Wang, Yi and Li, Yizhuo and Wang, Wenhai and Luo, Ping and Wang, Yali and Wang, Limin and Qiao, Yu	2,023	null	null	null	arXiv preprint arXiv:2305.06355	VideoChat: Chat-Centric Video Understanding	VideoChat : Chat-Centric Video Understanding	https://img.shlab.org.cn/pjlab/files/2023/06/638215855649090000.pdf	by KC Li · 2023 · Cited by 853 — VideoChat is an end-to-end chat-centric video understanding system integrating video and large language models, excelling in spatiotemporal reasoning and
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	zhang2023:video	\cite{zhang2023:video}	Video-llama: An instruction-tuned audio-visual language model for video understanding	null	null	true	false	Zhang, Hang and Li, Xin and Bing, Lidong	2,023	null	null	null	arXiv preprint arXiv:2306.02858	Video-llama: An instruction-tuned audio-visual language model for video understanding	[EMNLP 2023 Demo] Video-LLaMA: An Instruction-tuned Audio ...	https://github.com/DAMO-NLP-SG/Video-LLaMA	[EMNLP 2023 Demo] Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding # Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding The following checkpoints are the full weights (visual encoder + audio encoder + Q-Formers + language decoder) to launch Video-LLaMA: Firstly, set the `llama_model` (for the path to the language decoder), `imagebind_ckpt_path` (for the path to the audio encoder), `ckpt` (for the path to VL branch) and `ckpt_2` (for the path to AL branch) in eval\_configs/video\_llama\_eval\_withaudio.yaml accordingly. The training of each cross-modal branch (i.e., VL branch or AL branch) in Video-LLaMA consists of two stages, title = {Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding}, [EMNLP 2023 Demo] Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	lu2024:unified	\cite{lu2024:unified}	Unified-IO 2: Scaling Autoregressive Multimodal Models with Vision, Language, Audio, and Action	http://arxiv.org/abs/2312.17172v1	We present Unified-IO 2, the first autoregressive multimodal model that is capable of understanding and generating image, text, audio, and action. To unify different modalities, we tokenize inputs and outputs -- images, text, audio, action, bounding boxes, etc., into a shared semantic space and then process them with a single encoder-decoder transformer model. Since training with such diverse modalities is challenging, we propose various architectural improvements to stabilize model training. We train our model from scratch on a large multimodal pre-training corpus from diverse sources with a multimodal mixture of denoisers objective. To learn an expansive set of skills, such as following multimodal instructions, we construct and finetune on an ensemble of 120 datasets with prompts and augmentations. With a single unified model, Unified-IO 2 achieves state-of-the-art performance on the GRIT benchmark and strong results in more than 35 benchmarks, including image generation and understanding, natural language understanding, video and audio understanding, and robotic manipulation. We release all our models to the research community.	true	true	Lu, Jiasen and Clark, Christopher and Lee, Sangho and Zhang, Zichen and Khosla, Savya and Marten, Ryan and Hoiem, Derek and Kembhavi, Aniruddha	2,024	null	null	null	null	Unified-IO 2: Scaling Autoregressive Multimodal Models with Vision, Language, Audio, and Action	Unified-IO 2: Scaling Autoregressive Multimodal Models with ...	https://openaccess.thecvf.com/content/CVPR2024/papers/Lu_Unified-IO_2_Scaling_Autoregressive_Multimodal_Models_with_Vision_Language_Audio_CVPR_2024_paper.pdf	by J Lu · 2024 · Cited by 210 — UNIFIED-IO 2 is a model that understands and generates image, text, audio, and action, using a single encoder-decoder model.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	achiam2023:gpt	\cite{achiam2023:gpt}	Gpt-4 technical report	null	null	true	false	Achiam, Josh and Adler, Steven and Agarwal, Sandhini and Ahmad, Lama and Akkaya, Ilge and Aleman, Florencia Leoni and Almeida, Diogo and Altenschmidt, Janko and Altman, Sam and Anadkat, Shyamal and others	2,023	null	null	null	arXiv preprint arXiv:2303.08774	Gpt-4 technical report	GPT-4 Technical Report	http://arxiv.org/pdf/2303.08774v6	We report the development of GPT-4, a large-scale, multimodal model which can accept image and text inputs and produce text outputs. While less capable than humans in many real-world scenarios, GPT-4 exhibits human-level performance on various professional and academic benchmarks, including passing a simulated bar exam with a score around the top 10% of test takers. GPT-4 is a Transformer-based model pre-trained to predict the next token in a document. The post-training alignment process results in improved performance on measures of factuality and adherence to desired behavior. A core component of this project was developing infrastructure and optimization methods that behave predictably across a wide range of scales. This allowed us to accurately predict some aspects of GPT-4's performance based on models trained with no more than 1/1,000th the compute of GPT-4.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	busso2008:iemocap	\cite{busso2008:iemocap}	IEMOCAP: Interactive emotional dyadic motion capture database	null	null	true	false	Busso, Carlos and Bulut, Murtaza and Lee, Chi-Chun and Kazemzadeh, Abe and Mower, Emily and Kim, Samuel and Chang, Jeannette N and Lee, Sungbok and Narayanan, Shrikanth S	2,008	null	null	null	Language resources and evaluation	IEMOCAP: Interactive emotional dyadic motion capture database	IEMOCAP- Home	https://sail.usc.edu/iemocap/	The Interactive Emotional Dyadic Motion Capture (IEMOCAP) database is an acted, multimodal and multispeaker database, recently collected at SAIL lab at USC.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	zadeh2018:multimodal	\cite{zadeh2018:multimodal}	Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph	null	null	true	false	Zadeh, AmirAli Bagher and Liang, Paul Pu and Poria, Soujanya and Cambria, Erik and Morency, Louis-Philippe	2,018	null	null	null	null	Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph	The MOSEI Dataset and Interpretable Dynamic Fusion	https://pliang279.github.io/papers/dap2018_mosei.pdf	by PP Liang · Cited by 30 — In this paper we introduce CMU-Multimodal Opinion. Sentiment and Emotion Intensity (CMU-. MOSEI), the largest dataset for multimodal sentiment analysis and
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	poria2019:meld	\cite{poria2019:meld}	MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations	http://arxiv.org/abs/1810.02508v6	Emotion recognition in conversations is a challenging task that has recently gained popularity due to its potential applications. Until now, however, a large-scale multimodal multi-party emotional conversational database containing more than two speakers per dialogue was missing. Thus, we propose the Multimodal EmotionLines Dataset (MELD), an extension and enhancement of EmotionLines. MELD contains about 13,000 utterances from 1,433 dialogues from the TV-series Friends. Each utterance is annotated with emotion and sentiment labels, and encompasses audio, visual and textual modalities. We propose several strong multimodal baselines and show the importance of contextual and multimodal information for emotion recognition in conversations. The full dataset is available for use at http:// affective-meld.github.io.	true	true	Poria, Soujanya and Hazarika, Devamanyu and Majumder, Navonil and Naik, Gautam and Cambria, Erik and Mihalcea, Rada	2,019	null	null	null	null	MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations	MELD: A Multimodal Multi-Party Dataset for Emotion ...	https://github.com/declare-lab/MELD	* /data/MELD/train_sent_emo.csv - contains the utterances in the training set along with Sentiment and Emotion labels. * /data/MELD/dev_sent_emo.csv - contains the utterances in the dev set along with Sentiment and Emotion labels. * /data/MELD/test_sent_emo.csv - contains the utterances in the test set along with Sentiment and Emotion labels. * /data/MELD_Dyadic/train_sent_emo_dya.csv - contains the utterances in the training set of the dyadic variant of MELD along with Sentiment and Emotion labels. * /data/MELD_Dyadic/test_sent_emo_dya.csv - contains the utterances in the test set of the dyadic variant along with Sentiment and Emotion labels. Each utterance in a dialogue has been labeled by any of these seven emotions -- Neutral, Joyful, Peaceful, Powerful, Scared, Mad and Sad. The annotations are borrowed from the original dataset.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	han2023:champagne	\cite{han2023:champagne}	CHAMPAGNE: Learning Real-world Conversation from Large-Scale Web Videos	http://arxiv.org/abs/2303.09713v2	Visual information is central to conversation: body gestures and physical behaviour, for example, contribute to meaning that transcends words alone. To date, however, most neural conversational models are limited to just text. We introduce CHAMPAGNE, a generative model of conversations that can account for visual contexts. To train CHAMPAGNE, we collect and release YTD-18M, a large-scale corpus of 18M video-based dialogues. YTD-18M is constructed from web videos: crucial to our data collection pipeline is a pretrained language model that converts error-prone automatic transcripts to a cleaner dialogue format while maintaining meaning. Human evaluation reveals that YTD-18M is more sensible and specific than prior resources (MMDialog, 1M dialogues), while maintaining visual-groundedness. Experiments demonstrate that 1) CHAMPAGNE learns to conduct conversation from YTD-18M; and 2) when fine-tuned, it achieves state-of-the-art results on four vision-language tasks focused on real-world conversations. We release data, models, and code.	true	true	Han, Seungju and Hessel, Jack and Dziri, Nouha and Choi, Yejin and Yu, Youngjae	2,023	null	null	null	null	CHAMPAGNE: Learning Real-world Conversation from Large-Scale Web Videos	[PDF] Learning Real-world Conversation from Large-Scale Web Videos	https://openaccess.thecvf.com/content/ICCV2023/papers/Han_CHAMPAGNE_Learning_Real-world_Conversation_from_Large-Scale_Web_Videos_ICCV_2023_paper.pdf	Figure 1: CHAMPAGNE is a generative model of real-world conversational frames trained on. YTD-18M, a dataset of 18M video-based dialogues.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	park2024:let	\cite{park2024:let}	Let's Go Real Talk: Spoken Dialogue Model for Face-to-Face Conversation	http://arxiv.org/abs/2406.07867v2	In this paper, we introduce a novel Face-to-Face spoken dialogue model. It processes audio-visual speech from user input and generates audio-visual speech as the response, marking the initial step towards creating an avatar chatbot system without relying on intermediate text. To this end, we newly introduce MultiDialog, the first large-scale multimodal (i.e., audio and visual) spoken dialogue corpus containing 340 hours of approximately 9,000 dialogues, recorded based on the open domain dialogue dataset, TopicalChat. The MultiDialog contains parallel audio-visual recordings of conversation partners acting according to the given script with emotion annotations, which we expect to open up research opportunities in multimodal synthesis. Our Face-to-Face spoken dialogue model incorporates a textually pretrained large language model and adapts it into the audio-visual spoken dialogue domain by incorporating speech-text joint pretraining. Through extensive experiments, we validate the effectiveness of our model in facilitating a face-to-face conversation. Demo and data are available at https://multidialog.github.io and https://huggingface.co/datasets/IVLLab/MultiDialog, respectively.	true	true	Park, Se Jin and Kim, Chae Won and Rha, Hyeongseop and Kim, Minsu and Hong, Joanna and Yeo, Jeong Hun and Ro, Yong Man	2,024	null	null	null	arXiv preprint arXiv:2406.07867	Let's Go Real Talk: Spoken Dialogue Model for Face-to-Face Conversation	Let's Go Real Talk: Spoken Dialogue Model for Face-to-Face...	https://openreview.net/forum?id=zby4Ade9CCF	In this paper, we introduce a novel Face-to-Face spoken dialogue model. It processes audio-visual speech from user input and generates audio-visual speech as
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	shafique2023:nonverbal	\cite{shafique2023:nonverbal}	Nonverbal Communication Cue Recognition: A Pathway to More Accessible Communication	null	null	true	false	Shafique, Zoya and Wang, Haiyan and Tian, Yingli	2,023	null	null	null	null	Nonverbal Communication Cue Recognition: A Pathway to More Accessible Communication	[PDF] Nonverbal Communication Cue Recognition: A Pathway to More ...	https://openaccess.thecvf.com/content/CVPR2023W/WiCV/papers/Shafique_Nonverbal_Communication_Cue_Recognition_A_Pathway_to_More_Accessible_Communication_CVPRW_2023_paper.pdf	Nonverbal communication cues (NVCs) include body language, facial expressions, and hand gestures, conveying emotions and attitudes.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	zhang2023:learning	\cite{zhang2023:learning}	Learning Emotion Representations from Verbal and Nonverbal Communication	http://arxiv.org/abs/2305.13500v1	Emotion understanding is an essential but highly challenging component of artificial general intelligence. The absence of extensively annotated datasets has significantly impeded advancements in this field. We present EmotionCLIP, the first pre-training paradigm to extract visual emotion representations from verbal and nonverbal communication using only uncurated data. Compared to numerical labels or descriptions used in previous methods, communication naturally contains emotion information. Furthermore, acquiring emotion representations from communication is more congruent with the human learning process. We guide EmotionCLIP to attend to nonverbal emotion cues through subject-aware context encoding and verbal emotion cues using sentiment-guided contrastive learning. Extensive experiments validate the effectiveness and transferability of EmotionCLIP. Using merely linear-probe evaluation protocol, EmotionCLIP outperforms the state-of-the-art supervised visual emotion recognition methods and rivals many multimodal approaches across various benchmarks. We anticipate that the advent of EmotionCLIP will address the prevailing issue of data scarcity in emotion understanding, thereby fostering progress in related domains. The code and pre-trained models are available at https://github.com/Xeaver/EmotionCLIP.	true	true	Zhang, Sitao and Pan, Yimu and Wang, James Z	2,023	null	null	null	null	Learning Emotion Representations from Verbal and Nonverbal Communication	Learning Emotion Representations from Verbal and Nonverbal Communication	http://arxiv.org/pdf/2305.13500v1	Emotion understanding is an essential but highly challenging component of artificial general intelligence. The absence of extensively annotated datasets has significantly impeded advancements in this field. We present EmotionCLIP, the first pre-training paradigm to extract visual emotion representations from verbal and nonverbal communication using only uncurated data. Compared to numerical labels or descriptions used in previous methods, communication naturally contains emotion information. Furthermore, acquiring emotion representations from communication is more congruent with the human learning process. We guide EmotionCLIP to attend to nonverbal emotion cues through subject-aware context encoding and verbal emotion cues using sentiment-guided contrastive learning. Extensive experiments validate the effectiveness and transferability of EmotionCLIP. Using merely linear-probe evaluation protocol, EmotionCLIP outperforms the state-of-the-art supervised visual emotion recognition methods and rivals many multimodal approaches across various benchmarks. We anticipate that the advent of EmotionCLIP will address the prevailing issue of data scarcity in emotion understanding, thereby fostering progress in related domains. The code and pre-trained models are available at https://github.com/Xeaver/EmotionCLIP.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	cherakara2023:furchat	\cite{cherakara2023:furchat}	FurChat: An Embodied Conversational Agent using LLMs, Combining Open and Closed-Domain Dialogue with Facial Expressions	http://arxiv.org/abs/2308.15214v2	We demonstrate an embodied conversational agent that can function as a receptionist and generate a mixture of open and closed-domain dialogue along with facial expressions, by using a large language model (LLM) to develop an engaging conversation. We deployed the system onto a Furhat robot, which is highly expressive and capable of using both verbal and nonverbal cues during interaction. The system was designed specifically for the National Robotarium to interact with visitors through natural conversations, providing them with information about the facilities, research, news, upcoming events, etc. The system utilises the state-of-the-art GPT-3.5 model to generate such information along with domain-general conversations and facial expressions based on prompt engineering.	true	true	Cherakara, Neeraj and Varghese, Finny and Shabana, Sheena and Nelson, Nivan and Karukayil, Abhiram and Kulothungan, Rohith and Farhan, Mohammed Afil and Nesset, Birthe and Moujahid, Meriam and Dinkar, Tanvi and others	2,023	null	null	null	null	FurChat: An Embodied Conversational Agent using LLMs, Combining Open and Closed-Domain Dialogue with Facial Expressions	[PDF] FurChat: An Embodied Conversational Agent using LLMs ...	https://aclanthology.org/2023.sigdial-1.55.pdf	FurChat is an embodied conversational agent using LLMs, combining open and closed-domain dialogue with facial expressions, and can function as a receptionist.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	lee2023:developing	\cite{lee2023:developing}	Developing Social Robots with Empathetic Non-Verbal Cues Using Large Language Models	http://arxiv.org/abs/2308.16529v1	We propose augmenting the empathetic capacities of social robots by integrating non-verbal cues. Our primary contribution is the design and labeling of four types of empathetic non-verbal cues, abbreviated as SAFE: Speech, Action (gesture), Facial expression, and Emotion, in a social robot. These cues are generated using a Large Language Model (LLM). We developed an LLM-based conversational system for the robot and assessed its alignment with social cues as defined by human counselors. Preliminary results show distinct patterns in the robot's responses, such as a preference for calm and positive social emotions like 'joy' and 'lively', and frequent nodding gestures. Despite these tendencies, our approach has led to the development of a social robot capable of context-aware and more authentic interactions. Our work lays the groundwork for future studies on human-robot interactions, emphasizing the essential role of both verbal and non-verbal cues in creating social and empathetic robots.	true	true	Lee, Yoon Kyung and Jung, Yoonwon and Kang, Gyuyi and Hahn, Sowon	2,023	null	null	null	arXiv preprint arXiv:2308.16529	Developing Social Robots with Empathetic Non-Verbal Cues Using Large Language Models	Developing Social Robots with Empathetic Non-Verbal Cues Using ...	https://www.researchgate.net/publication/373552152_Developing_Social_Robots_with_Empathetic_Non-Verbal_Cues_Using_Large_Language_Models	We developed an LLM-based conversational system for the robot and assessed its alignment with social cues as defined by human counselors. Preliminary results
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	lin2023:one	\cite{lin2023:one}	One-Stage 3D Whole-Body Mesh Recovery with Component Aware Transformer	http://arxiv.org/abs/2303.16160v1	Whole-body mesh recovery aims to estimate the 3D human body, face, and hands parameters from a single image. It is challenging to perform this task with a single network due to resolution issues, i.e., the face and hands are usually located in extremely small regions. Existing works usually detect hands and faces, enlarge their resolution to feed in a specific network to predict the parameter, and finally fuse the results. While this copy-paste pipeline can capture the fine-grained details of the face and hands, the connections between different parts cannot be easily recovered in late fusion, leading to implausible 3D rotation and unnatural pose. In this work, we propose a one-stage pipeline for expressive whole-body mesh recovery, named OSX, without separate networks for each part. Specifically, we design a Component Aware Transformer (CAT) composed of a global body encoder and a local face/hand decoder. The encoder predicts the body parameters and provides a high-quality feature map for the decoder, which performs a feature-level upsample-crop scheme to extract high-resolution part-specific features and adopt keypoint-guided deformable attention to estimate hand and face precisely. The whole pipeline is simple yet effective without any manual post-processing and naturally avoids implausible prediction. Comprehensive experiments demonstrate the effectiveness of OSX. Lastly, we build a large-scale Upper-Body dataset (UBody) with high-quality 2D and 3D whole-body annotations. It contains persons with partially visible bodies in diverse real-life scenarios to bridge the gap between the basic task and downstream applications.	true	true	Lin, Jing and Zeng, Ailing and Wang, Haoqian and Zhang, Lei and Li, Yu	2,023	null	null	null	null	One-Stage 3D Whole-Body Mesh Recovery with Component Aware Transformer	IDEA-Research/OSX - GitHub	https://github.com/IDEA-Research/OSX	This repo is official PyTorch implementation of One-Stage 3D Whole-Body Mesh Recovery with Component Aware Transformer (CVPR2023). We propose the first one-
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	dwivedi2024:tokenhmr	\cite{dwivedi2024:tokenhmr}	TokenHMR: Advancing Human Mesh Recovery with a Tokenized Pose Representation	http://arxiv.org/abs/2404.16752v1	We address the problem of regressing 3D human pose and shape from a single image, with a focus on 3D accuracy. The current best methods leverage large datasets of 3D pseudo-ground-truth (p-GT) and 2D keypoints, leading to robust performance. With such methods, we observe a paradoxical decline in 3D pose accuracy with increasing 2D accuracy. This is caused by biases in the p-GT and the use of an approximate camera projection model. We quantify the error induced by current camera models and show that fitting 2D keypoints and p-GT accurately causes incorrect 3D poses. Our analysis defines the invalid distances within which minimizing 2D and p-GT losses is detrimental. We use this to formulate a new loss Threshold-Adaptive Loss Scaling (TALS) that penalizes gross 2D and p-GT losses but not smaller ones. With such a loss, there are many 3D poses that could equally explain the 2D evidence. To reduce this ambiguity we need a prior over valid human poses but such priors can introduce unwanted bias. To address this, we exploit a tokenized representation of human pose and reformulate the problem as token prediction. This restricts the estimated poses to the space of valid poses, effectively providing a uniform prior. Extensive experiments on the EMDB and 3DPW datasets show that our reformulated keypoint loss and tokenization allows us to train on in-the-wild data while improving 3D accuracy over the state-of-the-art. Our models and code are available for research at https://tokenhmr.is.tue.mpg.de.	true	true	Dwivedi, Sai Kumar and Sun, Yu and Patel, Priyanka and Feng, Yao and Black, Michael J	2,024	null	null	null	null	TokenHMR: Advancing Human Mesh Recovery with a Tokenized Pose Representation	TokenHMR: Advancing Human Mesh Recovery with a ...	https://github.com/saidwivedi/TokenHMR	Our method has two stages: Tokenization: The encoder maps continuous poses to discrete pose tokens. TokenHMR: During the training of human pose
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	danvevcek2022emoca	\cite{danvevcek2022emoca}	EMOCA: Emotion Driven Monocular Face Capture and Animation	http://arxiv.org/abs/2204.11312v1	As 3D facial avatars become more widely used for communication, it is critical that they faithfully convey emotion. Unfortunately, the best recent methods that regress parametric 3D face models from monocular images are unable to capture the full spectrum of facial expression, such as subtle or extreme emotions. We find the standard reconstruction metrics used for training (landmark reprojection error, photometric error, and face recognition loss) are insufficient to capture high-fidelity expressions. The result is facial geometries that do not match the emotional content of the input image. We address this with EMOCA (EMOtion Capture and Animation), by introducing a novel deep perceptual emotion consistency loss during training, which helps ensure that the reconstructed 3D expression matches the expression depicted in the input image. While EMOCA achieves 3D reconstruction errors that are on par with the current best methods, it significantly outperforms them in terms of the quality of the reconstructed expression and the perceived emotional content. We also directly regress levels of valence and arousal and classify basic expressions from the estimated 3D face parameters. On the task of in-the-wild emotion recognition, our purely geometric approach is on par with the best image-based methods, highlighting the value of 3D geometry in analyzing human behavior. The model and code are publicly available at https://emoca.is.tue.mpg.de.	true	true	Dan{\v{e}}{\v{c}}ek, Radek and Black, Michael J and Bolkart, Timo	2,022	null	null	null	null	EMOCA: Emotion Driven Monocular Face Capture and Animation	EMOCA: Emotion Driven Monocular Face Capture and Animation	http://arxiv.org/pdf/2204.11312v1	As 3D facial avatars become more widely used for communication, it is critical that they faithfully convey emotion. Unfortunately, the best recent methods that regress parametric 3D face models from monocular images are unable to capture the full spectrum of facial expression, such as subtle or extreme emotions. We find the standard reconstruction metrics used for training (landmark reprojection error, photometric error, and face recognition loss) are insufficient to capture high-fidelity expressions. The result is facial geometries that do not match the emotional content of the input image. We address this with EMOCA (EMOtion Capture and Animation), by introducing a novel deep perceptual emotion consistency loss during training, which helps ensure that the reconstructed 3D expression matches the expression depicted in the input image. While EMOCA achieves 3D reconstruction errors that are on par with the current best methods, it significantly outperforms them in terms of the quality of the reconstructed expression and the perceived emotional content. We also directly regress levels of valence and arousal and classify basic expressions from the estimated 3D face parameters. On the task of in-the-wild emotion recognition, our purely geometric approach is on par with the best image-based methods, highlighting the value of 3D geometry in analyzing human behavior. The model and code are publicly available at https://emoca.is.tue.mpg.de.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	yi2023:generating	\cite{yi2023:generating}	Generating Holistic 3D Human Motion from Speech	http://arxiv.org/abs/2212.04420v2	This work addresses the problem of generating 3D holistic body motions from human speech. Given a speech recording, we synthesize sequences of 3D body poses, hand gestures, and facial expressions that are realistic and diverse. To achieve this, we first build a high-quality dataset of 3D holistic body meshes with synchronous speech. We then define a novel speech-to-motion generation framework in which the face, body, and hands are modeled separately. The separated modeling stems from the fact that face articulation strongly correlates with human speech, while body poses and hand gestures are less correlated. Specifically, we employ an autoencoder for face motions, and a compositional vector-quantized variational autoencoder (VQ-VAE) for the body and hand motions. The compositional VQ-VAE is key to generating diverse results. Additionally, we propose a cross-conditional autoregressive model that generates body poses and hand gestures, leading to coherent and realistic motions. Extensive experiments and user studies demonstrate that our proposed approach achieves state-of-the-art performance both qualitatively and quantitatively. Our novel dataset and code will be released for research purposes at https://talkshow.is.tue.mpg.de.	true	true	Yi, Hongwei and Liang, Hualin and Liu, Yifei and Cao, Qiong and Wen, Yandong and Bolkart, Timo and Tao, Dacheng and Black, Michael J	2,023	null	null	null	null	Generating Holistic 3D Human Motion from Speech	Generating Holistic 3D Human Motion from Speech	http://arxiv.org/pdf/2212.04420v2	This work addresses the problem of generating 3D holistic body motions from human speech. Given a speech recording, we synthesize sequences of 3D body poses, hand gestures, and facial expressions that are realistic and diverse. To achieve this, we first build a high-quality dataset of 3D holistic body meshes with synchronous speech. We then define a novel speech-to-motion generation framework in which the face, body, and hands are modeled separately. The separated modeling stems from the fact that face articulation strongly correlates with human speech, while body poses and hand gestures are less correlated. Specifically, we employ an autoencoder for face motions, and a compositional vector-quantized variational autoencoder (VQ-VAE) for the body and hand motions. The compositional VQ-VAE is key to generating diverse results. Additionally, we propose a cross-conditional autoregressive model that generates body poses and hand gestures, leading to coherent and realistic motions. Extensive experiments and user studies demonstrate that our proposed approach achieves state-of-the-art performance both qualitatively and quantitatively. Our novel dataset and code will be released for research purposes at https://talkshow.is.tue.mpg.de.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	wu2024:motionllm	\cite{wu2024:motionllm}	MotionLLM: Multimodal Motion-Language Learning with Large Language Models	null	null	true	false	Wu, Qi and Zhao, Yubo and Wang, Yifan and Tai, Yu-Wing and Tang, Chi-Keung	2,024	null	null	null	arXiv preprint arXiv:2405.17013	MotionLLM: Multimodal Motion-Language Learning with Large Language Models	(PDF) MotionLLM: Multimodal Motion-Language Learning ...	https://www.researchgate.net/publication/380906869_MotionLLM_Multimodal_Motion-Language_Learning_with_Large_Language_Models	MotionGPT-2 accommodates multiple motion-relevant tasks and supporting multimodal control conditions through pre-trained Large Language Models (
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	lu2023:humantomato	\cite{lu2023:humantomato}	HumanTOMATO: Text-aligned Whole-body Motion Generation	http://arxiv.org/abs/2310.12978v1	This work targets a novel text-driven whole-body motion generation task, which takes a given textual description as input and aims at generating high-quality, diverse, and coherent facial expressions, hand gestures, and body motions simultaneously. Previous works on text-driven motion generation tasks mainly have two limitations: they ignore the key role of fine-grained hand and face controlling in vivid whole-body motion generation, and lack a good alignment between text and motion. To address such limitations, we propose a Text-aligned whOle-body Motion generATiOn framework, named HumanTOMATO, which is the first attempt to our knowledge towards applicable holistic motion generation in this research area. To tackle this challenging task, our solution includes two key designs: (1) a Holistic Hierarchical VQ-VAE (aka H$^2$VQ) and a Hierarchical-GPT for fine-grained body and hand motion reconstruction and generation with two structured codebooks; and (2) a pre-trained text-motion-alignment model to help generated motion align with the input textual description explicitly. Comprehensive experiments verify that our model has significant advantages in both the quality of generated motions and their alignment with text.	true	true	Lu, Shunlin and Chen, Ling-Hao and Zeng, Ailing and Lin, Jing and Zhang, Ruimao and Zhang, Lei and Shum, Heung-Yeung	2,023	null	null	null	arXiv preprint arXiv:2310.12978	HumanTOMATO: Text-aligned Whole-body Motion Generation	HumanTOMATO: Text-aligned Whole-body Motion ...	https://lhchen.top/HumanTOMATO/	The proposed HumanTOMATO model can generate text-aligned whole-body motions with vivid and harmonious face, hand, and body motion.
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	ng2023:can	\cite{ng2023:can}	Can Language Models Learn to Listen?	http://arxiv.org/abs/2308.10897v1	We present a framework for generating appropriate facial responses from a listener in dyadic social interactions based on the speaker's words. Given an input transcription of the speaker's words with their timestamps, our approach autoregressively predicts a response of a listener: a sequence of listener facial gestures, quantized using a VQ-VAE. Since gesture is a language component, we propose treating the quantized atomic motion elements as additional language token inputs to a transformer-based large language model. Initializing our transformer with the weights of a language model pre-trained only on text results in significantly higher quality listener responses than training a transformer from scratch. We show that our generated listener motion is fluent and reflective of language semantics through quantitative metrics and a qualitative user study. In our evaluation, we analyze the model's ability to utilize temporal and semantic aspects of spoken text. Project page: https://people.eecs.berkeley.edu/~evonne_ng/projects/text2listen/	true	true	Ng, Evonne and Subramanian, Sanjay and Klein, Dan and Kanazawa, Angjoo and Darrell, Trevor and Ginosar, Shiry	2,023	null	null	null	null	Can Language Models Learn to Listen?	Can Language Models Learn to Listen?	http://arxiv.org/pdf/2308.10897v1	We present a framework for generating appropriate facial responses from a listener in dyadic social interactions based on the speaker's words. Given an input transcription of the speaker's words with their timestamps, our approach autoregressively predicts a response of a listener: a sequence of listener facial gestures, quantized using a VQ-VAE. Since gesture is a language component, we propose treating the quantized atomic motion elements as additional language token inputs to a transformer-based large language model. Initializing our transformer with the weights of a language model pre-trained only on text results in significantly higher quality listener responses than training a transformer from scratch. We show that our generated listener motion is fluent and reflective of language semantics through quantitative metrics and a qualitative user study. In our evaluation, we analyze the model's ability to utilize temporal and semantic aspects of spoken text. Project page: https://people.eecs.berkeley.edu/~evonne_ng/projects/text2listen/
Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues	2506.00958v1	ng2022:learning	\cite{ng2022:learning}	Learning to Listen: Modeling Non-Deterministic Dyadic Facial Motion	http://arxiv.org/abs/2204.08451v1	We present a framework for modeling interactional communication in dyadic conversations: given multimodal inputs of a speaker, we autoregressively output multiple possibilities of corresponding listener motion. We combine the motion and speech audio of the speaker using a motion-audio cross attention transformer. Furthermore, we enable non-deterministic prediction by learning a discrete latent representation of realistic listener motion with a novel motion-encoding VQ-VAE. Our method organically captures the multimodal and non-deterministic nature of nonverbal dyadic interactions. Moreover, it produces realistic 3D listener facial motion synchronous with the speaker (see video). We demonstrate that our method outperforms baselines qualitatively and quantitatively via a rich suite of experiments. To facilitate this line of research, we introduce a novel and large in-the-wild dataset of dyadic conversations. Code, data, and videos available at https://evonneng.github.io/learning2listen/.	true	true	Ng, Evonne and Joo, Hanbyul and Hu, Liwen and Li, Hao and Darrell, Trevor and Kanazawa, Angjoo and Ginosar, Shiry	2,022	null	null	null	null	Learning to Listen: Modeling Non-Deterministic Dyadic Facial Motion	[PDF] Learning To Listen: Modeling Non-Deterministic Dyadic Facial Motion	https://openaccess.thecvf.com/content/CVPR2022/papers/Ng_Learning_To_Listen_Modeling_Non-Deterministic_Dyadic_Facial_Motion_CVPR_2022_paper.pdf	The method synthesizes listener motion from speaker video using a motion-audio transformer and a VQ-VAE, outputting multiple possibilities of listener motion.
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	strom2006expressive	\cite{strom2006expressive}	Expressive prosody for unit-selection speech synthesis.	null	null	true	false	Strom, Volker and Clark, Robert AJ and King, Simon	2,006	null	null	null	null	Expressive prosody for unit-selection speech synthesis.	Expressive Prosody for Unit-selection Speech Synthesis - CSTR	https://www.cstr.ed.ac.uk/downloads/publications/2006/strom06.pdf	by V Strom · Cited by 42 — The Festival unit selection speech synthesis system, Multisyn [1], achieves highly natural synthetic speech by avoiding use of an ex- plicit model of prosody in
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	ren2019fastspeech	\cite{ren2019fastspeech}	FastSpeech: Fast, Robust and Controllable Text to Speech	http://arxiv.org/abs/1905.09263v5	Neural network based end-to-end text to speech (TTS) has significantly improved the quality of synthesized speech. Prominent methods (e.g., Tacotron 2) usually first generate mel-spectrogram from text, and then synthesize speech from the mel-spectrogram using vocoder such as WaveNet. Compared with traditional concatenative and statistical parametric approaches, neural network based end-to-end models suffer from slow inference speed, and the synthesized speech is usually not robust (i.e., some words are skipped or repeated) and lack of controllability (voice speed or prosody control). In this work, we propose a novel feed-forward network based on Transformer to generate mel-spectrogram in parallel for TTS. Specifically, we extract attention alignments from an encoder-decoder based teacher model for phoneme duration prediction, which is used by a length regulator to expand the source phoneme sequence to match the length of the target mel-spectrogram sequence for parallel mel-spectrogram generation. Experiments on the LJSpeech dataset show that our parallel model matches autoregressive models in terms of speech quality, nearly eliminates the problem of word skipping and repeating in particularly hard cases, and can adjust voice speed smoothly. Most importantly, compared with autoregressive Transformer TTS, our model speeds up mel-spectrogram generation by 270x and the end-to-end speech synthesis by 38x. Therefore, we call our model FastSpeech.	true	true	Ren, Yi and Ruan, Yangjun and Tan, Xu and Qin, Tao and Zhao, Sheng and Zhao, Zhou and Liu, Tie-Yan	2,019	null	null	null	Advances in neural information processing systems	FastSpeech: Fast, Robust and Controllable Text to Speech	FastSpeech: Fast, Robust and Controllable Text to Speech	http://arxiv.org/pdf/1905.09263v5	Neural network based end-to-end text to speech (TTS) has significantly improved the quality of synthesized speech. Prominent methods (e.g., Tacotron 2) usually first generate mel-spectrogram from text, and then synthesize speech from the mel-spectrogram using vocoder such as WaveNet. Compared with traditional concatenative and statistical parametric approaches, neural network based end-to-end models suffer from slow inference speed, and the synthesized speech is usually not robust (i.e., some words are skipped or repeated) and lack of controllability (voice speed or prosody control). In this work, we propose a novel feed-forward network based on Transformer to generate mel-spectrogram in parallel for TTS. Specifically, we extract attention alignments from an encoder-decoder based teacher model for phoneme duration prediction, which is used by a length regulator to expand the source phoneme sequence to match the length of the target mel-spectrogram sequence for parallel mel-spectrogram generation. Experiments on the LJSpeech dataset show that our parallel model matches autoregressive models in terms of speech quality, nearly eliminates the problem of word skipping and repeating in particularly hard cases, and can adjust voice speed smoothly. Most importantly, compared with autoregressive Transformer TTS, our model speeds up mel-spectrogram generation by 270x and the end-to-end speech synthesis by 38x. Therefore, we call our model FastSpeech.
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	ren2020fastspeech	\cite{ren2020fastspeech}	FastSpeech 2: Fast and High-Quality End-to-End Text to Speech	http://arxiv.org/abs/2006.04558v8	Non-autoregressive text to speech (TTS) models such as FastSpeech can synthesize speech significantly faster than previous autoregressive models with comparable quality. The training of FastSpeech model relies on an autoregressive teacher model for duration prediction (to provide more information as input) and knowledge distillation (to simplify the data distribution in output), which can ease the one-to-many mapping problem (i.e., multiple speech variations correspond to the same text) in TTS. However, FastSpeech has several disadvantages: 1) the teacher-student distillation pipeline is complicated and time-consuming, 2) the duration extracted from the teacher model is not accurate enough, and the target mel-spectrograms distilled from teacher model suffer from information loss due to data simplification, both of which limit the voice quality. In this paper, we propose FastSpeech 2, which addresses the issues in FastSpeech and better solves the one-to-many mapping problem in TTS by 1) directly training the model with ground-truth target instead of the simplified output from teacher, and 2) introducing more variation information of speech (e.g., pitch, energy and more accurate duration) as conditional inputs. Specifically, we extract duration, pitch and energy from speech waveform and directly take them as conditional inputs in training and use predicted values in inference. We further design FastSpeech 2s, which is the first attempt to directly generate speech waveform from text in parallel, enjoying the benefit of fully end-to-end inference. Experimental results show that 1) FastSpeech 2 achieves a 3x training speed-up over FastSpeech, and FastSpeech 2s enjoys even faster inference speed; 2) FastSpeech 2 and 2s outperform FastSpeech in voice quality, and FastSpeech 2 can even surpass autoregressive models. Audio samples are available at https://speechresearch.github.io/fastspeech2/.	true	true	Ren, Yi and Hu, Chenxu and Tan, Xu and Qin, Tao and Zhao, Sheng and Zhao, Zhou and Liu, Tie-Yan	2,020	null	null	null	arXiv preprint arXiv:2006.04558	FastSpeech 2: Fast and High-Quality End-to-End Text to Speech	FastSpeech 2: Fast and High-Quality End-to-End Text to Speech	https://www.microsoft.com/en-us/research/lab/microsoft-research-asia/articles/fastspeech-2-fast-and-high-quality-end-to-end-text-to-speech/	FastSpeech 2 outperforms FastSpeech in voice quality and enjoys a much simpler training pipeline (3x training time reduction) while inheriting its advantages.
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	mohan2021ctrl	\cite{mohan2021ctrl}	Ctrl-P: Temporal control of prosodic variation for speech synthesis	null	null	true	false	Mohan, Devang S Ram and Hu, Vivian and Teh, Tian Huey and Torresquintero, Alexandra and Wallis, Christopher GR and Staib, Marlene and Foglianti, Lorenzo and Gao, Jiameng and King, Simon	2,021	null	null	null	arXiv preprint arXiv:2106.08352	Ctrl-P: Temporal control of prosodic variation for speech synthesis	Ctrl-P: Temporal Control of Prosodic Variation for Speech Synthesis	http://arxiv.org/pdf/2106.08352v1	Text does not fully specify the spoken form, so text-to-speech models must be able to learn from speech data that vary in ways not explained by the corresponding text. One way to reduce the amount of unexplained variation in training data is to provide acoustic information as an additional learning signal. When generating speech, modifying this acoustic information enables multiple distinct renditions of a text to be produced. Since much of the unexplained variation is in the prosody, we propose a model that generates speech explicitly conditioned on the three primary acoustic correlates of prosody: $F_{0}$, energy and duration. The model is flexible about how the values of these features are specified: they can be externally provided, or predicted from text, or predicted then subsequently modified. Compared to a model that employs a variational auto-encoder to learn unsupervised latent features, our model provides more interpretable, temporally-precise, and disentangled control. When automatically predicting the acoustic features from text, it generates speech that is more natural than that from a Tacotron 2 model with reference encoder. Subsequent human-in-the-loop modification of the predicted acoustic features can significantly further increase naturalness.
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	bandekar2023speaking	\cite{bandekar2023speaking}	Speaking rate attention-based duration prediction for speed control TTS	http://arxiv.org/abs/2310.08846v1	With the advent of high-quality speech synthesis, there is a lot of interest in controlling various prosodic attributes of speech. Speaking rate is an essential attribute towards modelling the expressivity of speech. In this work, we propose a novel approach to control the speaking rate for non-autoregressive TTS. We achieve this by conditioning the speaking rate inside the duration predictor, allowing implicit speaking rate control. We show the benefits of this approach by synthesising audio at various speaking rate factors and measuring the quality of speaking rate-controlled synthesised speech. Further, we study the effect of the speaking rate distribution of the training data towards effective rate control. Finally, we fine-tune a baseline pretrained TTS model to obtain speaking rate control TTS. We provide various analyses to showcase the benefits of using this proposed approach, along with objective as well as subjective metrics. We find that the proposed methods have higher subjective scores and lower speaker rate errors across many speaking rate factors over the baseline.	true	true	Bandekar, Jesuraj and Udupa, Sathvik and Singh, Abhayjeet and Jayakumar, Anjali and Badiger, Sandhya and Kumar, Saurabh and VH, Pooja and Ghosh, Prasanta Kumar and others	2,023	null	null	null	arXiv preprint arXiv:2310.08846	Speaking rate attention-based duration prediction for speed control TTS	Speaking Rate Control of end-to-end TTS Models by Direct ...	https://www.isca-archive.org/interspeech_2022/lenglet22_interspeech.pdf	by M Lenglet · 2022 · Cited by 8 — Evaluation was performed on the control of speaking rate on both attention-based (TC) and duration predictor based (FS) methods. Objective analyses showed
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	wang2018style	\cite{wang2018style}	Style Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech Synthesis	http://arxiv.org/abs/1803.09017v1	In this work, we propose "global style tokens" (GSTs), a bank of embeddings that are jointly trained within Tacotron, a state-of-the-art end-to-end speech synthesis system. The embeddings are trained with no explicit labels, yet learn to model a large range of acoustic expressiveness. GSTs lead to a rich set of significant results. The soft interpretable "labels" they generate can be used to control synthesis in novel ways, such as varying speed and speaking style - independently of the text content. They can also be used for style transfer, replicating the speaking style of a single audio clip across an entire long-form text corpus. When trained on noisy, unlabeled found data, GSTs learn to factorize noise and speaker identity, providing a path towards highly scalable but robust speech synthesis.	true	true	Wang, Yuxuan and Stanton, Daisy and Zhang, Yu and Ryan, RJ-Skerry and Battenberg, Eric and Shor, Joel and Xiao, Ying and Jia, Ye and Ren, Fei and Saurous, Rif A	2,018	null	null	null	null	Style Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech Synthesis	Unsupervised Style Modeling, Control and Transfer in End- ...	https://research.google/pubs/style-tokens-unsupervised-style-modeling-control-and-transfer-in-end-to-end-speech-synthesis/	by Y Wang · Cited by 1080 — In this work, we propose “global style tokens”(GSTs), a bank of embeddings that are jointly trained within Tacotron, a state-of-the-art end-to-end speech
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	skerry2018towards	\cite{skerry2018towards}	Towards End-to-End Prosody Transfer for Expressive Speech Synthesis with Tacotron	http://arxiv.org/abs/1803.09047v1	We present an extension to the Tacotron speech synthesis architecture that learns a latent embedding space of prosody, derived from a reference acoustic representation containing the desired prosody. We show that conditioning Tacotron on this learned embedding space results in synthesized audio that matches the prosody of the reference signal with fine time detail even when the reference and synthesis speakers are different. Additionally, we show that a reference prosody embedding can be used to synthesize text that is different from that of the reference utterance. We define several quantitative and subjective metrics for evaluating prosody transfer, and report results with accompanying audio samples from single-speaker and 44-speaker Tacotron models on a prosody transfer task.	true	true	Skerry-Ryan, RJ and Battenberg, Eric and Xiao, Ying and Wang, Yuxuan and Stanton, Daisy and Shor, Joel and Weiss, Ron and Clark, Rob and Saurous, Rif A	2,018	null	null	null	null	Towards End-to-End Prosody Transfer for Expressive Speech Synthesis with Tacotron	[PDF] Towards End-to-End Prosody Transfer for Expressive Speech ...	https://proceedings.mlr.press/v80/skerry-ryan18a/skerry-ryan18a.pdf	Abstract. We present an extension to the Tacotron speech synthesis architecture that learns a latent embed- ding space of prosody, derived from a reference.
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	hsu2018hierarchical	\cite{hsu2018hierarchical}	Hierarchical Generative Modeling for Controllable Speech Synthesis	http://arxiv.org/abs/1810.07217v2	This paper proposes a neural sequence-to-sequence text-to-speech (TTS) model which can control latent attributes in the generated speech that are rarely annotated in the training data, such as speaking style, accent, background noise, and recording conditions. The model is formulated as a conditional generative model based on the variational autoencoder (VAE) framework, with two levels of hierarchical latent variables. The first level is a categorical variable, which represents attribute groups (e.g. clean/noisy) and provides interpretability. The second level, conditioned on the first, is a multivariate Gaussian variable, which characterizes specific attribute configurations (e.g. noise level, speaking rate) and enables disentangled fine-grained control over these attributes. This amounts to using a Gaussian mixture model (GMM) for the latent distribution. Extensive evaluation demonstrates its ability to control the aforementioned attributes. In particular, we train a high-quality controllable TTS model on real found data, which is capable of inferring speaker and style attributes from a noisy utterance and use it to synthesize clean speech with controllable speaking style.	true	true	Hsu, Wei-Ning and Zhang, Yu and Weiss, Ron J and Zen, Heiga and Wu, Yonghui and Wang, Yuxuan and Cao, Yuan and Jia, Ye and Chen, Zhifeng and Shen, Jonathan and others	2,018	null	null	null	arXiv preprint arXiv:1810.07217	Hierarchical Generative Modeling for Controllable Speech Synthesis	Hierarchical Generative Modeling for Controllable Speech Synthesis	http://arxiv.org/pdf/1810.07217v2	This paper proposes a neural sequence-to-sequence text-to-speech (TTS) model which can control latent attributes in the generated speech that are rarely annotated in the training data, such as speaking style, accent, background noise, and recording conditions. The model is formulated as a conditional generative model based on the variational autoencoder (VAE) framework, with two levels of hierarchical latent variables. The first level is a categorical variable, which represents attribute groups (e.g. clean/noisy) and provides interpretability. The second level, conditioned on the first, is a multivariate Gaussian variable, which characterizes specific attribute configurations (e.g. noise level, speaking rate) and enables disentangled fine-grained control over these attributes. This amounts to using a Gaussian mixture model (GMM) for the latent distribution. Extensive evaluation demonstrates its ability to control the aforementioned attributes. In particular, we train a high-quality controllable TTS model on real found data, which is capable of inferring speaker and style attributes from a noisy utterance and use it to synthesize clean speech with controllable speaking style.
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	lenglet2022speaking	\cite{lenglet2022speaking}	Speaking Rate Control of end-to-end TTS Models by Direct Manipulation of the Encoder's Output Embeddings	null	null	true	false	Lenglet, Martin and Perrotin, Olivier and Bailly, G{\'e}rard	2,022	null	null	null	null	Speaking Rate Control of end-to-end TTS Models by Direct Manipulation of the Encoder's Output Embeddings	Speaking Rate Control of end-to-end TTS Models by ... - ISCA Archive	https://www.isca-archive.org/interspeech_2022/lenglet22_interspeech.html	Experimental results show that the control provided by embeddings reproduces a behaviour closer to natural speech data.
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	zhang2020unified	\cite{zhang2020unified}	Unified Mandarin TTS Front-end Based on Distilled BERT Model	http://arxiv.org/abs/2012.15404v1	The front-end module in a typical Mandarin text-to-speech system (TTS) is composed of a long pipeline of text processing components, which requires extensive efforts to build and is prone to large accumulative model size and cascade errors. In this paper, a pre-trained language model (PLM) based model is proposed to simultaneously tackle the two most important tasks in TTS front-end, i.e., prosodic structure prediction (PSP) and grapheme-to-phoneme (G2P) conversion. We use a pre-trained Chinese BERT[1] as the text encoder and employ multi-task learning technique to adapt it to the two TTS front-end tasks. Then, the BERT encoder is distilled into a smaller model by employing a knowledge distillation technique called TinyBERT[2], making the whole model size 25% of that of benchmark pipeline models while maintaining competitive performance on both tasks. With the proposed the methods, we are able to run the whole TTS front-end module in a light and unified manner, which is more friendly to deployment on mobile devices.	true	true	Zhang, Yang and Deng, Liqun and Wang, Yasheng	2,020	null	null	null	arXiv preprint arXiv:2012.15404	Unified Mandarin TTS Front-end Based on Distilled BERT Model	Unified Mandarin TTS Front-end Based on Distilled BERT Model	https://arxiv.org/abs/2012.15404	We use a pre-trained Chinese BERT[1] as the text encoder and employ multi-task learning technique to adapt it to the two TTS front-end tasks.
Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models	2506.00832v1	fong2022speech	\cite{fong2022speech}	Speech Audio Corrector: using speech from non-target speakers for one-off correction of mispronunciations in grapheme-input text-to-speech	null	null	true	false	Fong, Jason and Lyth, Daniel and Henter, Gustav Eje and Tang, Hao and King, Simon	2,022	null	null	null	null	Speech Audio Corrector: using speech from non-target speakers for one-off correction of mispronunciations in grapheme-input text-to-speech	[PDF] using speech from non-target speakers for one-off correction of ...	https://www.research.ed.ac.uk/files/364801102/Speech_Audio_Corrector_FONG_DOA13062022_VOR.pdf	Missing: 04/08/2025
Dual Debiasing for Noisy In-Context Learning for Text Generation	2506.00418v1	yoo2022ground	\cite{yoo2022ground}	Ground-Truth Labels Matter: A Deeper Look into Input-Label Demonstrations	http://arxiv.org/abs/2205.12685v2	Despite recent explosion of interests in in-context learning, the underlying mechanism and the precise impact of the quality of demonstrations remain elusive. Intuitively, ground-truth labels should have as much impact in in-context learning (ICL) as supervised learning, but recent work reported that the input-label correspondence is significantly less important than previously thought. Intrigued by this counter-intuitive observation, we re-examine the importance of ground-truth labels in in-context learning. With the introduction of two novel metrics, namely Label-Correctness Sensitivity and Ground-truth Label Effect Ratio (GLER), we were able to conduct quantifiable analysis on the impact of ground-truth label demonstrations. Through extensive analyses, we find that the correct input-label mappings can have varying impacts on the downstream in-context learning performances, depending on the experimental configuration. Through additional studies, we identify key components, such as the verbosity of prompt templates and the language model size, as the controlling factor to achieve more noise-resilient ICL.	true	true	Yoo, Kang Min and Kim, Junyeob and Kim, Hyuhng Joon and Cho, Hyunsoo and Jo, Hwiyeol and Lee, Sang-Woo and Lee, Sang-goo and Kim, Taeuk	2,022	null	null	null	null	Ground-Truth Labels Matter: A Deeper Look into Input-Label Demonstrations	Ground-Truth Labels Matter: A Deeper Look into Input- ...	https://aclanthology.org/2022.emnlp-main.155.pdf	by KM Yoo · 2022 · Cited by 100 — We propose two new quantifiable metrics, sen- sitivity and GLER, to measure the impact of ground-truth label demonstrations on ICL. • We conduct
Dual Debiasing for Noisy In-Context Learning for Text Generation	2506.00418v1	o2023contrastive	\cite{o2023contrastive}	Contrastive Decoding Improves Reasoning in Large Language Models	http://arxiv.org/abs/2309.09117v2	We demonstrate that Contrastive Decoding -- a simple, computationally light, and training-free text generation method proposed by Li et al 2022 -- achieves large out-of-the-box improvements over greedy decoding on a variety of reasoning tasks. Originally shown to improve the perceived quality of long-form text generation, Contrastive Decoding searches for strings that maximize a weighted difference in likelihood between strong and weak models. We show that Contrastive Decoding leads LLaMA-65B to outperform LLaMA 2, GPT-3.5 and PaLM 2-L on the HellaSwag commonsense reasoning benchmark, and to outperform LLaMA 2, GPT-3.5 and PaLM-540B on the GSM8K math word reasoning benchmark, in addition to improvements on a collection of other tasks. Analysis suggests that Contrastive Decoding improves over existing methods by preventing some abstract reasoning errors, as well as by avoiding simpler modes such as copying sections of the input during chain-of-thought. Overall, Contrastive Decoding outperforms nucleus sampling for long-form generation and greedy decoding for reasoning tasks, making it a powerful general purpose method for generating text from language models.	true	true	O'Brien, Sean and Lewis, Mike	2,023	null	null	null	arXiv preprint arXiv:2309.09117	Contrastive Decoding Improves Reasoning in Large Language Models	Contrastive Decoding Improves Reasoning in Large Language Models	http://arxiv.org/pdf/2309.09117v2	We demonstrate that Contrastive Decoding -- a simple, computationally light, and training-free text generation method proposed by Li et al 2022 -- achieves large out-of-the-box improvements over greedy decoding on a variety of reasoning tasks. Originally shown to improve the perceived quality of long-form text generation, Contrastive Decoding searches for strings that maximize a weighted difference in likelihood between strong and weak models. We show that Contrastive Decoding leads LLaMA-65B to outperform LLaMA 2, GPT-3.5 and PaLM 2-L on the HellaSwag commonsense reasoning benchmark, and to outperform LLaMA 2, GPT-3.5 and PaLM-540B on the GSM8K math word reasoning benchmark, in addition to improvements on a collection of other tasks. Analysis suggests that Contrastive Decoding improves over existing methods by preventing some abstract reasoning errors, as well as by avoiding simpler modes such as copying sections of the input during chain-of-thought. Overall, Contrastive Decoding outperforms nucleus sampling for long-form generation and greedy decoding for reasoning tasks, making it a powerful general purpose method for generating text from language models.
Dual Debiasing for Noisy In-Context Learning for Text Generation	2506.00418v1	li2023unified	\cite{li2023unified}	Unified Demonstration Retriever for In-Context Learning	http://arxiv.org/abs/2305.04320v2	In-context learning is a new learning paradigm where a language model conditions on a few input-output pairs (demonstrations) and a test input, and directly outputs the prediction. It has been shown highly dependent on the provided demonstrations and thus promotes the research of demonstration retrieval: given a test input, relevant examples are retrieved from the training set to serve as informative demonstrations for in-context learning. While previous works focus on training task-specific retrievers for several tasks separately, these methods are often hard to transfer and scale on various tasks, and separately trained retrievers incur a lot of parameter storage and deployment cost. In this paper, we propose Unified Demonstration Retriever (\textbf{UDR}), a single model to retrieve demonstrations for a wide range of tasks. To train UDR, we cast various tasks' training signals into a unified list-wise ranking formulation by language model's feedback. Then we propose a multi-task list-wise ranking training framework, with an iterative mining strategy to find high-quality candidates, which can help UDR fully incorporate various tasks' signals. Experiments on 30+ tasks across 13 task families and multiple data domains show that UDR significantly outperforms baselines. Further analyses show the effectiveness of each proposed component and UDR's strong ability in various scenarios including different LMs (1.3B - 175B), unseen datasets, varying demonstration quantities, etc.	true	true	Li, Xiaonan and Lv, Kai and Yan, Hang and Lin, Tianyang and Zhu, Wei and Ni, Yuan and Xie, Guotong and Wang, Xiaoling and Qiu, Xipeng	2,023	null	null	null	null	Unified Demonstration Retriever for In-Context Learning	Unified Demonstration Retriever for In-Context Learning	https://aclanthology.org/2023.acl-long.256/	In this paper, we propose Unified Demonstration Retriever (UDR), a single model to retrieve demonstrations for a wide range of tasks.
Dual Debiasing for Noisy In-Context Learning for Text Generation	2506.00418v1	liucontext	\cite{liucontext}	In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering	http://arxiv.org/abs/2311.06668v3	Large language models (LLMs) demonstrate emergent in-context learning capabilities, where they adapt to new tasks based on example demonstrations. However, in-context learning has seen limited effectiveness in many settings, is difficult to quantitatively control and takes up context window space. To overcome these limitations, we propose an alternative approach that recasts in-context learning as in-context vectors (ICV). Using ICV has two steps. We first use a forward pass on demonstration examples to create the in-context vector from the latent embedding of the LLM. This vector captures essential information about the intended task. On a new query, instead of adding demonstrations to the prompt, we shift the latent states of the LLM using the ICV. The ICV approach has several benefits: 1) it enables the LLM to more effectively follow the demonstration examples; 2) it's easy to control by adjusting the magnitude of the ICV; 3) it reduces the length of the prompt by removing the in-context demonstrations; 4) ICV is computationally much more efficient than fine-tuning. We demonstrate that ICV achieves better performance compared to standard in-context learning and fine-tuning on diverse tasks including safety, style transfer, role-playing and formatting. Moreover, we show that we can flexibly teach LLM to simultaneously follow different types of instructions by simple vector arithmetics on the corresponding ICVs.	true	true	Liu, Sheng and Ye, Haotian and Xing, Lei and Zou, James Y	null	null	null	null	null	In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering	Making In Context Learning More Effective and ...	https://consensus.app/papers/incontext-vectors-making-in-context-learning-more-zou-liu/20a28c8387155fa1ac876aad9841f1ee	Key takeaway: 'In-context vectors (ICV) improve in-context learning effectiveness, controllability, and computational efficiency in large

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

NBERw21340

\cite{NBERw21340}

Effective Policy for Reducing Inequality? The Earned Income Tax Credit and the Distribution of Income

null

true

false

Hoynes, Hilary W and Patel, Ankur J

2,015

July

http://www.nber.org/papers/w21340

10.3386/w21340

null

Effective Policy for Reducing Inequality? The Earned Income Tax Credit and the Distribution of Income

Effective Policy for Reducing Inequality? The Earned Income

https://ideas.repec.org/p/nbr/nberwo/21340.html

Our results show that a policy-induced $1000 increase in the EITC leads to a 7.3 percentage point increase in employment and a 9.4 percentage point reduction

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

NBERw21211

\cite{NBERw21211}

The Earned Income Tax Credit (EITC)

null

true

false

Nichols, Austin and Rothstein, Jesse

2,015

May

http://www.nber.org/papers/w21211

10.3386/w21211

null

The Earned Income Tax Credit (EITC)

What is the earned income tax credit? - Tax Policy Center

https://taxpolicycenter.org/briefing-book/what-earned-income-tax-credit

The earned income tax credit (EITC) provides substantial support to low- and moderate-income working parents who claim a qualifying child.

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

Foo2019ProcessAC

\cite{Foo2019ProcessAC}

Process and Critical Approaches to Solving the Systemic Climate Change Governance Problem

null

true

false

Check Woo Foo

2,019

null

https://api.semanticscholar.org/CorpusID:235319207

null

Politics \& Energy eJournal

Process and Critical Approaches to Solving the Systemic Climate Change Governance Problem

Process and Critical Approaches to Solving the Systemic Climate ...

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3608501

The most important and urgent task, besides avoiding nuclear war, is abatement of the existential threat of systemic climate change,

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

Patjoshi2015DesignAD

\cite{Patjoshi2015DesignAD}

Design and Development of Advanced Control strategies for Power Quality Enhancement at Distribution Level

null

true

false

Rajesh Kumar Patjoshi

2,015

null

https://api.semanticscholar.org/CorpusID:112918597

null

Design and Development of Advanced Control strategies for Power Quality Enhancement at Distribution Level

(PDF) Advanced Control Strategies for UPQC to Improve ...

https://www.researchgate.net/publication/279289697_Advanced_Control_Strategies_for_UPQC_to_Improve_Power_Quality_of_Power_Distribution_Systems

PDF | On Jul 2, 2014, Quoc Nam Trinh published Advanced Control Strategies for UPQC to Improve Power Quality of Power Distribution Systems

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

10.1257/jep.25.4.165

\cite{10.1257/jep.25.4.165}

The Case for a Progressive Tax: From Basic Research to Policy Recommendations

null

true

false

Diamond, Peter and Saez, Emmanuel

2,011

December

https://www.aeaweb.org/articles?id=10.1257/jep.25.4.165

10.1257/jep.25.4.165

Journal of Economic Perspectives

The Case for a Progressive Tax: From Basic Research to Policy Recommendations

The Case for a Progressive Tax

https://economics.mit.edu/sites/default/files/2022-09/jep.25.4.165.pdf

Therefore, optimal income tax theory is fi rst a normative theory that shows how a social welfare objective combines with constraints arising from theory that shows how a social welfare objective combines with constraints arising from limits on resources and behavioral responses to taxation in order to derive specifi c limits on resources and behavioral responses to taxation in order to derive specifi c The Case for a Progressive Tax: From Basic Research to Policy Recommendations † ■ ■ Peter Diamond is Professor Emeritus of Economics, Massachusetts Institute of Tech-nology, Cambridge Massachusetts. doi=10.1257/jep.25.4.165 Peter Diamond and Emmanuel Saez 166 Journal of Economic Perspectives tax policy recommendations. In addition, optimal income tax theory can be used to tax policy recommendations.

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

10.2307/2296779

\cite{10.2307/2296779}

An Exploration in the Theory of Optimum Income Taxation12

null

true

false

Mirrlees, J. A.

1,971

04

https://doi.org/10.2307/2296779

10.2307/2296779

The Review of Economic Studies

An Exploration in the Theory of Optimum Income Taxation12

Exploration in the Theory of Optimum Income Taxation12

https://academic.oup.com/restud/article-abstract/38/2/175/1527903

by JA Mirrlees · 1971 · Cited by 7415 — J. A. Mirrlees; An Exploration in the Theory of Optimum Income Taxation12, The Review of Economic Studies, Volume 38, Issue 2, 1 April 1971, Pages 175–208,

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

RePEc:aea:aecrev:v:61:y:1971:i:1:p:8-27

\cite{RePEc:aea:aecrev:v:61:y:1971:i:1:p:8-27}

Optimal Taxation and Public Production: I--Production Efficiency

null

true

false

Diamond, Peter and Mirrlees, James

1,971

null

https://EconPapers.repec.org/RePEc:aea:aecrev:v:61:y:1971:i:1:p:8-27

null

American Economic Review

Optimal Taxation and Public Production: I--Production Efficiency

[PDF] Optimal Taxation and Public Production I: Production Efficiency

http://hassler-j.iies.su.se/Courses/DynPubFin/Papers/DiamondMirrlees.pdf

Theories of optimal production in a planned economy have usually assumed that the tax system can allow the govern- ment to achieve any desired redistribution of

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

10.1111/1467-937X.00166

\cite{10.1111/1467-937X.00166}

Using Elasticities to Derive Optimal Income Tax Rates

null

true

false

Saez, Emmanuel

2,001

01

https://doi.org/10.1111/1467-937X.00166

10.1111/1467-937X.00166

The Review of Economic Studies

Using Elasticities to Derive Optimal Income Tax Rates

https://academic.oup.com/restud/article/68/1/205/1568609

by E Saez · 2001 · Cited by 1885 — This paper derives optimal income tax formulas using compensated and uncompensated elasticities of earnings with respect to tax rates.

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

10.1257/pol.6.1.230

\cite{10.1257/pol.6.1.230}

Optimal Taxation of Top Labor Incomes: A Tale of Three Elasticities

null

true

false

Piketty, Thomas and Saez, Emmanuel and Stantcheva, Stefanie

2,014

February

https://www.aeaweb.org/articles?id=10.1257/pol.6.1.230

10.1257/pol.6.1.230

American Economic Journal: Economic Policy

Optimal Taxation of Top Labor Incomes: A Tale of Three Elasticities

https://www.nber.org/papers/w17616

This paper presents a model of optimal labor income taxation where top incomes respond to marginal tax rates through three channels.

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

10.1257/pol.20180033

\cite{10.1257/pol.20180033}

Optimal Income Taxation with Unemployment and Wage Responses: A Sufficient Statistics Approach

null

true

false

Kroft, Kory and Kucko, Kavan and Lehmann, Etienne and Schmieder, Johannes

2,020

February

https://www.aeaweb.org/articles?id=10.1257/pol.20180033

10.1257/pol.20180033

American Economic Journal: Economic Policy

Optimal Income Taxation with Unemployment and Wage Responses: A Sufficient Statistics Approach

Optimal Income Taxation with Unemployment and Wage Responses

https://www.aeaweb.org/articles?id=10.1257/pol.20180033

We derive a sufficient statistics tax formula in a model that incorporates unemployment and endogenous wages to study the shape of the optimal income tax. Key

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

zheng2020aieconomistimprovingequality

\cite{zheng2020aieconomistimprovingequality}

The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies

http://arxiv.org/abs/2004.13332v1

Tackling real-world socio-economic challenges requires designing and testing economic policies. However, this is hard in practice, due to a lack of appropriate (micro-level) economic data and limited opportunity to experiment. In this work, we train social planners that discover tax policies in dynamic economies that can effectively trade-off economic equality and productivity. We propose a two-level deep reinforcement learning approach to learn dynamic tax policies, based on economic simulations in which both agents and a government learn and adapt. Our data-driven approach does not make use of economic modeling assumptions, and learns from observational data alone. We make four main contributions. First, we present an economic simulation environment that features competitive pressures and market dynamics. We validate the simulation by showing that baseline tax systems perform in a way that is consistent with economic theory, including in regard to learned agent behaviors and specializations. Second, we show that AI-driven tax policies improve the trade-off between equality and productivity by 16% over baseline policies, including the prominent Saez tax framework. Third, we showcase several emergent features: AI-driven tax policies are qualitatively different from baselines, setting a higher top tax rate and higher net subsidies for low incomes. Moreover, AI-driven tax policies perform strongly in the face of emergent tax-gaming strategies learned by AI agents. Lastly, AI-driven tax policies are also effective when used in experiments with human participants. In experiments conducted on MTurk, an AI tax policy provides an equality-productivity trade-off that is similar to that provided by the Saez framework along with higher inverse-income weighted social welfare.

true

Stephan Zheng and Alexander Trott and Sunil Srinivasa and Nikhil Naik and Melvin Gruesbeck and David C. Parkes and Richard Socher

2,020

null

https://arxiv.org/abs/2004.13332

null

The AI Economist: Improving Equality and Productivity with AI-Driven Tax Policies

[PDF] Improving Equality and Productivity with AI-Driven Tax Policies - arXiv

http://arxiv.org/pdf/2004.13332

The AI Economist uses AI to discover tax policies that improve the trade-off between equality and productivity, achieving a 16% improvement

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

NBERc14009

\cite{NBERc14009}

The Impact of Machine Learning on Economics

null

true

false

Susan Athey

2,018

January

http://www.nber.org/chapters/c14009

null

The Impact of Machine Learning on Economics

https://www.gsb.stanford.edu/faculty-research/publications/impact-machine-learning-economics

# The Impact of Machine Learning on Economics This paper provides an assessment of the early contributions of machine learning to economics, as well as predictions about its future contributions. It begins by briefly overviewing some themes from the literature on machine learning, and then draws some contrasts with traditional approaches to estimating the impact of counterfactual policies in economics. Next, we review some of the initial “off-the-shelf” applications of machine learning to economics, including applications in analyzing text and images. Finally, we overview a set of broader predictions about the future impact of machine learning on economics, including its impacts on the nature of collaboration, funding, research tools, and research questions. ## Footer contact links ## Footer 1 ## Footer 2 ## Footer legal links

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

AxtellFarmer2022

\cite{AxtellFarmer2022}

Agent Based Modeling in Economics and Finance: Past, Present, and Future

null

true

false

Axtell, R. and Farmer, J.

2,022

null

Journal of Economic Literature

Agent Based Modeling in Economics and Finance: Past, Present, and Future

[PDF] Agent-Based Modeling in Economics and Finance: Past, Present ...

https://complexityhandbook.uni-hohenheim.de/fileadmin/einrichtungen/complexityhandbook/AXTELL_Robert.pdf

Abstract. Agent-based modeling is a novel computational methodology for representing the behavior of individuals in order to study social phenomena.

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

DelliGatti2018

\cite{DelliGatti2018}

Contents

null

true

false

Delli Gatti, Domenico and Fagiolo, Giorgio and Gallegati, Mauro and Richiardi, Matteo and Russo, Alberto

2,018

null

Contents

CONTENTS | definition in the Cambridge English Dictionary

https://dictionary.cambridge.org/us/dictionary/english/contents

everything that is contained within something: contents of The contents of his bag spilled all over the floor. He didn't need to open the box because

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

shen2025phyxdoesmodelwits

\cite{shen2025phyxdoesmodelwits}

PhyX: Does Your Model Have the "Wits" for Physical Reasoning?

http://arxiv.org/abs/2505.15929v2

Existing benchmarks fail to capture a crucial aspect of intelligence: physical reasoning, the integrated ability to combine domain knowledge, symbolic reasoning, and understanding of real-world constraints. To address this gap, we introduce PhyX: the first large-scale benchmark designed to assess models capacity for physics-grounded reasoning in visual scenarios. PhyX includes 3K meticulously curated multimodal questions spanning 6 reasoning types across 25 sub-domains and 6 core physics domains: thermodynamics, electromagnetism, mechanics, modern physics, optics, and wave\&acoustics. In our comprehensive evaluation, even state-of-the-art models struggle significantly with physical reasoning. GPT-4o, Claude3.7-Sonnet, and GPT-o4-mini achieve only 32.5%, 42.2%, and 45.8% accuracy respectively-performance gaps exceeding 29% compared to human experts. Our analysis exposes critical limitations in current models: over-reliance on memorized disciplinary knowledge, excessive dependence on mathematical formulations, and surface-level visual pattern matching rather than genuine physical understanding. We provide in-depth analysis through fine-grained statistics, detailed case studies, and multiple evaluation paradigms to thoroughly examine physical reasoning capabilities. To ensure reproducibility, we implement a compatible evaluation protocol based on widely-used toolkits such as VLMEvalKit, enabling one-click evaluation. More details are available on our project page: https://phyx-bench.github.io/.

true

Hui Shen and Taiqiang Wu and Qi Han and Yunta Hsieh and Jizhou Wang and Yuyue Zhang and Yuxin Cheng and Zijian Hao and Yuansheng Ni and Xin Wang and Zhongwei Wan and Kai Zhang and Wendong Xu and Jing Xiong and Ping Luo and Wenhu Chen and Chaofan Tao and Zhuoqing Mao and Ngai Wong

2,025

null

https://arxiv.org/abs/2505.15929

null

PhyX: Does Your Model Have the "Wits" for Physical Reasoning?

http://arxiv.org/pdf/2505.15929v2

Existing benchmarks fail to capture a crucial aspect of intelligence: physical reasoning, the integrated ability to combine domain knowledge, symbolic reasoning, and understanding of real-world constraints. To address this gap, we introduce PhyX: the first large-scale benchmark designed to assess models capacity for physics-grounded reasoning in visual scenarios. PhyX includes 3K meticulously curated multimodal questions spanning 6 reasoning types across 25 sub-domains and 6 core physics domains: thermodynamics, electromagnetism, mechanics, modern physics, optics, and wave\&acoustics. In our comprehensive evaluation, even state-of-the-art models struggle significantly with physical reasoning. GPT-4o, Claude3.7-Sonnet, and GPT-o4-mini achieve only 32.5%, 42.2%, and 45.8% accuracy respectively-performance gaps exceeding 29% compared to human experts. Our analysis exposes critical limitations in current models: over-reliance on memorized disciplinary knowledge, excessive dependence on mathematical formulations, and surface-level visual pattern matching rather than genuine physical understanding. We provide in-depth analysis through fine-grained statistics, detailed case studies, and multiple evaluation paradigms to thoroughly examine physical reasoning capabilities. To ensure reproducibility, we implement a compatible evaluation protocol based on widely-used toolkits such as VLMEvalKit, enabling one-click evaluation. More details are available on our project page: https://phyx-bench.github.io/.

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

zhao2024competeaiunderstandingcompetitiondynamics

\cite{zhao2024competeaiunderstandingcompetitiondynamics}

CompeteAI: Understanding the Competition Dynamics in Large Language Model-based Agents

http://arxiv.org/abs/2310.17512v2

Large language models (LLMs) have been widely used as agents to complete different tasks, such as personal assistance or event planning. While most of the work has focused on cooperation and collaboration between agents, little work explores competition, another important mechanism that promotes the development of society and economy. In this paper, we seek to examine the competition dynamics in LLM-based agents. We first propose a general framework for studying the competition between agents. Then, we implement a practical competitive environment using GPT-4 to simulate a virtual town with two types of agents, restaurant agents and customer agents. Specifically, the restaurant agents compete with each other to attract more customers, where competition encourages them to transform, such as cultivating new operating strategies. Simulation experiments reveal several interesting findings at the micro and macro levels, which align well with existing market and sociological theories. We hope that the framework and environment can be a promising testbed to study competition that fosters understanding of society. Code is available at: https://github.com/microsoft/competeai.

true

Qinlin Zhao and Jindong Wang and Yixuan Zhang and Yiqiao Jin and Kaijie Zhu and Hao Chen and Xing Xie

2,024

null

https://arxiv.org/abs/2310.17512

null

CompeteAI: Understanding the Competition Dynamics in Large Language Model-based Agents

CompeteAI: Understanding the Competition Dynamics in Large ...

https://arxiv.org/abs/2310.17512

In this paper, we seek to examine the competition dynamics in LLM-based agents. We first propose a general framework for studying the competition between

TaxAgent: How Large Language Model Designs Fiscal Policy

2506.02838v1

nie2024surveylargelanguagemodels

\cite{nie2024surveylargelanguagemodels}

A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges

null

true

false

Yuqi Nie and Yaxuan Kong and Xiaowen Dong and John M. Mulvey and H. Vincent Poor and Qingsong Wen and Stefan Zohren

2,024

null

https://arxiv.org/abs/2406.11903

null

A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges

http://arxiv.org/pdf/2406.11903v1

Recent advances in large language models (LLMs) have unlocked novel opportunities for machine learning applications in the financial domain. These models have demonstrated remarkable capabilities in understanding context, processing vast amounts of data, and generating human-preferred contents. In this survey, we explore the application of LLMs on various financial tasks, focusing on their potential to transform traditional practices and drive innovation. We provide a discussion of the progress and advantages of LLMs in financial contexts, analyzing their advanced technologies as well as prospective capabilities in contextual understanding, transfer learning flexibility, complex emotion detection, etc. We then highlight this survey for categorizing the existing literature into key application areas, including linguistic tasks, sentiment analysis, financial time series, financial reasoning, agent-based modeling, and other applications. For each application area, we delve into specific methodologies, such as textual analysis, knowledge-based analysis, forecasting, data augmentation, planning, decision support, and simulations. Furthermore, a comprehensive collection of datasets, model assets, and useful codes associated with mainstream applications are presented as resources for the researchers and practitioners. Finally, we outline the challenges and opportunities for future research, particularly emphasizing a number of distinctive aspects in this field. We hope our work can help facilitate the adoption and further development of LLMs in the financial sector.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

vllm

\cite{vllm}

Efficient Memory Management for Large Language Model Serving with PagedAttention

http://arxiv.org/abs/2309.06180v1

High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2-4$\times$ with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm

true

Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph Gonzalez and Hao Zhang and Ion Stoica

2,023

null

https://doi.org/10.1145/3600006.3613165

10.1145/3600006.3613165

null

Efficient Memory Management for Large Language Model Serving with PagedAttention

Efficient Memory Management for Large Language Model ...

https://arxiv.org/pdf/2309.06180

Efficient Memory Management for Large Language Model Serving with PagedAttention Woosuk Kwon 1,∗ Zhuohan Li 1,∗ Siyuan Zhuang 1 Ying Sheng 1,2 Lianmin Zheng 1 Cody Hao Yu 3 Joseph E. Gonzalez 1 Hao Zhang 4 Ion Stoica 1 1 UC Berkeley 2Stanford University 3Independent Researcher 4UC San Diego Abstract High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. To address this problem, we propose PagedAttention, an attention al-gorithm inspired by the classical virtual memory and pag-ing techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce mem-ory usage. To address the above limitations, we propose PagedAt-tention , an attention algorithm inspired by the operating system’s (OS) solution to memory fragmentation and shar-ing: virtual memory with paging . In this work, we build vLLM , a high-throughput distributed LLM serving engine on top of PagedAttention that achieves near-zero waste in KV cache memory.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

chunkattention

\cite{chunkattention}

ChunkAttention: Efficient Self-Attention with Prefix-Aware KV Cache and Two-Phase Partition

http://arxiv.org/abs/2402.15220v4

Self-attention is an essential component of large language models (LLM) but a significant source of inference latency for long sequences. In multi-tenant LLM serving scenarios, the compute and memory operation cost of self-attention can be optimized by using the probability that multiple LLM requests have shared system prompts in prefixes. In this paper, we introduce ChunkAttention, a prefix-aware self-attention module that can detect matching prompt prefixes across multiple requests and share their key/value tensors in memory at runtime to improve the memory utilization of KV cache. This is achieved by breaking monolithic key/value tensors into smaller chunks and structuring them into the auxiliary prefix tree. Consequently, on top of the prefix-tree based KV cache, we design an efficient self-attention kernel, where a two-phase partition algorithm is implemented to improve the data locality during self-attention computation in the presence of shared system prompts. Experiments show that ChunkAttention can speed up the self-attention kernel by 3.2-4.8$\times$ compared to the state-of-the-art implementation, with the length of the system prompt ranging from 1024 to 4096.

true

Lu Ye and Ze Tao and Yong Huang and Yang Li

2,024

null

https://aclanthology.org/2024.acl-long.623

null

ChunkAttention: Efficient Self-Attention with Prefix-Aware KV Cache and Two-Phase Partition

[PDF] Efficient Self-Attention with Prefix-Aware KV Cache and Two-Phase ...

https://aclanthology.org/2024.acl-long.623.pdf

ChunkAttention is a prefix-aware self-attention module that uses a prefix-aware KV cache and two-phase partition to improve memory utilization

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

cachedattention

\cite{cachedattention}

Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttention

null

true

false

Bin Gao and Zhuomin He and Puru Sharma and Qingxuan Kang and Djordje Jevdjic and Junbo Deng and Xingkun Yang and Zhou Yu and Pengfei Zuo

2,024

null

https://www.usenix.org/conference/atc24/presentation/gao-bin-cost

null

Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttention

Cost-Efficient Large Language Model Serving for Multi-turn ... - arXiv

https://arxiv.org/abs/2403.19708

This paper proposes CachedAttention, a new attention mechanism that enables reuse of KV caches across multi-turn conversations, significantly reducing the

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

promptcache

\cite{promptcache}

Prompt Cache: Modular Attention Reuse for Low-Latency Inference

http://arxiv.org/abs/2311.04934v2

We present Prompt Cache, an approach for accelerating inference for large language models (LLM) by reusing attention states across different LLM prompts. Many input prompts have overlapping text segments, such as system messages, prompt templates, and documents provided for context. Our key insight is that by precomputing and storing the attention states of these frequently occurring text segments on the inference server, we can efficiently reuse them when these segments appear in user prompts. Prompt Cache employs a schema to explicitly define such reusable text segments, called prompt modules. The schema ensures positional accuracy during attention state reuse and provides users with an interface to access cached states in their prompt. Using a prototype implementation, we evaluate Prompt Cache across several LLMs. We show that Prompt Cache significantly reduce latency in time-to-first-token, especially for longer prompts such as document-based question answering and recommendations. The improvements range from 8x for GPU-based inference to 60x for CPU-based inference, all while maintaining output accuracy and without the need for model parameter modifications.

true

In Gim and Guojun Chen and Seung{-}Seob Lee and Nikhil Sarda and Anurag Khandelwal and Lin Zhong

2,024

null

https://proceedings.mlsys.org/paper\_files/paper/2024/hash/a66caa1703fe34705a4368c3014c1966-Abstract-Conference.html

null

Prompt Cache: Modular Attention Reuse for Low-Latency Inference

[PDF] Prompt Cache: Modular Attention Reuse for Low-Latency Inference

https://proceedings.mlsys.org/paper_files/paper/2024/file/a66caa1703fe34705a4368c3014c1966-Paper-Conference.pdf

Prompt Cache accelerates LLM inference by reusing attention states of frequently occurring text segments, precomputed and stored in memory.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

sglang

\cite{sglang}

Efficiently Programming Large Language Models using SGLang

null

true

false

Lianmin Zheng and Liangsheng Yin and Zhiqiang Xie and Jeff Huang and Chuyue Sun and Cody Hao Yu and Shiyi Cao and Christos Kozyrakis and Ion Stoica and Joseph E. Gonzalez and Clark W. Barrett and Ying Sheng

2,023

null

https://doi.org/10.48550/arXiv.2312.07104

10.48550/ARXIV.2312.07104

CoRR

Efficiently Programming Large Language Models using SGLang

https://arxiv.org/html/2312.07104v1

SGLang simplifies the writing of LLM programs and boosts execution efficiency. Our experiments demonstrate that SGLang can speed up common LLM tasks by up to 5

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

cacheblend

\cite{cacheblend}

CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion

http://arxiv.org/abs/2405.16444v3

Large language models (LLMs) often incorporate multiple text chunks in their inputs to provide the necessary contexts. To speed up the prefill of the long LLM inputs, one can pre-compute the KV cache of a text and re-use the KV cache when the context is reused as the prefix of another LLM input. However, the reused text chunks are not always the input prefix, which makes precomputed KV caches not directly usable since they ignore the text's cross-attention with the preceding texts. Thus, the benefits of reusing KV caches remain largely unrealized. This paper tackles just one challenge: when an LLM input contains multiple text chunks, how to quickly combine their precomputed KV caches in order to achieve the same generation quality as the expensive full prefill (i.e., without reusing KV cache)? This challenge naturally arises in retrieval-augmented generation (RAG) where the input is supplemented with multiple retrieved texts as the context. We present CacheBlend, a scheme that reuses the precomputed KV caches, regardless prefix or not, and selectively recomputes the KV values of a small subset of tokens to partially update each reused KV cache. In the meantime, the small extra delay for recomputing some tokens can be pipelined with the retrieval of KV caches within the same job, allowing CacheBlend to store KV caches in slower devices with more storage capacity while retrieving them without increasing the inference delay. By comparing CacheBlend with the state-of-the-art KV cache reusing schemes on three open-source LLMs of various sizes and four popular benchmark datasets of different tasks, we show that CacheBlend reduces time-to-first-token (TTFT) by 2.2-3.3x and increases the inference throughput by 2.8-5x from full KV recompute without compromising generation quality. The code is available at https://github.com/LMCache/LMCache.

true

Jiayi Yao and Hanchen Li and Yuhan Liu and Siddhant Ray and Yihua Cheng and Qizheng Zhang and Kuntai Du and Shan Lu and Junchen Jiang

2,024

null

https://doi.org/10.48550/arXiv.2405.16444

10.48550/ARXIV.2405.16444

CoRR

CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion

CacheBlend: Fast Large Language Model Serving for RAG ... - arXiv

https://arxiv.org/abs/2405.16444

Image 4: arxiv logo>cs> arXiv:2405.16444 View a PDF of the paper titled CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion, by Jiayi Yao and 8 other authors View a PDF of the paper titled CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion, by Jiayi Yao and 8 other authors - [x] Bibliographic Explorer Toggle - [x] Connected Papers Toggle - [x] Litmaps Toggle - [x] alphaXiv Toggle - [x] Links to Code Toggle - [x] DagsHub Toggle - [x] GotitPub Toggle - [x] Huggingface Toggle - [x] Links to Code Toggle - [x] ScienceCast Toggle - [x] Replicate Toggle - [x] Spaces Toggle - [x] Core recommender toggle - [x] IArxiv recommender toggle

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

openaiapi

\cite{openaiapi}

OpenAI developer platform

null

true

false

OpenAI

null

OpenAI developer platform

Introducing Verdi, an AI dev platform powered by GPT-4o - OpenAI

https://openai.com/index/mercado-libre/

Verdi, a development platform layer using GPT-4o, GPT-4o mini, and GPT-3.5 Turbo, which is transforming how Mercado Libre handles customer service and other

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

genimiapi

\cite{genimiapi}

Gemini API

null

true

false

Google

2,025

null

Gemini API

Gemini Developer API | Gemma open models | Google AI for ...

https://ai.google.dev/

Gemini Developer API | Gemma open models | Google AI for Developers - Gemini Showcase - Gemini Showcase ### Integrate Google AI models with an API key Build with cutting-edge AI models, like Gemini, Imagen, and Veo, from Google DeepMind Integrate Google AI models with an API key Unlock AI capabilities for your apps with a simple call to the Gemini API. Integrate AI models like Gemini Nano into web apps with Chrome's built-in web platform APIs. Build trusted and secure AI with guidance for responsible design, development, and deployment of models and applications. See how the Ruby-based AI agent framework empowers developer teams to be more productive with the power of Gemini models.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

claudeapi

\cite{claudeapi}

Claude API

null

true

false

Anthropic

2,025

null

Claude API

Anthropic API

https://docs.anthropic.com/en/home

Home - Anthropic Claude Documentation Learn how to get started with the Anthropic API, the Console, and Claude Code. Explore the advanced features and capabilities now available in Claude.## API reference Integrate and scale using our API and SDKs.## Anthropic Console Learn about changes and new features in Claude and the API.## Upgrade to Claude 4 Upgrade to the latest model to access new tools and features available in Claude 4. ## Claude Code ## Claude Code quickstart Get started with Claude Code.## Claude Code reference Consult the Claude Code reference documentation for details on feature implementation and configuration.## Claude Code release notes Learn about changes and new features in Claude Code. See replicable code samples and implementations.## Anthropic Quickstarts

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

mooncake

\cite{mooncake}

Mooncake Trace

null

true

false

null

2,025

null

Mooncake Trace

kvcache-ai/Mooncake - GitHub

https://github.com/kvcache-ai/Mooncake

Moonshot AI. Now both the Transfer Engine and Mooncake Store are open-sourced! This repository also hosts its technical report and the open sourced traces.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

hu2024epic

\cite{hu2024epic}

EPIC: Efficient Position-Independent Context Caching for Serving Large Language Models

null

true

false

Junhao Hu and Wenrui Huang and Haoyi Wang and Weidong Wang and Tiancheng Hu and Qin Zhang and Hao Feng and Xusheng Chen and Yizhou Shan and Tao Xie

2,024

null

https://arxiv.org/abs/2410.15332

null

EPIC: Efficient Position-Independent Context Caching for Serving Large Language Models

EPIC: Efficient Position-Independent Caching for Serving Large...

https://openreview.net/forum?id=qjd3ZUiHRT&referrer=%5Bthe%20profile%20of%20Yizhou%20Shan%5D(%2Fprofile%3Fid%3D~Yizhou_Shan2)

Summary: This paper introduces PICI, an efficient position-independent context caching system for serving large language models. The system pre-computes the KV

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

streamingllm

\cite{streamingllm}

Efficient Streaming Language Models with Attention Sinks

http://arxiv.org/abs/2309.17453v4

Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory. Secondly, popular LLMs cannot generalize to longer texts than the training sequence length. Window attention, where only the most recent KVs are cached, is a natural approach -- but we show that it fails when the text length surpasses the cache size. We observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a "sink" even if they are not semantically important. Based on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence lengths without any fine-tuning. We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more. In addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2x speedup. Code and datasets are provided at https://github.com/mit-han-lab/streaming-llm.

true

Guangxuan Xiao and Yuandong Tian and Beidi Chen and Song Han and Mike Lewis

2,024

null

https://openreview.net/forum?id=NG7sS51zVF

null

Efficient Streaming Language Models with Attention Sinks

http://arxiv.org/pdf/2309.17453v4

Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory. Secondly, popular LLMs cannot generalize to longer texts than the training sequence length. Window attention, where only the most recent KVs are cached, is a natural approach -- but we show that it fails when the text length surpasses the cache size. We observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a "sink" even if they are not semantically important. Based on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence lengths without any fine-tuning. We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more. In addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2x speedup. Code and datasets are provided at https://github.com/mit-han-lab/streaming-llm.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

h2o

\cite{h2o}

{H2O:} Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models

null

true

false

Zhenyu Zhang and Ying Sheng and Tianyi Zhou and Tianlong Chen and Lianmin Zheng and Ruisi Cai and Zhao Song and Yuandong Tian and Christopher R{\'{e}} and Clark W. Barrett and Zhangyang Wang and Beidi Chen

2,023

null

http://papers.nips.cc/paper\_files/paper/2023/hash/6ceefa7b15572587b78ecfcebb2827f8-Abstract-Conference.html

null

{H2O:} Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models

Hogwild! Inference: Parallel LLM Generation via Concurrent Attention

https://arxiv.org/html/2504.06261v1

H2o: Heavy-hitter oracle for efficient generative inference of large language models. Advances in Neural Information Processing Systems, 36

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

infinigen

\cite{infinigen}

InfiniGen: Efficient Generative Inference of Large Language Models with Dynamic KV Cache Management

http://arxiv.org/abs/2406.19707v1

Transformer-based large language models (LLMs) demonstrate impressive performance across various natural language processing tasks. Serving LLM inference for generating long contents, however, poses a challenge due to the enormous memory footprint of the transient state, known as the key-value (KV) cache, which scales with the sequence length and batch size. In this paper, we present InfiniGen, a novel KV cache management framework tailored for long-text generation, which synergistically works with modern offloading-based inference systems. InfiniGen leverages the key insight that a few important tokens that are essential for computing the subsequent attention layer in the Transformer can be speculated by performing a minimal rehearsal with the inputs of the current layer and part of the query weight and key cache of the subsequent layer. This allows us to prefetch only the essential KV cache entries (without fetching them all), thereby mitigating the fetch overhead from the host memory in offloading-based LLM serving systems. Our evaluation on several representative LLMs shows that InfiniGen improves the overall performance of a modern offloading-based system by up to 3.00x compared to prior KV cache management methods while offering substantially better model accuracy.

true

Wonbeom Lee and Jungi Lee and Junghwan Seo and Jaewoong Sim

2,024

null

https://www.usenix.org/conference/osdi24/presentation/lee

null

InfiniGen: Efficient Generative Inference of Large Language Models with Dynamic KV Cache Management

InfiniGen: Efficient Generative Inference of Large Language Models ...

https://arxiv.org/abs/2406.19707

In this paper, we present InfiniGen, a novel KV cache management framework tailored for long-text generation, which synergistically works with modern

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

pyramidkv

\cite{pyramidkv}

PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling

http://arxiv.org/abs/2406.02069v4

In this study, we investigate whether attention-based information flow inside large language models (LLMs) is aggregated through noticeable patterns for long context processing. Our observations reveal that LLMs aggregate information through Pyramidal Information Funneling where attention is scattering widely in lower layers, progressively consolidating within specific contexts, and ultimately focusing on critical tokens (a.k.a massive activation or attention sink) in higher layers. Motivated by these insights, we developed PyramidKV, a novel and effective KV cache compression method. This approach dynamically adjusts the KV cache size across different layers, allocating more cache in lower layers and less in higher ones, diverging from traditional methods that maintain a uniform KV cache size. Our experimental evaluations, utilizing the LongBench benchmark, show that PyramidKV matches the performance of models with a full KV cache while retaining only 12% of the KV cache, thus significantly reducing memory usage. In scenarios emphasizing memory efficiency, where only 0.7% of the KV cache is maintained, PyramidKV surpasses other KV cache compression techniques, achieving up to a 20.5 absolute accuracy improvement on TREC dataset. In the Needle-in-a-Haystack experiment, PyramidKV outperforms competing methods in maintaining long-context comprehension in LLMs; notably, retaining just 128 KV cache entries enables the LLAMA-3-70B model to achieve 100.0 Acc. performance.

true

Zefan Cai and Yichi Zhang and Bofei Gao and Yuliang Liu and Tianyu Liu and Keming Lu and Wayne Xiong and Yue Dong and Baobao Chang and Junjie Hu and Wen Xiao

2,024

null

https://doi.org/10.48550/arXiv.2406.02069

10.48550/ARXIV.2406.02069

CoRR

PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling

PyramidKV: Dynamic KV Cache Compression based on Pyramidal...

https://openreview.net/forum?id=jZVNmDiU86

We developed PyramidKV, a novel and effective KV cache compression method. This approach dynamically adjusts the KV cache size across different layers.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

KVQuant

\cite{KVQuant}

KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization

http://arxiv.org/abs/2401.18079v6

LLMs are seeing growing use for applications which require large context windows, and with these large context windows KV cache activations surface as the dominant contributor to memory consumption during inference. Quantization is a promising approach for compressing KV cache activations; however, existing solutions fail to represent activations accurately in sub-4-bit precision. Our work, KVQuant, facilitates low precision KV cache quantization by incorporating several novel methods: (i) Per-Channel Key Quantization, where we adjust the dimension along which we quantize the Key activations to better match the distribution; (ii) Pre-RoPE Key Quantization, where we quantize Key activations before the rotary positional embedding to mitigate its impact on quantization; (iii) Non-Uniform KV Cache Quantization, where we derive per-layer sensitivity-weighted non-uniform datatypes that better represent the distributions; and (iv) Per-Vector Dense-and-Sparse Quantization, where we isolate outliers separately for each vector to minimize skews in quantization ranges. By applying our method to the LLaMA, Llama-2, Llama-3, and Mistral models, we achieve < 0.1 perplexity degradation with 3-bit quantization on both Wikitext-2 and C4, outperforming existing approaches. Our method enables serving LLaMA-7B with a context length of up to 1 million on a single A100-80GB GPU and up to 10 million on an 8-GPU system. We develop custom CUDA kernels for KVQuant, showing that we can achieve up to ~1.7x speedups, compared to baseline fp16 matrix-vector multiplications, for the LLaMA-7B model.

true

Coleman Hooper and Sehoon Kim and Hiva Mohammadzadeh and Michael W. Mahoney and Yakun Sophia Shao and Kurt Keutzer and Amir Gholami

2,024

null

http://papers.nips.cc/paper\_files/paper/2024/hash/028fcbcf85435d39a40c4d61b42c99a4-Abstract-Conference.html

null

KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization

KVQuant: Towards 10 Million Context Length LLM Inference with KV ...

https://github.com/SqueezeAILab/KVQuant

GitHub - SqueezeAILab/KVQuant: [NeurIPS 2024] KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization [NeurIPS 2024] KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization [Paper] KVQuant is a methodology for efficient KV cache quantization that incorporates several innovations to acheive accurate low-precision quantization, thereby enabling efficient long context length inference. TLDR: KVQuant addresses the memory bottleneck with long context length inference by quantizing the KV cache to low precision. title={KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization}, [NeurIPS 2024] KVQuant: Towards 10 Million Context Length LLM Inference with KV Cache Quantization

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

lruk

\cite{lruk}

The {LRU-K} Page Replacement Algorithm For Database Disk Buffering

null

true

false

Elizabeth J. O'Neil and Patrick E. O'Neil and Gerhard Weikum

1,993

null

https://doi.org/10.1145/170035.170081

10.1145/170035.170081

null

The {LRU-K} Page Replacement Algorithm For Database Disk Buffering

[PDF] The LRU-K Page Replacement Algorithm For Database Disk Buffering

https://www.cs.cmu.edu/~natassa/courses/15-721/papers/p297-o_neil.pdf

The basic idea of. LRU-K is to keep track of the times of the last K references to popular database pages, using this information to statis- tieall y estimate

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

slru

\cite{slru}

Caching Strategies to Improve Disk System Performance

null

true

false

Ramakrishna Karedla and J. Spencer Love and Bradley G. Wherry

1,994

null

https://doi.org/10.1109/2.268884

10.1109/2.268884

Computer

Caching Strategies to Improve Disk System Performance

Caching strategies to improve disk system performance - IEEE Xplore

http://ieeexplore.ieee.org/document/268884/

In this article, we examine the use of caching as a means to increase system response time and improve the data throughput of the disk subsystem.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

twoq

\cite{twoq}

2Q: {A} Low Overhead High Performance Buffer Management Replacement Algorithm

null

true

false

Theodore Johnson and Dennis E. Shasha

1,994

null

http://www.vldb.org/conf/1994/P439.PDF

null

2Q: {A} Low Overhead High Performance Buffer Management Replacement Algorithm

2Q: A Low Overhead High Performance Buffer Management ...

https://dl.acm.org/doi/10.5555/645920.672996

2Q: A Low Overhead High Performance Buffer Management Replacement Algorithm. Authors: Theodore Johnson.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

eelru

\cite{eelru}

{EELRU:} Simple and Effective Adaptive Page Replacement

null

true

false

Yannis Smaragdakis and Scott F. Kaplan and Paul R. Wilson

1,999

null

https://doi.org/10.1145/301453.301486

10.1145/301453.301486

null

{EELRU:} Simple and Effective Adaptive Page Replacement

EELRU: Simple and Effective Adaptive Page Replacement

https://www.researchgate.net/publication/2822757_EELRU_Simple_and_Effective_Adaptive_Page_Replacement

EELRU is a simple adaptive replacement algorithm, which uses only the kind of information needed by LRU---how recently each page has been touched relative to

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

lrfu

\cite{lrfu}

{LRFU:} {A} Spectrum of Policies that Subsumes the Least Recently Used and Least Frequently Used Policies

null

true

false

Donghee Lee and Jongmoo Choi and Jong{-}Hun Kim and Sam H. Noh and Sang Lyul Min and Yookun Cho and Chong{-}Sang Kim

2,001

null

https://doi.org/10.1109/TC.2001.970573

10.1109/TC.2001.970573

{IEEE} Trans. Computers

{LRFU:} {A} Spectrum of Policies that Subsumes the Least Recently Used and Least Frequently Used Policies

[PDF] LRFU: a spectrum of policies that subsumes the least recently used ...

https://www.openu.ac.il/home/wiseman/2os/lru/lrfu.pdf

Of these, the Least Recently Used (LRU) and the. Least Frequently Used (LFU) block replacement policies constitute the two main streams. The LRU policy and its.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

lirs

\cite{lirs}

{LIRS:} an efficient low inter-reference recency set replacement policy to improve buffer cache performance

null

true

false

Song Jiang and Xiaodong Zhang

2,002

null

https://doi.org/10.1145/511334.511340

10.1145/511334.511340

null

{LIRS:} an efficient low inter-reference recency set replacement policy to improve buffer cache performance

LIRS: an efficient low inter-reference recency set replacement policy ...

https://www.researchgate.net/publication/367088056_LIRS_an_efficient_low_inter-reference_recency_set_replacement_policy_to_improve_buffer_cache_performance

Many studies are focused on cache replacement algorithms, such as FIFO, LRU, LFU, and some advanced cache algorithms like ARC [19], LIRS [15] and 2Q [16].

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

arc

\cite{arc}

{ARC:} {A} Self-Tuning, Low Overhead Replacement Cache

null

true

false

Nimrod Megiddo and Dharmendra S. Modha

2,003

null

http://www.usenix.org/events/fast03/tech/megiddo.html

null

{ARC:} {A} Self-Tuning, Low Overhead Replacement Cache

[PDF] ARC: A Self-Tuning, Low Overhead Replacement Cache

https://www.cs.cmu.edu/~natassa/courses/15-721/papers/arcfast.pdf

We propose a new cache management policy, namely, Adaptive. Replacement Cache (ARC), that has several advantages. In response to evolving and changing access

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

mq

\cite{mq}

Second-Level Buffer Cache Management

null

true

false

Yuanyuan Zhou and Zhifeng Chen and Kai Li

2,004

null

https://doi.org/10.1109/TPDS.2004.13

10.1109/TPDS.2004.13

{IEEE} Trans. Parallel Distributed Syst.

Second-Level Buffer Cache Management

[PDF] Second-Level Buffer Cache Management

https://www.openu.ac.il/home/wiseman/2os/lru/mq.pdf

This is a local cache replacement algorithm because it manages an L2 buffer cache without any information from first-level.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

car

\cite{car}

{CAR:} Clock with Adaptive Replacement

null

true

false

Sorav Bansal and Dharmendra S. Modha

2,004

null

http://www.usenix.org/events/fast04/tech/bansal.html

null

{CAR:} Clock with Adaptive Replacement

CAR: Clock with Adaptive Replacement - Stanford CS Theory

http://theory.stanford.edu/~sbansal/pubs/fast04.pdf

by S Bansal · Cited by 412 — CAR is a new algorithm that improves upon CLOCK by being scan-resistant, self-tuning, and adaptively capturing recency and frequency features.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

clockpro

\cite{clockpro}

CLOCK-Pro: An Effective Improvement of the {CLOCK} Replacement

null

true

false

Song Jiang and Feng Chen and Xiaodong Zhang

2,005

null

http://www.usenix.org/events/usenix05/tech/general/jiang.html

null

CLOCK-Pro: An Effective Improvement of the {CLOCK} Replacement

CLOCK-Pro: An Effective Improvement of the CLOCK Replacement

https://www.usenix.org/conference/2005-usenix-annual-technical-conference/clock-pro-effective-improvement-clock-replacement

We propose an improved CLOCK replacement policy, called CLOCK-Pro. By additionally keeping track of a limited number of replaced pages, CLOCK-Pro works in a

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

DBLP:journals/tos/EinzigerEFM22

\cite{DBLP:journals/tos/EinzigerEFM22}

Lightweight Robust Size Aware Cache Management

http://arxiv.org/abs/2105.08770v2

Modern key-value stores, object stores, Internet proxy caches, as well as Content Delivery Networks (CDN) often manage objects of diverse sizes, e.g., blobs, video files of different lengths, images with varying resolution, and small documents. In such workloads, size-aware cache policies outperform size-oblivious algorithms. Unfortunately, existing size-aware algorithms tend to be overly complicated and computationally~expensive. Our work follows a more approachable pattern; we extend the prevalent (size-oblivious) TinyLFU cache admission policy to handle variable sized items. Implementing our approach inside two popular caching libraries only requires minor changes. We show that our algorithms yield competitive or better hit-ratios and byte hit-ratios compared to the state of the art size-aware algorithms such as AdaptSize, LHD, LRB, and GDSF. Further, a runtime comparison indicates that our implementation is faster by up to x3 compared to the best alternative, i.e., it imposes much lower CPU overhead.

true

Gil Einziger and Ohad Eytan and Roy Friedman and Benjamin Manes

2,022

null

https://doi.org/10.1145/3507920

10.1145/3507920

{ACM} Trans. Storage

Lightweight Robust Size Aware Cache Management

http://arxiv.org/pdf/2105.08770v2

Modern key-value stores, object stores, Internet proxy caches, as well as Content Delivery Networks (CDN) often manage objects of diverse sizes, e.g., blobs, video files of different lengths, images with varying resolution, and small documents. In such workloads, size-aware cache policies outperform size-oblivious algorithms. Unfortunately, existing size-aware algorithms tend to be overly complicated and computationally~expensive. Our work follows a more approachable pattern; we extend the prevalent (size-oblivious) TinyLFU cache admission policy to handle variable sized items. Implementing our approach inside two popular caching libraries only requires minor changes. We show that our algorithms yield competitive or better hit-ratios and byte hit-ratios compared to the state of the art size-aware algorithms such as AdaptSize, LHD, LRB, and GDSF. Further, a runtime comparison indicates that our implementation is faster by up to x3 compared to the best alternative, i.e., it imposes much lower CPU overhead.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

lhd

\cite{lhd}

{LHD:} Improving Cache Hit Rate by Maximizing Hit Density

null

true

false

Nathan Beckmann and Haoxian Chen and Asaf Cidon

2,018

null

https://www.usenix.org/conference/nsdi18/presentation/beckmann

null

{LHD:} Improving Cache Hit Rate by Maximizing Hit Density

LHD: improving cache hit rate by maximizing hit density

https://dl.acm.org/doi/10.5555/3307441.3307475

We introduce least hit density (LHD), a novel eviction policy for key-value caches. LHD predicts each object's expected hits-per-space-consumed (hit density).

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

cacheus

\cite{cacheus}

Learning Cache Replacement with {CACHEUS}

null

true

false

Liana V. Rodriguez and Farzana Beente Yusuf and Steven Lyons and Eysler Paz and Raju Rangaswami and Jason Liu and Ming Zhao and Giri Narasimhan

2,021

null

https://www.usenix.org/conference/fast21/presentation/rodriguez

null

Learning Cache Replacement with {CACHEUS}

Learning Cache Replacement with Cacheus

https://www.usenix.org/system/files/fast21-rodriguez.pdf

by LV Rodriguez · 2021 · Cited by 125 — Furthermore, CACHEUS enables augmenting state-of-the-art algorithms (e.g., LIRS, ARC) by combining it with a complementary cache replacement

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

sieve

\cite{sieve}

{SIEVE} is Simpler than {LRU:} an Efficient Turn-Key Eviction Algorithm for Web Caches

null

true

false

Yazhuo Zhang and Juncheng Yang and Yao Yue and Ymir Vigfusson and K. V. Rashmi

2,024

null

https://www.usenix.org/conference/nsdi24/presentation/zhang-yazhuo

null

{SIEVE} is Simpler than {LRU:} an Efficient Turn-Key Eviction Algorithm for Web Caches

SIEVE - An Efficient Turn-Key Eviction Algorithm for Web Caches

https://www.classcentral.com/course/youtube-nsdi-24-sieve-is-simpler-than-lru-an-efficient-turn-key-eviction-algorithm-for-web-caches-294624

Discover how SIEVE outperforms traditional algorithms like LRU in simplicity, efficiency, and scalability for web cache workloads. Learn about the algorithm's

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

cherkasova1998improving

\cite{cherkasova1998improving}

Improving WWW proxies performance with greedy-dual-size-frequency caching policy

null

true

false

Cherkasova, Ludmila

1,998

null

Improving WWW proxies performance with greedy-dual-size-frequency caching policy

Improving WWW proxies performance with Greedy-Dual- ...

https://www.researchgate.net/publication/228542715_Improving_WWW_proxies_performance_with_Greedy-Dual-Size-Frequency_caching_policy

This paper introduces the Greedy-Dual-Size-Frequency caching policy to maximize hit and byte hit rates for WWW proxies. Proposed caching strategy incorporates

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

yang2020twemcache

\cite{yang2020twemcache}

A large scale analysis of hundreds of in-memory cache clusters at Twitter

null

true

false

Juncheng Yang and Yao Yue and K. V. Rashmi

2,020

null

https://www.usenix.org/conference/osdi20/presentation/yang

null

A large scale analysis of hundreds of in-memory cache clusters at Twitter

[PDF] A large scale analysis of hundreds of in-memory cache clusters at ...

https://www.usenix.org/system/files/osdi20-yang.pdf

This paper is included in the Proceedings of the 14th USENIX Symposium on Operating Systems Design and Implementation November 4–6, 2020 978-1-939133-19-9 Open access to the Proceedings of the 14th USENIX Symposium on Operating Systems Design and Implementation is sponsored by USENIX A large scale analysis of hundreds of in-memory cache clusters at Twitter Juncheng Yang, Carnegie Mellon University; Yao Yue, Twitter; K. When memory is full, 192 14th USENIX Symposium on Operating Systems Design and Implementation USENIX Association #cluster request rate cache size cpu cores 0.00 0.25 0.50 0.75 1.00 Fraction of use case storage computation transient Figure 2: Resources consumed for the three cache use cases.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

berg2020cachelib

\cite{berg2020cachelib}

The {CacheLib} Caching Engine: Design and Experiences at Scale

null

true

false

Benjamin Berg and Daniel S. Berger and Sara McAllister and Isaac Grosof and Sathya Gunasekar and Jimmy Lu and Michael Uhlar and Jim Carrig and Nathan Beckmann and Mor Harchol-Balter and Gregory R. Ganger

2,020

null

https://www.usenix.org/conference/osdi20/presentation/berg

null

The {CacheLib} Caching Engine: Design and Experiences at Scale

The CacheLib Caching Engine: Design and Experiences at Scale

https://www.usenix.org/conference/osdi20/presentation/berg

CacheLib is a general-purpose caching engine, designed based on experiences with a range of caching use cases at Facebook, that facilitates the easy

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

icebreaker

\cite{icebreaker}

IceBreaker: warming serverless functions better with heterogeneity

null

true

false

Rohan Basu Roy and Tirthak Patel and Devesh Tiwari

2,022

null

https://doi.org/10.1145/3503222.3507750

10.1145/3503222.3507750

null

IceBreaker: warming serverless functions better with heterogeneity

[PDF] IceBreaker: Warming Serverless Functions Better with Heterogeneity

http://www1.ece.neu.edu/~ningfang/SimPaper/icebreaker-ASPLOS22.pdf

IceBreaker is a novel function pre-warming and keep-alive scheme for serverless functions that exploit server-heterogeneity to lower the keep-alive cost and

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

fasscache

\cite{fasscache}

FaasCache: keeping serverless computing alive with greedy-dual caching

null

true

false

Alexander Fuerst and Prateek Sharma

2,021

null

https://doi.org/10.1145/3445814.3446757

10.1145/3445814.3446757

null

FaasCache: keeping serverless computing alive with greedy-dual caching

[PDF] FaasCache: Keeping Serverless Computing Alive with Greedy-Dual ...

https://afuerst.github.io/assets/FaasCache.pdf

Keep-alive policies must keep functions alive based on their resource and usage characteristics, which is challenging due to the diversity in FaaS workloads.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

DBLP:conf/osdi/ZhongLCHZL0024

\cite{DBLP:conf/osdi/ZhongLCHZL0024}

DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving

http://arxiv.org/abs/2401.09670v3

DistServe improves the performance of large language models (LLMs) serving by disaggregating the prefill and decoding computation. Existing LLM serving systems colocate the two phases and batch the computation of prefill and decoding across all users and requests. We find that this strategy not only leads to strong prefill-decoding interferences but also couples the resource allocation and parallelism plans for both phases. LLM applications often emphasize individual latency for each phase: time to first token (TTFT) for the prefill phase and time per output token (TPOT) of each request for the decoding phase. In the presence of stringent latency requirements, existing systems have to prioritize one latency over the other, or over-provision compute resources to meet both. DistServe assigns prefill and decoding computation to different GPUs, hence eliminating prefill-decoding interferences. Given the application's TTFT and TPOT requirements, DistServe co-optimizes the resource allocation and parallelism strategy tailored for each phase. DistServe also places the two phases according to the serving cluster's bandwidth to minimize the communication caused by disaggregation. As a result, DistServe significantly improves LLM serving performance in terms of the maximum rate that can be served within both TTFT and TPOT constraints on each GPU. Our evaluations show that on various popular LLMs, applications, and latency requirements, DistServe can serve 7.4x more requests or 12.6x tighter SLO, compared to state-of-the-art systems, while staying within latency constraints for > 90% of requests.

true

Yinmin Zhong and Shengyu Liu and Junda Chen and Jianbo Hu and Yibo Zhu and Xuanzhe Liu and Xin Jin and Hao Zhang

2,024

null

https://www.usenix.org/conference/osdi24/presentation/zhong-yinmin

null

DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving

[PDF] DistServe: Disaggregating Prefill and Decoding for Goodput ...

https://www.usenix.org/system/files/osdi24-zhong-yinmin.pdf

July 10–12, 2024 • Santa Clara, CA, USA 978-1-939133-40-3 Open access to the Proceedings of the 18th USENIX Symposium on Operating Systems Design and Implementation is sponsored by DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving Yinmin Zhong and Shengyu Liu, Peking University; Junda Chen, UC San Diego; Jianbo Hu, Peking University; Yibo Zhu, StepFun; Xuanzhe Liu and Xin Jin, Peking University; Hao Zhang, UC San Diego DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving Yinmin Zhong1 Shengyu Liu1 Junda Chen3 Jianbo Hu1 Yibo Zhu2 Xuanzhe Liu1 Xin Jin1 Hao Zhang3 1School of Computer Science, Peking University 2StepFun 3UC San Diego Abstract DistServe improves the performance of large language mod-els (LLMs) serving by disaggregating the prefill and decoding computation.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

DBLP:journals/corr/abs-2404-09526

\cite{DBLP:journals/corr/abs-2404-09526}

LoongServe: Efficiently Serving Long-Context Large Language Models with Elastic Sequence Parallelism

http://arxiv.org/abs/2404.09526v2

The context window of large language models (LLMs) is rapidly increasing, leading to a huge variance in resource usage between different requests as well as between different phases of the same request. Restricted by static parallelism strategies, existing LLM serving systems cannot efficiently utilize the underlying resources to serve variable-length requests in different phases. To address this problem, we propose a new parallelism paradigm, elastic sequence parallelism (ESP), to elastically adapt to the variance between different requests and phases. Based on ESP, we design and build LoongServe, an LLM serving system that (1) improves computation efficiency by elastically adjusting the degree of parallelism in real-time, (2) improves communication efficiency by reducing key-value cache migration overhead and overlapping partial decoding communication with computation, and (3) improves GPU memory efficiency by reducing key-value cache fragmentation across instances. Our evaluation under diverse real-world datasets shows that LoongServe improves the maximum throughput by up to 3.85$\times$ compared to the chunked prefill and 5.81$\times$ compared to the prefill-decoding disaggregation.

true

Bingyang Wu and Shengyu Liu and Yinmin Zhong and Peng Sun and Xuanzhe Liu and Xin Jin

2,024

null

https://doi.org/10.48550/arXiv.2404.09526

10.48550/ARXIV.2404.09526

CoRR

LoongServe: Efficiently Serving Long-Context Large Language Models with Elastic Sequence Parallelism

LoongServe: Efficiently Serving Long-Context Large Language ...

https://colab.ws/articles/10.1145%2F3694715.3695948

LoongServe: Efficiently Serving Long-Context Large Language Models with Elastic Sequence Parallelism. Bingyang Wu 1. ,. Shengyu Liu 1. ,. Yinmin

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

DBLP:conf/sosp/KwonLZ0ZY0ZS23

\cite{DBLP:conf/sosp/KwonLZ0ZY0ZS23}

Efficient Memory Management for Large Language Model Serving with PagedAttention

http://arxiv.org/abs/2309.06180v1

High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. However, existing systems struggle because the key-value cache (KV cache) memory for each request is huge and grows and shrinks dynamically. When managed inefficiently, this memory can be significantly wasted by fragmentation and redundant duplication, limiting the batch size. To address this problem, we propose PagedAttention, an attention algorithm inspired by the classical virtual memory and paging techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce memory usage. Our evaluations show that vLLM improves the throughput of popular LLMs by 2-4$\times$ with the same level of latency compared to the state-of-the-art systems, such as FasterTransformer and Orca. The improvement is more pronounced with longer sequences, larger models, and more complex decoding algorithms. vLLM's source code is publicly available at https://github.com/vllm-project/vllm

true

Woosuk Kwon and Zhuohan Li and Siyuan Zhuang and Ying Sheng and Lianmin Zheng and Cody Hao Yu and Joseph Gonzalez and Hao Zhang and Ion Stoica

2,023

null

https://doi.org/10.1145/3600006.3613165

10.1145/3600006.3613165

null

Efficient Memory Management for Large Language Model Serving with PagedAttention

Efficient Memory Management for Large Language Model ...

https://arxiv.org/pdf/2309.06180

Efficient Memory Management for Large Language Model Serving with PagedAttention Woosuk Kwon 1,∗ Zhuohan Li 1,∗ Siyuan Zhuang 1 Ying Sheng 1,2 Lianmin Zheng 1 Cody Hao Yu 3 Joseph E. Gonzalez 1 Hao Zhang 4 Ion Stoica 1 1 UC Berkeley 2Stanford University 3Independent Researcher 4UC San Diego Abstract High throughput serving of large language models (LLMs) requires batching sufficiently many requests at a time. To address this problem, we propose PagedAttention, an attention al-gorithm inspired by the classical virtual memory and pag-ing techniques in operating systems. On top of it, we build vLLM, an LLM serving system that achieves (1) near-zero waste in KV cache memory and (2) flexible sharing of KV cache within and across requests to further reduce mem-ory usage. To address the above limitations, we propose PagedAt-tention , an attention algorithm inspired by the operating system’s (OS) solution to memory fragmentation and shar-ing: virtual memory with paging . In this work, we build vLLM , a high-throughput distributed LLM serving engine on top of PagedAttention that achieves near-zero waste in KV cache memory.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

alpaserve

\cite{alpaserve}

AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving

http://arxiv.org/abs/2302.11665v2

Model parallelism is conventionally viewed as a method to scale a single large deep learning model beyond the memory limits of a single device. In this paper, we demonstrate that model parallelism can be additionally used for the statistical multiplexing of multiple devices when serving multiple models, even when a single model can fit into a single device. Our work reveals a fundamental trade-off between the overhead introduced by model parallelism and the opportunity to exploit statistical multiplexing to reduce serving latency in the presence of bursty workloads. We explore the new trade-off space and present a novel serving system, AlpaServe, that determines an efficient strategy for placing and parallelizing collections of large deep learning models across a distributed cluster. Evaluation results on production workloads show that AlpaServe can process requests at up to 10x higher rates or 6x more burstiness while staying within latency constraints for more than 99% of requests.

true

Zhuohan Li and Lianmin Zheng and Yinmin Zhong and Vincent Liu and Ying Sheng and Xin Jin and Yanping Huang and Zhifeng Chen and Hao Zhang and Joseph E. Gonzalez and Ion Stoica

2,023

null

https://www.usenix.org/conference/osdi23/presentation/li-zhouhan

null

AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving

alpa-projects/mms: AlpaServe - GitHub

https://github.com/alpa-projects/mms

This is the official implementation of our OSDI'23 paper: AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving. To reproduce

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

DBLP:conf/osdi/YuJKKC22

\cite{DBLP:conf/osdi/YuJKKC22}

Orca: {A} Distributed Serving System for Transformer-Based Generative Models

null

true

false

Gyeong{-}In Yu and Joo Seong Jeong and Geon{-}Woo Kim and Soojeong Kim and Byung{-}Gon Chun

2,022

null

https://www.usenix.org/conference/osdi22/presentation/yu

null

Orca: {A} Distributed Serving System for Transformer-Based Generative Models

Orca: A Distributed Serving System for Transformer-Based ... - USENIX

https://www.usenix.org/conference/osdi22/presentation/yu

We have implemented a distributed serving system called ORCA, with additional designs for scalability to models with hundreds of billions of parameters.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

DBLP:conf/isca/PatelCZSGMB24

\cite{DBLP:conf/isca/PatelCZSGMB24}

Splitwise: Efficient generative LLM inference using phase splitting

http://arxiv.org/abs/2311.18677v2

Recent innovations in generative large language models (LLMs) have made their applications and use-cases ubiquitous. This has led to large-scale deployments of these models, using complex, expensive, and power-hungry AI accelerators, most commonly GPUs. These developments make LLM inference efficiency an important challenge. Based on our extensive characterization, we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a memory-intensive token generation, each with distinct latency, throughput, memory, and power characteristics. Despite state-of-the-art batching and scheduling, the token generation phase underutilizes compute resources. Specifically, unlike compute-intensive prompt computation phases, token generation phases do not require the compute capability of the latest GPUs, and can be run with lower power and cost. With Splitwise, we propose splitting the two phases of a LLM inference request on to separate machines. This allows us to use hardware that is well-suited for each phase, and provision resources independently per phase. However, splitting an inference request across machines requires state transfer from the machine running prompt computation over to the machine generating tokens. We implement and optimize this state transfer using the fast back-plane interconnects available in today's GPU clusters. We use the Splitwise technique to design LLM inference clusters using the same or different types of machines for the prompt computation and token generation phases. Our clusters are optimized for three key objectives: throughput, cost, and power. In particular, we show that we can achieve 1.4x higher throughput at 20% lower cost than current designs. Alternatively, we can achieve 2.35x more throughput with the same cost and power budgets.

true

Pratyush Patel and Esha Choukse and Chaojie Zhang and Aashaka Shah and {\'{I}}{\~{n}}igo Goiri and Saeed Maleki and Ricardo Bianchini

2,024

null

https://doi.org/10.1109/ISCA59077.2024.00019

10.1109/ISCA59077.2024.00019

null

Splitwise: Efficient generative LLM inference using phase splitting

http://arxiv.org/pdf/2311.18677v2

Recent innovations in generative large language models (LLMs) have made their applications and use-cases ubiquitous. This has led to large-scale deployments of these models, using complex, expensive, and power-hungry AI accelerators, most commonly GPUs. These developments make LLM inference efficiency an important challenge. Based on our extensive characterization, we find that there are two main phases during an LLM inference request: a compute-intensive prompt computation, and a memory-intensive token generation, each with distinct latency, throughput, memory, and power characteristics. Despite state-of-the-art batching and scheduling, the token generation phase underutilizes compute resources. Specifically, unlike compute-intensive prompt computation phases, token generation phases do not require the compute capability of the latest GPUs, and can be run with lower power and cost. With Splitwise, we propose splitting the two phases of a LLM inference request on to separate machines. This allows us to use hardware that is well-suited for each phase, and provision resources independently per phase. However, splitting an inference request across machines requires state transfer from the machine running prompt computation over to the machine generating tokens. We implement and optimize this state transfer using the fast back-plane interconnects available in today's GPU clusters. We use the Splitwise technique to design LLM inference clusters using the same or different types of machines for the prompt computation and token generation phases. Our clusters are optimized for three key objectives: throughput, cost, and power. In particular, we show that we can achieve 1.4x higher throughput at 20% lower cost than current designs. Alternatively, we can achieve 2.35x more throughput with the same cost and power budgets.

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

298501

\cite{298501}

{Cost-Efficient} Large Language Model Serving for Multi-turn Conversations with {CachedAttention}

null

true

false

Bin Gao and Zhuomin He and Puru Sharma and Qingxuan Kang and Djordje Jevdjic and Junbo Deng and Xingkun Yang and Zhou Yu and Pengfei Zuo

2,024

null

https://www.usenix.org/conference/atc24/presentation/gao-bin-cost

null

{Cost-Efficient} Large Language Model Serving for Multi-turn Conversations with {CachedAttention}

Cost-Efficient Large Language Model Serving for Multi-turn ... - arXiv

https://arxiv.org/abs/2403.19708

View a PDF of the paper titled Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttention, by Bin Gao and 8 other authors To address the problem, this paper proposes CachedAttention, a new attention mechanism that enables reuse of KV caches across multi-turn conversations, significantly reducing the repetitive computation overheads. View a PDF of the paper titled Cost-Efficient Large Language Model Serving for Multi-turn Conversations with CachedAttention, by Bin Gao and 8 other authors - [x] Bibliographic Explorer Toggle - [x] Connected Papers Toggle - [x] Litmaps Toggle - [x] alphaXiv Toggle - [x] Links to Code Toggle - [x] DagsHub Toggle - [x] GotitPub Toggle - [x] Links to Code Toggle - [x] ScienceCast Toggle - [x] Replicate Toggle

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

DBLP:journals/corr/abs-2412-17246

\cite{DBLP:journals/corr/abs-2412-17246}

Fast and Live Model Auto Scaling with {O(1)} Host Caching

null

true

false

Dingyan Zhang and Haotian Wang and Yang Liu and Xingda Wei and Yizhou Shan and Rong Chen and Haibo Chen

2,024

null

https://doi.org/10.48550/arXiv.2412.17246

10.48550/ARXIV.2412.17246

CoRR

Fast and Live Model Auto Scaling with {O(1)} Host Caching

Fast and Live Model Auto Scaling with 𝑂⁢(1) Host Caching

https://arxiv.org/html/2412.17246v1

Model autoscaling is the key mechanism to achieve serverless model-as-a-service, but it faces a fundamental trade-off between scaling speed and storage/memory

KVCache Cache in the Wild: Characterizing and Optimizing KVCache Cache at a Large Cloud Provider

2506.02634v1

shahrad2020serverless

\cite{shahrad2020serverless}

Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider

http://arxiv.org/abs/2003.03423v3

Function as a Service (FaaS) has been gaining popularity as a way to deploy computations to serverless backends in the cloud. This paradigm shifts the complexity of allocating and provisioning resources to the cloud provider, which has to provide the illusion of always-available resources (i.e., fast function invocations without cold starts) at the lowest possible resource cost. Doing so requires the provider to deeply understand the characteristics of the FaaS workload. Unfortunately, there has been little to no public information on these characteristics. Thus, in this paper, we first characterize the entire production FaaS workload of Azure Functions. We show for example that most functions are invoked very infrequently, but there is an 8-order-of-magnitude range of invocation frequencies. Using observations from our characterization, we then propose a practical resource management policy that significantly reduces the number of function coldstarts,while spending fewerresources than state-of-the-practice policies.

true

Mohammad Shahrad and Rodrigo Fonseca and Inigo Goiri and Gohar Chaudhry and Paul Batum and Jason Cooke and Eduardo Laureano and Colby Tresness and Mark Russinovich and Ricardo Bianchini

2,020

null

https://www.usenix.org/conference/atc20/presentation/shahrad

null

Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider

Characterizing and Optimizing the Serverless Workload at ...

https://www.usenix.org/system/files/atc20-shahrad.pdf

by M Shahrad · 2020 · Cited by 879 — This paper characterizes Azure Functions' serverless workload, showing most functions are invoked infrequently, and proposes a resource

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

liu2024:visual

\cite{liu2024:visual}

Visual Instruction Tuning

http://arxiv.org/abs/2304.08485v2

Instruction tuning large language models (LLMs) using machine-generated instruction-following data has improved zero-shot capabilities on new tasks, but the idea is less explored in the multimodal field. In this paper, we present the first attempt to use language-only GPT-4 to generate multimodal language-image instruction-following data. By instruction tuning on such generated data, we introduce LLaVA: Large Language and Vision Assistant, an end-to-end trained large multimodal model that connects a vision encoder and LLM for general-purpose visual and language understanding.Our early experiments show that LLaVA demonstrates impressive multimodel chat abilities, sometimes exhibiting the behaviors of multimodal GPT-4 on unseen images/instructions, and yields a 85.1% relative score compared with GPT-4 on a synthetic multimodal instruction-following dataset. When fine-tuned on Science QA, the synergy of LLaVA and GPT-4 achieves a new state-of-the-art accuracy of 92.53%. We make GPT-4 generated visual instruction tuning data, our model and code base publicly available.

true

Liu, Haotian and Li, Chunyuan and Wu, Qingyang and Lee, Yong Jae

2,024

null

Advances in neural information processing systems

Visual Instruction Tuning

http://arxiv.org/pdf/2304.08485v2

Instruction tuning large language models (LLMs) using machine-generated instruction-following data has improved zero-shot capabilities on new tasks, but the idea is less explored in the multimodal field. In this paper, we present the first attempt to use language-only GPT-4 to generate multimodal language-image instruction-following data. By instruction tuning on such generated data, we introduce LLaVA: Large Language and Vision Assistant, an end-to-end trained large multimodal model that connects a vision encoder and LLM for general-purpose visual and language understanding.Our early experiments show that LLaVA demonstrates impressive multimodel chat abilities, sometimes exhibiting the behaviors of multimodal GPT-4 on unseen images/instructions, and yields a 85.1% relative score compared with GPT-4 on a synthetic multimodal instruction-following dataset. When fine-tuned on Science QA, the synergy of LLaVA and GPT-4 achieves a new state-of-the-art accuracy of 92.53%. We make GPT-4 generated visual instruction tuning data, our model and code base publicly available.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

bai2023:qwen

\cite{bai2023:qwen}

Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond

null

true

false

Bai, Jinze and Bai, Shuai and Yang, Shusheng and Wang, Shijie and Tan, Sinan and Wang, Peng and Lin, Junyang and Zhou, Chang and Zhou, Jingren

2,023

null

Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond

Qwen-VL: A Versatile Vision-Language Model for Understanding...

https://openreview.net/forum?id=qrGjFJVl3m

Despite the effort in open-sourcing the model and its weights, the reviewers find QWEN-VL lacking in significant research contributions and technical novelty. * _**Open-source:**_ Qwen-VL is an open-sourced large vision-language model that excels in **(i)** achieving leading performance across a wide range of vision-language understanding and generation tasks, **(ii)** offering multi-lingual support, particularly in English and Chinese, **(iii)** accommodating multi-image and high-resolution inputs, and **(iv)** demonstrating fine-grained visual perception abilities, particularly in scene text-oriented visual question-answering and visual grounding. Unlike previous representative vision-language models like PaLI-X, which leverages proprietary in-house data and utilize publicly inaccessible model weights (_e.g._, ViT-22B), along with significantly high training costs, our Qwen-VL's training process is more practical and holds considerable referential significance for future research.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

chen2023:sharegpt4v

\cite{chen2023:sharegpt4v}

ShareGPT4V: Improving Large Multi-Modal Models with Better Captions

http://arxiv.org/abs/2311.12793v2

In the realm of large multi-modal models (LMMs), efficient modality alignment is crucial yet often constrained by the scarcity of high-quality image-text data. To address this bottleneck, we introduce the ShareGPT4V dataset, a pioneering large-scale resource featuring 1.2 million highly descriptive captions, which surpasses existing datasets in diversity and information content, covering world knowledge, object properties, spatial relationships, and aesthetic evaluations. Specifically, ShareGPT4V originates from a curated 100K high-quality captions collected from advanced GPT4-Vision and has been expanded to 1.2M with a superb caption model trained on this subset. ShareGPT4V first demonstrates its effectiveness for the Supervised Fine-Tuning (SFT) phase, by substituting an equivalent quantity of detailed captions in existing SFT datasets with a subset of our high-quality captions, significantly enhancing the LMMs like LLaVA-7B, LLaVA-1.5-13B, and Qwen-VL-Chat-7B on the MME and MMBench benchmarks, with respective gains of 222.8/22.0/22.3 and 2.7/1.3/1.5. We further incorporate ShareGPT4V data into both the pre-training and SFT phases, obtaining ShareGPT4V-7B, a superior LMM based on a simple architecture that has remarkable performance across a majority of the multi-modal benchmarks. This project is available at https://ShareGPT4V.github.io to serve as a pivotal resource for advancing the LMMs community.

true

Chen, Lin and Li, Jisong and Dong, Xiaoyi and Zhang, Pan and He, Conghui and Wang, Jiaqi and Zhao, Feng and Lin, Dahua

2,023

null

arXiv preprint arXiv:2311.12793

ShareGPT4V: Improving Large Multi-Modal Models with Better Captions

Improving Large Multi-Modal Models with Better Captions - arXiv

https://arxiv.org/abs/2311.12793

Image 4: arxiv logo>cs> arXiv:2311.12793 arXiv:2311.12793 (cs) View a PDF of the paper titled ShareGPT4V: Improving Large Multi-Modal Models with Better Captions, by Lin Chen and 7 other authors View a PDF of the paper titled ShareGPT4V: Improving Large Multi-Modal Models with Better Captions, by Lin Chen and 7 other authors - [x] Bibliographic Explorer Toggle - [x] Connected Papers Toggle - [x] Litmaps Toggle - [x] scite.ai Toggle - [x] alphaXiv Toggle - [x] Links to Code Toggle - [x] DagsHub Toggle - [x] GotitPub Toggle - [x] Huggingface Toggle - [x] Links to Code Toggle - [x] ScienceCast Toggle - [x] Replicate Toggle - [x] Spaces Toggle - [x] Spaces Toggle - [x] Core recommender toggle

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

li2023:videochat

\cite{li2023:videochat}

VideoChat: Chat-Centric Video Understanding

http://arxiv.org/abs/2305.06355v2

In this paper, we initiate an attempt of developing an end-to-end chat-centric video understanding system, coined as VideoChat. It integrates video foundation models and large language models via a learnable neural interface, excelling in spatiotemporal reasoning, event localization, and causal relationship inference. To instructively tune this system, we build a video-centric instruction dataset, composed of thousands of videos associated with detailed descriptions and conversations. This dataset emphasizes spatiotemporal reasoning and captures causal relationships, providing a valuable asset for training our chat-centric video understanding system. Preliminary qualitative experiments demonstrate the potential of our system across a broad spectrum of video applications, which could serve as a simple prototype system for future research on chat-centric video understanding. Access our code and data at https://github.com/OpenGVLab/Ask-Anything

true

Li, KunChang and He, Yinan and Wang, Yi and Li, Yizhuo and Wang, Wenhai and Luo, Ping and Wang, Yali and Wang, Limin and Qiao, Yu

2,023

null

arXiv preprint arXiv:2305.06355

VideoChat: Chat-Centric Video Understanding

VideoChat : Chat-Centric Video Understanding

https://img.shlab.org.cn/pjlab/files/2023/06/638215855649090000.pdf

by KC Li · 2023 · Cited by 853 — VideoChat is an end-to-end chat-centric video understanding system integrating video and large language models, excelling in spatiotemporal reasoning and

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

zhang2023:video

\cite{zhang2023:video}

Video-llama: An instruction-tuned audio-visual language model for video understanding

null

true

false

Zhang, Hang and Li, Xin and Bing, Lidong

2,023

null

arXiv preprint arXiv:2306.02858

Video-llama: An instruction-tuned audio-visual language model for video understanding

[EMNLP 2023 Demo] Video-LLaMA: An Instruction-tuned Audio ...

https://github.com/DAMO-NLP-SG/Video-LLaMA

[EMNLP 2023 Demo] Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding # Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding The following checkpoints are the full weights (visual encoder + audio encoder + Q-Formers + language decoder) to launch Video-LLaMA: Firstly, set the `llama_model` (for the path to the language decoder), `imagebind_ckpt_path` (for the path to the audio encoder), `ckpt` (for the path to VL branch) and `ckpt_2` (for the path to AL branch) in eval\_configs/video\_llama\_eval\_withaudio.yaml accordingly. The training of each cross-modal branch (i.e., VL branch or AL branch) in Video-LLaMA consists of two stages, title = {Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding}, [EMNLP 2023 Demo] Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

lu2024:unified

\cite{lu2024:unified}

Unified-IO 2: Scaling Autoregressive Multimodal Models with Vision, Language, Audio, and Action

http://arxiv.org/abs/2312.17172v1

We present Unified-IO 2, the first autoregressive multimodal model that is capable of understanding and generating image, text, audio, and action. To unify different modalities, we tokenize inputs and outputs -- images, text, audio, action, bounding boxes, etc., into a shared semantic space and then process them with a single encoder-decoder transformer model. Since training with such diverse modalities is challenging, we propose various architectural improvements to stabilize model training. We train our model from scratch on a large multimodal pre-training corpus from diverse sources with a multimodal mixture of denoisers objective. To learn an expansive set of skills, such as following multimodal instructions, we construct and finetune on an ensemble of 120 datasets with prompts and augmentations. With a single unified model, Unified-IO 2 achieves state-of-the-art performance on the GRIT benchmark and strong results in more than 35 benchmarks, including image generation and understanding, natural language understanding, video and audio understanding, and robotic manipulation. We release all our models to the research community.

true

Lu, Jiasen and Clark, Christopher and Lee, Sangho and Zhang, Zichen and Khosla, Savya and Marten, Ryan and Hoiem, Derek and Kembhavi, Aniruddha

2,024

null

Unified-IO 2: Scaling Autoregressive Multimodal Models with Vision, Language, Audio, and Action

Unified-IO 2: Scaling Autoregressive Multimodal Models with ...

https://openaccess.thecvf.com/content/CVPR2024/papers/Lu_Unified-IO_2_Scaling_Autoregressive_Multimodal_Models_with_Vision_Language_Audio_CVPR_2024_paper.pdf

by J Lu · 2024 · Cited by 210 — UNIFIED-IO 2 is a model that understands and generates image, text, audio, and action, using a single encoder-decoder model.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

achiam2023:gpt

\cite{achiam2023:gpt}

Gpt-4 technical report

null

true

false

Achiam, Josh and Adler, Steven and Agarwal, Sandhini and Ahmad, Lama and Akkaya, Ilge and Aleman, Florencia Leoni and Almeida, Diogo and Altenschmidt, Janko and Altman, Sam and Anadkat, Shyamal and others

2,023

null

arXiv preprint arXiv:2303.08774

Gpt-4 technical report

GPT-4 Technical Report

http://arxiv.org/pdf/2303.08774v6

We report the development of GPT-4, a large-scale, multimodal model which can accept image and text inputs and produce text outputs. While less capable than humans in many real-world scenarios, GPT-4 exhibits human-level performance on various professional and academic benchmarks, including passing a simulated bar exam with a score around the top 10% of test takers. GPT-4 is a Transformer-based model pre-trained to predict the next token in a document. The post-training alignment process results in improved performance on measures of factuality and adherence to desired behavior. A core component of this project was developing infrastructure and optimization methods that behave predictably across a wide range of scales. This allowed us to accurately predict some aspects of GPT-4's performance based on models trained with no more than 1/1,000th the compute of GPT-4.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

busso2008:iemocap

\cite{busso2008:iemocap}

IEMOCAP: Interactive emotional dyadic motion capture database

null

true

false

Busso, Carlos and Bulut, Murtaza and Lee, Chi-Chun and Kazemzadeh, Abe and Mower, Emily and Kim, Samuel and Chang, Jeannette N and Lee, Sungbok and Narayanan, Shrikanth S

2,008

null

Language resources and evaluation

IEMOCAP: Interactive emotional dyadic motion capture database

IEMOCAP- Home

https://sail.usc.edu/iemocap/

The Interactive Emotional Dyadic Motion Capture (IEMOCAP) database is an acted, multimodal and multispeaker database, recently collected at SAIL lab at USC.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

zadeh2018:multimodal

\cite{zadeh2018:multimodal}

Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph

null

true

false

Zadeh, AmirAli Bagher and Liang, Paul Pu and Poria, Soujanya and Cambria, Erik and Morency, Louis-Philippe

2,018

null

Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph

The MOSEI Dataset and Interpretable Dynamic Fusion

https://pliang279.github.io/papers/dap2018_mosei.pdf

by PP Liang · Cited by 30 — In this paper we introduce CMU-Multimodal Opinion. Sentiment and Emotion Intensity (CMU-. MOSEI), the largest dataset for multimodal sentiment analysis and

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

poria2019:meld

\cite{poria2019:meld}

MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations

http://arxiv.org/abs/1810.02508v6

Emotion recognition in conversations is a challenging task that has recently gained popularity due to its potential applications. Until now, however, a large-scale multimodal multi-party emotional conversational database containing more than two speakers per dialogue was missing. Thus, we propose the Multimodal EmotionLines Dataset (MELD), an extension and enhancement of EmotionLines. MELD contains about 13,000 utterances from 1,433 dialogues from the TV-series Friends. Each utterance is annotated with emotion and sentiment labels, and encompasses audio, visual and textual modalities. We propose several strong multimodal baselines and show the importance of contextual and multimodal information for emotion recognition in conversations. The full dataset is available for use at http:// affective-meld.github.io.

true

Poria, Soujanya and Hazarika, Devamanyu and Majumder, Navonil and Naik, Gautam and Cambria, Erik and Mihalcea, Rada

2,019

null

MELD: A Multimodal Multi-Party Dataset for Emotion Recognition in Conversations

MELD: A Multimodal Multi-Party Dataset for Emotion ...

https://github.com/declare-lab/MELD

* /data/MELD/train_sent_emo.csv - contains the utterances in the training set along with Sentiment and Emotion labels. * /data/MELD/dev_sent_emo.csv - contains the utterances in the dev set along with Sentiment and Emotion labels. * /data/MELD/test_sent_emo.csv - contains the utterances in the test set along with Sentiment and Emotion labels. * /data/MELD_Dyadic/train_sent_emo_dya.csv - contains the utterances in the training set of the dyadic variant of MELD along with Sentiment and Emotion labels. * /data/MELD_Dyadic/test_sent_emo_dya.csv - contains the utterances in the test set of the dyadic variant along with Sentiment and Emotion labels. Each utterance in a dialogue has been labeled by any of these seven emotions -- Neutral, Joyful, Peaceful, Powerful, Scared, Mad and Sad. The annotations are borrowed from the original dataset.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

han2023:champagne

\cite{han2023:champagne}

CHAMPAGNE: Learning Real-world Conversation from Large-Scale Web Videos

http://arxiv.org/abs/2303.09713v2

Visual information is central to conversation: body gestures and physical behaviour, for example, contribute to meaning that transcends words alone. To date, however, most neural conversational models are limited to just text. We introduce CHAMPAGNE, a generative model of conversations that can account for visual contexts. To train CHAMPAGNE, we collect and release YTD-18M, a large-scale corpus of 18M video-based dialogues. YTD-18M is constructed from web videos: crucial to our data collection pipeline is a pretrained language model that converts error-prone automatic transcripts to a cleaner dialogue format while maintaining meaning. Human evaluation reveals that YTD-18M is more sensible and specific than prior resources (MMDialog, 1M dialogues), while maintaining visual-groundedness. Experiments demonstrate that 1) CHAMPAGNE learns to conduct conversation from YTD-18M; and 2) when fine-tuned, it achieves state-of-the-art results on four vision-language tasks focused on real-world conversations. We release data, models, and code.

true

Han, Seungju and Hessel, Jack and Dziri, Nouha and Choi, Yejin and Yu, Youngjae

2,023

null

CHAMPAGNE: Learning Real-world Conversation from Large-Scale Web Videos

[PDF] Learning Real-world Conversation from Large-Scale Web Videos

https://openaccess.thecvf.com/content/ICCV2023/papers/Han_CHAMPAGNE_Learning_Real-world_Conversation_from_Large-Scale_Web_Videos_ICCV_2023_paper.pdf

Figure 1: CHAMPAGNE is a generative model of real-world conversational frames trained on. YTD-18M, a dataset of 18M video-based dialogues.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

park2024:let

\cite{park2024:let}

Let's Go Real Talk: Spoken Dialogue Model for Face-to-Face Conversation

http://arxiv.org/abs/2406.07867v2

In this paper, we introduce a novel Face-to-Face spoken dialogue model. It processes audio-visual speech from user input and generates audio-visual speech as the response, marking the initial step towards creating an avatar chatbot system without relying on intermediate text. To this end, we newly introduce MultiDialog, the first large-scale multimodal (i.e., audio and visual) spoken dialogue corpus containing 340 hours of approximately 9,000 dialogues, recorded based on the open domain dialogue dataset, TopicalChat. The MultiDialog contains parallel audio-visual recordings of conversation partners acting according to the given script with emotion annotations, which we expect to open up research opportunities in multimodal synthesis. Our Face-to-Face spoken dialogue model incorporates a textually pretrained large language model and adapts it into the audio-visual spoken dialogue domain by incorporating speech-text joint pretraining. Through extensive experiments, we validate the effectiveness of our model in facilitating a face-to-face conversation. Demo and data are available at https://multidialog.github.io and https://huggingface.co/datasets/IVLLab/MultiDialog, respectively.

true

Park, Se Jin and Kim, Chae Won and Rha, Hyeongseop and Kim, Minsu and Hong, Joanna and Yeo, Jeong Hun and Ro, Yong Man

2,024

null

arXiv preprint arXiv:2406.07867

Let's Go Real Talk: Spoken Dialogue Model for Face-to-Face Conversation

Let's Go Real Talk: Spoken Dialogue Model for Face-to-Face...

https://openreview.net/forum?id=zby4Ade9CCF

In this paper, we introduce a novel Face-to-Face spoken dialogue model. It processes audio-visual speech from user input and generates audio-visual speech as

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

shafique2023:nonverbal

\cite{shafique2023:nonverbal}

Nonverbal Communication Cue Recognition: A Pathway to More Accessible Communication

null

true

false

Shafique, Zoya and Wang, Haiyan and Tian, Yingli

2,023

null

Nonverbal Communication Cue Recognition: A Pathway to More Accessible Communication

[PDF] Nonverbal Communication Cue Recognition: A Pathway to More ...

https://openaccess.thecvf.com/content/CVPR2023W/WiCV/papers/Shafique_Nonverbal_Communication_Cue_Recognition_A_Pathway_to_More_Accessible_Communication_CVPRW_2023_paper.pdf

Nonverbal communication cues (NVCs) include body language, facial expressions, and hand gestures, conveying emotions and attitudes.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

zhang2023:learning

\cite{zhang2023:learning}

Learning Emotion Representations from Verbal and Nonverbal Communication

http://arxiv.org/abs/2305.13500v1

Emotion understanding is an essential but highly challenging component of artificial general intelligence. The absence of extensively annotated datasets has significantly impeded advancements in this field. We present EmotionCLIP, the first pre-training paradigm to extract visual emotion representations from verbal and nonverbal communication using only uncurated data. Compared to numerical labels or descriptions used in previous methods, communication naturally contains emotion information. Furthermore, acquiring emotion representations from communication is more congruent with the human learning process. We guide EmotionCLIP to attend to nonverbal emotion cues through subject-aware context encoding and verbal emotion cues using sentiment-guided contrastive learning. Extensive experiments validate the effectiveness and transferability of EmotionCLIP. Using merely linear-probe evaluation protocol, EmotionCLIP outperforms the state-of-the-art supervised visual emotion recognition methods and rivals many multimodal approaches across various benchmarks. We anticipate that the advent of EmotionCLIP will address the prevailing issue of data scarcity in emotion understanding, thereby fostering progress in related domains. The code and pre-trained models are available at https://github.com/Xeaver/EmotionCLIP.

true

Zhang, Sitao and Pan, Yimu and Wang, James Z

2,023

null

Learning Emotion Representations from Verbal and Nonverbal Communication

http://arxiv.org/pdf/2305.13500v1

Emotion understanding is an essential but highly challenging component of artificial general intelligence. The absence of extensively annotated datasets has significantly impeded advancements in this field. We present EmotionCLIP, the first pre-training paradigm to extract visual emotion representations from verbal and nonverbal communication using only uncurated data. Compared to numerical labels or descriptions used in previous methods, communication naturally contains emotion information. Furthermore, acquiring emotion representations from communication is more congruent with the human learning process. We guide EmotionCLIP to attend to nonverbal emotion cues through subject-aware context encoding and verbal emotion cues using sentiment-guided contrastive learning. Extensive experiments validate the effectiveness and transferability of EmotionCLIP. Using merely linear-probe evaluation protocol, EmotionCLIP outperforms the state-of-the-art supervised visual emotion recognition methods and rivals many multimodal approaches across various benchmarks. We anticipate that the advent of EmotionCLIP will address the prevailing issue of data scarcity in emotion understanding, thereby fostering progress in related domains. The code and pre-trained models are available at https://github.com/Xeaver/EmotionCLIP.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

cherakara2023:furchat

\cite{cherakara2023:furchat}

FurChat: An Embodied Conversational Agent using LLMs, Combining Open and Closed-Domain Dialogue with Facial Expressions

http://arxiv.org/abs/2308.15214v2

We demonstrate an embodied conversational agent that can function as a receptionist and generate a mixture of open and closed-domain dialogue along with facial expressions, by using a large language model (LLM) to develop an engaging conversation. We deployed the system onto a Furhat robot, which is highly expressive and capable of using both verbal and nonverbal cues during interaction. The system was designed specifically for the National Robotarium to interact with visitors through natural conversations, providing them with information about the facilities, research, news, upcoming events, etc. The system utilises the state-of-the-art GPT-3.5 model to generate such information along with domain-general conversations and facial expressions based on prompt engineering.

true

Cherakara, Neeraj and Varghese, Finny and Shabana, Sheena and Nelson, Nivan and Karukayil, Abhiram and Kulothungan, Rohith and Farhan, Mohammed Afil and Nesset, Birthe and Moujahid, Meriam and Dinkar, Tanvi and others

2,023

null

FurChat: An Embodied Conversational Agent using LLMs, Combining Open and Closed-Domain Dialogue with Facial Expressions

[PDF] FurChat: An Embodied Conversational Agent using LLMs ...

https://aclanthology.org/2023.sigdial-1.55.pdf

FurChat is an embodied conversational agent using LLMs, combining open and closed-domain dialogue with facial expressions, and can function as a receptionist.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

lee2023:developing

\cite{lee2023:developing}

Developing Social Robots with Empathetic Non-Verbal Cues Using Large Language Models

http://arxiv.org/abs/2308.16529v1

We propose augmenting the empathetic capacities of social robots by integrating non-verbal cues. Our primary contribution is the design and labeling of four types of empathetic non-verbal cues, abbreviated as SAFE: Speech, Action (gesture), Facial expression, and Emotion, in a social robot. These cues are generated using a Large Language Model (LLM). We developed an LLM-based conversational system for the robot and assessed its alignment with social cues as defined by human counselors. Preliminary results show distinct patterns in the robot's responses, such as a preference for calm and positive social emotions like 'joy' and 'lively', and frequent nodding gestures. Despite these tendencies, our approach has led to the development of a social robot capable of context-aware and more authentic interactions. Our work lays the groundwork for future studies on human-robot interactions, emphasizing the essential role of both verbal and non-verbal cues in creating social and empathetic robots.

true

Lee, Yoon Kyung and Jung, Yoonwon and Kang, Gyuyi and Hahn, Sowon

2,023

null

arXiv preprint arXiv:2308.16529

Developing Social Robots with Empathetic Non-Verbal Cues Using Large Language Models

Developing Social Robots with Empathetic Non-Verbal Cues Using ...

https://www.researchgate.net/publication/373552152_Developing_Social_Robots_with_Empathetic_Non-Verbal_Cues_Using_Large_Language_Models

We developed an LLM-based conversational system for the robot and assessed its alignment with social cues as defined by human counselors. Preliminary results

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

lin2023:one

\cite{lin2023:one}

One-Stage 3D Whole-Body Mesh Recovery with Component Aware Transformer

http://arxiv.org/abs/2303.16160v1

Whole-body mesh recovery aims to estimate the 3D human body, face, and hands parameters from a single image. It is challenging to perform this task with a single network due to resolution issues, i.e., the face and hands are usually located in extremely small regions. Existing works usually detect hands and faces, enlarge their resolution to feed in a specific network to predict the parameter, and finally fuse the results. While this copy-paste pipeline can capture the fine-grained details of the face and hands, the connections between different parts cannot be easily recovered in late fusion, leading to implausible 3D rotation and unnatural pose. In this work, we propose a one-stage pipeline for expressive whole-body mesh recovery, named OSX, without separate networks for each part. Specifically, we design a Component Aware Transformer (CAT) composed of a global body encoder and a local face/hand decoder. The encoder predicts the body parameters and provides a high-quality feature map for the decoder, which performs a feature-level upsample-crop scheme to extract high-resolution part-specific features and adopt keypoint-guided deformable attention to estimate hand and face precisely. The whole pipeline is simple yet effective without any manual post-processing and naturally avoids implausible prediction. Comprehensive experiments demonstrate the effectiveness of OSX. Lastly, we build a large-scale Upper-Body dataset (UBody) with high-quality 2D and 3D whole-body annotations. It contains persons with partially visible bodies in diverse real-life scenarios to bridge the gap between the basic task and downstream applications.

true

Lin, Jing and Zeng, Ailing and Wang, Haoqian and Zhang, Lei and Li, Yu

2,023

null

One-Stage 3D Whole-Body Mesh Recovery with Component Aware Transformer

IDEA-Research/OSX - GitHub

https://github.com/IDEA-Research/OSX

This repo is official PyTorch implementation of One-Stage 3D Whole-Body Mesh Recovery with Component Aware Transformer (CVPR2023). We propose the first one-

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

dwivedi2024:tokenhmr

\cite{dwivedi2024:tokenhmr}

TokenHMR: Advancing Human Mesh Recovery with a Tokenized Pose Representation

http://arxiv.org/abs/2404.16752v1

We address the problem of regressing 3D human pose and shape from a single image, with a focus on 3D accuracy. The current best methods leverage large datasets of 3D pseudo-ground-truth (p-GT) and 2D keypoints, leading to robust performance. With such methods, we observe a paradoxical decline in 3D pose accuracy with increasing 2D accuracy. This is caused by biases in the p-GT and the use of an approximate camera projection model. We quantify the error induced by current camera models and show that fitting 2D keypoints and p-GT accurately causes incorrect 3D poses. Our analysis defines the invalid distances within which minimizing 2D and p-GT losses is detrimental. We use this to formulate a new loss Threshold-Adaptive Loss Scaling (TALS) that penalizes gross 2D and p-GT losses but not smaller ones. With such a loss, there are many 3D poses that could equally explain the 2D evidence. To reduce this ambiguity we need a prior over valid human poses but such priors can introduce unwanted bias. To address this, we exploit a tokenized representation of human pose and reformulate the problem as token prediction. This restricts the estimated poses to the space of valid poses, effectively providing a uniform prior. Extensive experiments on the EMDB and 3DPW datasets show that our reformulated keypoint loss and tokenization allows us to train on in-the-wild data while improving 3D accuracy over the state-of-the-art. Our models and code are available for research at https://tokenhmr.is.tue.mpg.de.

true

Dwivedi, Sai Kumar and Sun, Yu and Patel, Priyanka and Feng, Yao and Black, Michael J

2,024

null

TokenHMR: Advancing Human Mesh Recovery with a Tokenized Pose Representation

TokenHMR: Advancing Human Mesh Recovery with a ...

https://github.com/saidwivedi/TokenHMR

Our method has two stages: Tokenization: The encoder maps continuous poses to discrete pose tokens. TokenHMR: During the training of human pose

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

danvevcek2022emoca

\cite{danvevcek2022emoca}

EMOCA: Emotion Driven Monocular Face Capture and Animation

http://arxiv.org/abs/2204.11312v1

As 3D facial avatars become more widely used for communication, it is critical that they faithfully convey emotion. Unfortunately, the best recent methods that regress parametric 3D face models from monocular images are unable to capture the full spectrum of facial expression, such as subtle or extreme emotions. We find the standard reconstruction metrics used for training (landmark reprojection error, photometric error, and face recognition loss) are insufficient to capture high-fidelity expressions. The result is facial geometries that do not match the emotional content of the input image. We address this with EMOCA (EMOtion Capture and Animation), by introducing a novel deep perceptual emotion consistency loss during training, which helps ensure that the reconstructed 3D expression matches the expression depicted in the input image. While EMOCA achieves 3D reconstruction errors that are on par with the current best methods, it significantly outperforms them in terms of the quality of the reconstructed expression and the perceived emotional content. We also directly regress levels of valence and arousal and classify basic expressions from the estimated 3D face parameters. On the task of in-the-wild emotion recognition, our purely geometric approach is on par with the best image-based methods, highlighting the value of 3D geometry in analyzing human behavior. The model and code are publicly available at https://emoca.is.tue.mpg.de.

true

Dan{\v{e}}{\v{c}}ek, Radek and Black, Michael J and Bolkart, Timo

2,022

null

EMOCA: Emotion Driven Monocular Face Capture and Animation

http://arxiv.org/pdf/2204.11312v1

As 3D facial avatars become more widely used for communication, it is critical that they faithfully convey emotion. Unfortunately, the best recent methods that regress parametric 3D face models from monocular images are unable to capture the full spectrum of facial expression, such as subtle or extreme emotions. We find the standard reconstruction metrics used for training (landmark reprojection error, photometric error, and face recognition loss) are insufficient to capture high-fidelity expressions. The result is facial geometries that do not match the emotional content of the input image. We address this with EMOCA (EMOtion Capture and Animation), by introducing a novel deep perceptual emotion consistency loss during training, which helps ensure that the reconstructed 3D expression matches the expression depicted in the input image. While EMOCA achieves 3D reconstruction errors that are on par with the current best methods, it significantly outperforms them in terms of the quality of the reconstructed expression and the perceived emotional content. We also directly regress levels of valence and arousal and classify basic expressions from the estimated 3D face parameters. On the task of in-the-wild emotion recognition, our purely geometric approach is on par with the best image-based methods, highlighting the value of 3D geometry in analyzing human behavior. The model and code are publicly available at https://emoca.is.tue.mpg.de.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

yi2023:generating

\cite{yi2023:generating}

Generating Holistic 3D Human Motion from Speech

http://arxiv.org/abs/2212.04420v2

This work addresses the problem of generating 3D holistic body motions from human speech. Given a speech recording, we synthesize sequences of 3D body poses, hand gestures, and facial expressions that are realistic and diverse. To achieve this, we first build a high-quality dataset of 3D holistic body meshes with synchronous speech. We then define a novel speech-to-motion generation framework in which the face, body, and hands are modeled separately. The separated modeling stems from the fact that face articulation strongly correlates with human speech, while body poses and hand gestures are less correlated. Specifically, we employ an autoencoder for face motions, and a compositional vector-quantized variational autoencoder (VQ-VAE) for the body and hand motions. The compositional VQ-VAE is key to generating diverse results. Additionally, we propose a cross-conditional autoregressive model that generates body poses and hand gestures, leading to coherent and realistic motions. Extensive experiments and user studies demonstrate that our proposed approach achieves state-of-the-art performance both qualitatively and quantitatively. Our novel dataset and code will be released for research purposes at https://talkshow.is.tue.mpg.de.

true

Yi, Hongwei and Liang, Hualin and Liu, Yifei and Cao, Qiong and Wen, Yandong and Bolkart, Timo and Tao, Dacheng and Black, Michael J

2,023

null

Generating Holistic 3D Human Motion from Speech

http://arxiv.org/pdf/2212.04420v2

This work addresses the problem of generating 3D holistic body motions from human speech. Given a speech recording, we synthesize sequences of 3D body poses, hand gestures, and facial expressions that are realistic and diverse. To achieve this, we first build a high-quality dataset of 3D holistic body meshes with synchronous speech. We then define a novel speech-to-motion generation framework in which the face, body, and hands are modeled separately. The separated modeling stems from the fact that face articulation strongly correlates with human speech, while body poses and hand gestures are less correlated. Specifically, we employ an autoencoder for face motions, and a compositional vector-quantized variational autoencoder (VQ-VAE) for the body and hand motions. The compositional VQ-VAE is key to generating diverse results. Additionally, we propose a cross-conditional autoregressive model that generates body poses and hand gestures, leading to coherent and realistic motions. Extensive experiments and user studies demonstrate that our proposed approach achieves state-of-the-art performance both qualitatively and quantitatively. Our novel dataset and code will be released for research purposes at https://talkshow.is.tue.mpg.de.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

wu2024:motionllm

\cite{wu2024:motionllm}

MotionLLM: Multimodal Motion-Language Learning with Large Language Models

null

true

false

Wu, Qi and Zhao, Yubo and Wang, Yifan and Tai, Yu-Wing and Tang, Chi-Keung

2,024

null

arXiv preprint arXiv:2405.17013

MotionLLM: Multimodal Motion-Language Learning with Large Language Models

(PDF) MotionLLM: Multimodal Motion-Language Learning ...

https://www.researchgate.net/publication/380906869_MotionLLM_Multimodal_Motion-Language_Learning_with_Large_Language_Models

MotionGPT-2 accommodates multiple motion-relevant tasks and supporting multimodal control conditions through pre-trained Large Language Models (

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

lu2023:humantomato

\cite{lu2023:humantomato}

HumanTOMATO: Text-aligned Whole-body Motion Generation

http://arxiv.org/abs/2310.12978v1

This work targets a novel text-driven whole-body motion generation task, which takes a given textual description as input and aims at generating high-quality, diverse, and coherent facial expressions, hand gestures, and body motions simultaneously. Previous works on text-driven motion generation tasks mainly have two limitations: they ignore the key role of fine-grained hand and face controlling in vivid whole-body motion generation, and lack a good alignment between text and motion. To address such limitations, we propose a Text-aligned whOle-body Motion generATiOn framework, named HumanTOMATO, which is the first attempt to our knowledge towards applicable holistic motion generation in this research area. To tackle this challenging task, our solution includes two key designs: (1) a Holistic Hierarchical VQ-VAE (aka H$^2$VQ) and a Hierarchical-GPT for fine-grained body and hand motion reconstruction and generation with two structured codebooks; and (2) a pre-trained text-motion-alignment model to help generated motion align with the input textual description explicitly. Comprehensive experiments verify that our model has significant advantages in both the quality of generated motions and their alignment with text.

true

Lu, Shunlin and Chen, Ling-Hao and Zeng, Ailing and Lin, Jing and Zhang, Ruimao and Zhang, Lei and Shum, Heung-Yeung

2,023

null

arXiv preprint arXiv:2310.12978

HumanTOMATO: Text-aligned Whole-body Motion Generation

HumanTOMATO: Text-aligned Whole-body Motion ...

https://lhchen.top/HumanTOMATO/

The proposed HumanTOMATO model can generate text-aligned whole-body motions with vivid and harmonious face, hand, and body motion.

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

ng2023:can

\cite{ng2023:can}

Can Language Models Learn to Listen?

http://arxiv.org/abs/2308.10897v1

We present a framework for generating appropriate facial responses from a listener in dyadic social interactions based on the speaker's words. Given an input transcription of the speaker's words with their timestamps, our approach autoregressively predicts a response of a listener: a sequence of listener facial gestures, quantized using a VQ-VAE. Since gesture is a language component, we propose treating the quantized atomic motion elements as additional language token inputs to a transformer-based large language model. Initializing our transformer with the weights of a language model pre-trained only on text results in significantly higher quality listener responses than training a transformer from scratch. We show that our generated listener motion is fluent and reflective of language semantics through quantitative metrics and a qualitative user study. In our evaluation, we analyze the model's ability to utilize temporal and semantic aspects of spoken text. Project page: https://people.eecs.berkeley.edu/~evonne_ng/projects/text2listen/

true

Ng, Evonne and Subramanian, Sanjay and Klein, Dan and Kanazawa, Angjoo and Darrell, Trevor and Ginosar, Shiry

2,023

null

Can Language Models Learn to Listen?

http://arxiv.org/pdf/2308.10897v1

We present a framework for generating appropriate facial responses from a listener in dyadic social interactions based on the speaker's words. Given an input transcription of the speaker's words with their timestamps, our approach autoregressively predicts a response of a listener: a sequence of listener facial gestures, quantized using a VQ-VAE. Since gesture is a language component, we propose treating the quantized atomic motion elements as additional language token inputs to a transformer-based large language model. Initializing our transformer with the weights of a language model pre-trained only on text results in significantly higher quality listener responses than training a transformer from scratch. We show that our generated listener motion is fluent and reflective of language semantics through quantitative metrics and a qualitative user study. In our evaluation, we analyze the model's ability to utilize temporal and semantic aspects of spoken text. Project page: https://people.eecs.berkeley.edu/~evonne_ng/projects/text2listen/

Speaking Beyond Language: A Large-Scale Multimodal Dataset for Learning Nonverbal Cues from Video-Grounded Dialogues

2506.00958v1

ng2022:learning

\cite{ng2022:learning}

Learning to Listen: Modeling Non-Deterministic Dyadic Facial Motion

http://arxiv.org/abs/2204.08451v1

We present a framework for modeling interactional communication in dyadic conversations: given multimodal inputs of a speaker, we autoregressively output multiple possibilities of corresponding listener motion. We combine the motion and speech audio of the speaker using a motion-audio cross attention transformer. Furthermore, we enable non-deterministic prediction by learning a discrete latent representation of realistic listener motion with a novel motion-encoding VQ-VAE. Our method organically captures the multimodal and non-deterministic nature of nonverbal dyadic interactions. Moreover, it produces realistic 3D listener facial motion synchronous with the speaker (see video). We demonstrate that our method outperforms baselines qualitatively and quantitatively via a rich suite of experiments. To facilitate this line of research, we introduce a novel and large in-the-wild dataset of dyadic conversations. Code, data, and videos available at https://evonneng.github.io/learning2listen/.

true

Ng, Evonne and Joo, Hanbyul and Hu, Liwen and Li, Hao and Darrell, Trevor and Kanazawa, Angjoo and Ginosar, Shiry

2,022

null

Learning to Listen: Modeling Non-Deterministic Dyadic Facial Motion

[PDF] Learning To Listen: Modeling Non-Deterministic Dyadic Facial Motion

https://openaccess.thecvf.com/content/CVPR2022/papers/Ng_Learning_To_Listen_Modeling_Non-Deterministic_Dyadic_Facial_Motion_CVPR_2022_paper.pdf

The method synthesizes listener motion from speaker video using a motion-audio transformer and a VQ-VAE, outputting multiple possibilities of listener motion.

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

strom2006expressive

\cite{strom2006expressive}

Expressive prosody for unit-selection speech synthesis.

null

true

false

Strom, Volker and Clark, Robert AJ and King, Simon

2,006

null

Expressive prosody for unit-selection speech synthesis.

Expressive Prosody for Unit-selection Speech Synthesis - CSTR

https://www.cstr.ed.ac.uk/downloads/publications/2006/strom06.pdf

by V Strom · Cited by 42 — The Festival unit selection speech synthesis system, Multisyn [1], achieves highly natural synthetic speech by avoiding use of an ex- plicit model of prosody in

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

ren2019fastspeech

\cite{ren2019fastspeech}

FastSpeech: Fast, Robust and Controllable Text to Speech

http://arxiv.org/abs/1905.09263v5

Neural network based end-to-end text to speech (TTS) has significantly improved the quality of synthesized speech. Prominent methods (e.g., Tacotron 2) usually first generate mel-spectrogram from text, and then synthesize speech from the mel-spectrogram using vocoder such as WaveNet. Compared with traditional concatenative and statistical parametric approaches, neural network based end-to-end models suffer from slow inference speed, and the synthesized speech is usually not robust (i.e., some words are skipped or repeated) and lack of controllability (voice speed or prosody control). In this work, we propose a novel feed-forward network based on Transformer to generate mel-spectrogram in parallel for TTS. Specifically, we extract attention alignments from an encoder-decoder based teacher model for phoneme duration prediction, which is used by a length regulator to expand the source phoneme sequence to match the length of the target mel-spectrogram sequence for parallel mel-spectrogram generation. Experiments on the LJSpeech dataset show that our parallel model matches autoregressive models in terms of speech quality, nearly eliminates the problem of word skipping and repeating in particularly hard cases, and can adjust voice speed smoothly. Most importantly, compared with autoregressive Transformer TTS, our model speeds up mel-spectrogram generation by 270x and the end-to-end speech synthesis by 38x. Therefore, we call our model FastSpeech.

true

Ren, Yi and Ruan, Yangjun and Tan, Xu and Qin, Tao and Zhao, Sheng and Zhao, Zhou and Liu, Tie-Yan

2,019

null

Advances in neural information processing systems

FastSpeech: Fast, Robust and Controllable Text to Speech

http://arxiv.org/pdf/1905.09263v5

Neural network based end-to-end text to speech (TTS) has significantly improved the quality of synthesized speech. Prominent methods (e.g., Tacotron 2) usually first generate mel-spectrogram from text, and then synthesize speech from the mel-spectrogram using vocoder such as WaveNet. Compared with traditional concatenative and statistical parametric approaches, neural network based end-to-end models suffer from slow inference speed, and the synthesized speech is usually not robust (i.e., some words are skipped or repeated) and lack of controllability (voice speed or prosody control). In this work, we propose a novel feed-forward network based on Transformer to generate mel-spectrogram in parallel for TTS. Specifically, we extract attention alignments from an encoder-decoder based teacher model for phoneme duration prediction, which is used by a length regulator to expand the source phoneme sequence to match the length of the target mel-spectrogram sequence for parallel mel-spectrogram generation. Experiments on the LJSpeech dataset show that our parallel model matches autoregressive models in terms of speech quality, nearly eliminates the problem of word skipping and repeating in particularly hard cases, and can adjust voice speed smoothly. Most importantly, compared with autoregressive Transformer TTS, our model speeds up mel-spectrogram generation by 270x and the end-to-end speech synthesis by 38x. Therefore, we call our model FastSpeech.

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

ren2020fastspeech

\cite{ren2020fastspeech}

FastSpeech 2: Fast and High-Quality End-to-End Text to Speech

http://arxiv.org/abs/2006.04558v8

Non-autoregressive text to speech (TTS) models such as FastSpeech can synthesize speech significantly faster than previous autoregressive models with comparable quality. The training of FastSpeech model relies on an autoregressive teacher model for duration prediction (to provide more information as input) and knowledge distillation (to simplify the data distribution in output), which can ease the one-to-many mapping problem (i.e., multiple speech variations correspond to the same text) in TTS. However, FastSpeech has several disadvantages: 1) the teacher-student distillation pipeline is complicated and time-consuming, 2) the duration extracted from the teacher model is not accurate enough, and the target mel-spectrograms distilled from teacher model suffer from information loss due to data simplification, both of which limit the voice quality. In this paper, we propose FastSpeech 2, which addresses the issues in FastSpeech and better solves the one-to-many mapping problem in TTS by 1) directly training the model with ground-truth target instead of the simplified output from teacher, and 2) introducing more variation information of speech (e.g., pitch, energy and more accurate duration) as conditional inputs. Specifically, we extract duration, pitch and energy from speech waveform and directly take them as conditional inputs in training and use predicted values in inference. We further design FastSpeech 2s, which is the first attempt to directly generate speech waveform from text in parallel, enjoying the benefit of fully end-to-end inference. Experimental results show that 1) FastSpeech 2 achieves a 3x training speed-up over FastSpeech, and FastSpeech 2s enjoys even faster inference speed; 2) FastSpeech 2 and 2s outperform FastSpeech in voice quality, and FastSpeech 2 can even surpass autoregressive models. Audio samples are available at https://speechresearch.github.io/fastspeech2/.

true

Ren, Yi and Hu, Chenxu and Tan, Xu and Qin, Tao and Zhao, Sheng and Zhao, Zhou and Liu, Tie-Yan

2,020

null

arXiv preprint arXiv:2006.04558

FastSpeech 2: Fast and High-Quality End-to-End Text to Speech

https://www.microsoft.com/en-us/research/lab/microsoft-research-asia/articles/fastspeech-2-fast-and-high-quality-end-to-end-text-to-speech/

FastSpeech 2 outperforms FastSpeech in voice quality and enjoys a much simpler training pipeline (3x training time reduction) while inheriting its advantages.

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

mohan2021ctrl

\cite{mohan2021ctrl}

Ctrl-P: Temporal control of prosodic variation for speech synthesis

null

true

false

Mohan, Devang S Ram and Hu, Vivian and Teh, Tian Huey and Torresquintero, Alexandra and Wallis, Christopher GR and Staib, Marlene and Foglianti, Lorenzo and Gao, Jiameng and King, Simon

2,021

null

arXiv preprint arXiv:2106.08352

Ctrl-P: Temporal control of prosodic variation for speech synthesis

Ctrl-P: Temporal Control of Prosodic Variation for Speech Synthesis

http://arxiv.org/pdf/2106.08352v1

Text does not fully specify the spoken form, so text-to-speech models must be able to learn from speech data that vary in ways not explained by the corresponding text. One way to reduce the amount of unexplained variation in training data is to provide acoustic information as an additional learning signal. When generating speech, modifying this acoustic information enables multiple distinct renditions of a text to be produced. Since much of the unexplained variation is in the prosody, we propose a model that generates speech explicitly conditioned on the three primary acoustic correlates of prosody: $F_{0}$, energy and duration. The model is flexible about how the values of these features are specified: they can be externally provided, or predicted from text, or predicted then subsequently modified. Compared to a model that employs a variational auto-encoder to learn unsupervised latent features, our model provides more interpretable, temporally-precise, and disentangled control. When automatically predicting the acoustic features from text, it generates speech that is more natural than that from a Tacotron 2 model with reference encoder. Subsequent human-in-the-loop modification of the predicted acoustic features can significantly further increase naturalness.

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

bandekar2023speaking

\cite{bandekar2023speaking}

Speaking rate attention-based duration prediction for speed control TTS

http://arxiv.org/abs/2310.08846v1

With the advent of high-quality speech synthesis, there is a lot of interest in controlling various prosodic attributes of speech. Speaking rate is an essential attribute towards modelling the expressivity of speech. In this work, we propose a novel approach to control the speaking rate for non-autoregressive TTS. We achieve this by conditioning the speaking rate inside the duration predictor, allowing implicit speaking rate control. We show the benefits of this approach by synthesising audio at various speaking rate factors and measuring the quality of speaking rate-controlled synthesised speech. Further, we study the effect of the speaking rate distribution of the training data towards effective rate control. Finally, we fine-tune a baseline pretrained TTS model to obtain speaking rate control TTS. We provide various analyses to showcase the benefits of using this proposed approach, along with objective as well as subjective metrics. We find that the proposed methods have higher subjective scores and lower speaker rate errors across many speaking rate factors over the baseline.

true

Bandekar, Jesuraj and Udupa, Sathvik and Singh, Abhayjeet and Jayakumar, Anjali and Badiger, Sandhya and Kumar, Saurabh and VH, Pooja and Ghosh, Prasanta Kumar and others

2,023

null

arXiv preprint arXiv:2310.08846

Speaking rate attention-based duration prediction for speed control TTS

Speaking Rate Control of end-to-end TTS Models by Direct ...

https://www.isca-archive.org/interspeech_2022/lenglet22_interspeech.pdf

by M Lenglet · 2022 · Cited by 8 — Evaluation was performed on the control of speaking rate on both attention-based (TC) and duration predictor based (FS) methods. Objective analyses showed

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

wang2018style

\cite{wang2018style}

Style Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech Synthesis

http://arxiv.org/abs/1803.09017v1

In this work, we propose "global style tokens" (GSTs), a bank of embeddings that are jointly trained within Tacotron, a state-of-the-art end-to-end speech synthesis system. The embeddings are trained with no explicit labels, yet learn to model a large range of acoustic expressiveness. GSTs lead to a rich set of significant results. The soft interpretable "labels" they generate can be used to control synthesis in novel ways, such as varying speed and speaking style - independently of the text content. They can also be used for style transfer, replicating the speaking style of a single audio clip across an entire long-form text corpus. When trained on noisy, unlabeled found data, GSTs learn to factorize noise and speaker identity, providing a path towards highly scalable but robust speech synthesis.

true

Wang, Yuxuan and Stanton, Daisy and Zhang, Yu and Ryan, RJ-Skerry and Battenberg, Eric and Shor, Joel and Xiao, Ying and Jia, Ye and Ren, Fei and Saurous, Rif A

2,018

null

Style Tokens: Unsupervised Style Modeling, Control and Transfer in End-to-End Speech Synthesis

Unsupervised Style Modeling, Control and Transfer in End- ...

https://research.google/pubs/style-tokens-unsupervised-style-modeling-control-and-transfer-in-end-to-end-speech-synthesis/

by Y Wang · Cited by 1080 — In this work, we propose “global style tokens”(GSTs), a bank of embeddings that are jointly trained within Tacotron, a state-of-the-art end-to-end speech

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

skerry2018towards

\cite{skerry2018towards}

Towards End-to-End Prosody Transfer for Expressive Speech Synthesis with Tacotron

http://arxiv.org/abs/1803.09047v1

We present an extension to the Tacotron speech synthesis architecture that learns a latent embedding space of prosody, derived from a reference acoustic representation containing the desired prosody. We show that conditioning Tacotron on this learned embedding space results in synthesized audio that matches the prosody of the reference signal with fine time detail even when the reference and synthesis speakers are different. Additionally, we show that a reference prosody embedding can be used to synthesize text that is different from that of the reference utterance. We define several quantitative and subjective metrics for evaluating prosody transfer, and report results with accompanying audio samples from single-speaker and 44-speaker Tacotron models on a prosody transfer task.

true

Skerry-Ryan, RJ and Battenberg, Eric and Xiao, Ying and Wang, Yuxuan and Stanton, Daisy and Shor, Joel and Weiss, Ron and Clark, Rob and Saurous, Rif A

2,018

null

Towards End-to-End Prosody Transfer for Expressive Speech Synthesis with Tacotron

[PDF] Towards End-to-End Prosody Transfer for Expressive Speech ...

https://proceedings.mlr.press/v80/skerry-ryan18a/skerry-ryan18a.pdf

Abstract. We present an extension to the Tacotron speech synthesis architecture that learns a latent embed- ding space of prosody, derived from a reference.

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

hsu2018hierarchical

\cite{hsu2018hierarchical}

Hierarchical Generative Modeling for Controllable Speech Synthesis

http://arxiv.org/abs/1810.07217v2

This paper proposes a neural sequence-to-sequence text-to-speech (TTS) model which can control latent attributes in the generated speech that are rarely annotated in the training data, such as speaking style, accent, background noise, and recording conditions. The model is formulated as a conditional generative model based on the variational autoencoder (VAE) framework, with two levels of hierarchical latent variables. The first level is a categorical variable, which represents attribute groups (e.g. clean/noisy) and provides interpretability. The second level, conditioned on the first, is a multivariate Gaussian variable, which characterizes specific attribute configurations (e.g. noise level, speaking rate) and enables disentangled fine-grained control over these attributes. This amounts to using a Gaussian mixture model (GMM) for the latent distribution. Extensive evaluation demonstrates its ability to control the aforementioned attributes. In particular, we train a high-quality controllable TTS model on real found data, which is capable of inferring speaker and style attributes from a noisy utterance and use it to synthesize clean speech with controllable speaking style.

true

Hsu, Wei-Ning and Zhang, Yu and Weiss, Ron J and Zen, Heiga and Wu, Yonghui and Wang, Yuxuan and Cao, Yuan and Jia, Ye and Chen, Zhifeng and Shen, Jonathan and others

2,018

null

arXiv preprint arXiv:1810.07217

Hierarchical Generative Modeling for Controllable Speech Synthesis

http://arxiv.org/pdf/1810.07217v2

This paper proposes a neural sequence-to-sequence text-to-speech (TTS) model which can control latent attributes in the generated speech that are rarely annotated in the training data, such as speaking style, accent, background noise, and recording conditions. The model is formulated as a conditional generative model based on the variational autoencoder (VAE) framework, with two levels of hierarchical latent variables. The first level is a categorical variable, which represents attribute groups (e.g. clean/noisy) and provides interpretability. The second level, conditioned on the first, is a multivariate Gaussian variable, which characterizes specific attribute configurations (e.g. noise level, speaking rate) and enables disentangled fine-grained control over these attributes. This amounts to using a Gaussian mixture model (GMM) for the latent distribution. Extensive evaluation demonstrates its ability to control the aforementioned attributes. In particular, we train a high-quality controllable TTS model on real found data, which is capable of inferring speaker and style attributes from a noisy utterance and use it to synthesize clean speech with controllable speaking style.

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

lenglet2022speaking

\cite{lenglet2022speaking}

Speaking Rate Control of end-to-end TTS Models by Direct Manipulation of the Encoder's Output Embeddings

null

true

false

Lenglet, Martin and Perrotin, Olivier and Bailly, G{\'e}rard

2,022

null

Speaking Rate Control of end-to-end TTS Models by Direct Manipulation of the Encoder's Output Embeddings

Speaking Rate Control of end-to-end TTS Models by ... - ISCA Archive

https://www.isca-archive.org/interspeech_2022/lenglet22_interspeech.html

Experimental results show that the control provided by embeddings reproduces a behaviour closer to natural speech data.

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

zhang2020unified

\cite{zhang2020unified}

Unified Mandarin TTS Front-end Based on Distilled BERT Model

http://arxiv.org/abs/2012.15404v1

The front-end module in a typical Mandarin text-to-speech system (TTS) is composed of a long pipeline of text processing components, which requires extensive efforts to build and is prone to large accumulative model size and cascade errors. In this paper, a pre-trained language model (PLM) based model is proposed to simultaneously tackle the two most important tasks in TTS front-end, i.e., prosodic structure prediction (PSP) and grapheme-to-phoneme (G2P) conversion. We use a pre-trained Chinese BERT[1] as the text encoder and employ multi-task learning technique to adapt it to the two TTS front-end tasks. Then, the BERT encoder is distilled into a smaller model by employing a knowledge distillation technique called TinyBERT[2], making the whole model size 25% of that of benchmark pipeline models while maintaining competitive performance on both tasks. With the proposed the methods, we are able to run the whole TTS front-end module in a light and unified manner, which is more friendly to deployment on mobile devices.

true

Zhang, Yang and Deng, Liqun and Wang, Yasheng

2,020

null

arXiv preprint arXiv:2012.15404

Unified Mandarin TTS Front-end Based on Distilled BERT Model

https://arxiv.org/abs/2012.15404

We use a pre-trained Chinese BERT[1] as the text encoder and employ multi-task learning technique to adapt it to the two TTS front-end tasks.

Counterfactual Activation Editing for Post-hoc Prosody and Mispronunciation Correction in TTS Models

2506.00832v1

fong2022speech

\cite{fong2022speech}

Speech Audio Corrector: using speech from non-target speakers for one-off correction of mispronunciations in grapheme-input text-to-speech

null

true

false

Fong, Jason and Lyth, Daniel and Henter, Gustav Eje and Tang, Hao and King, Simon

2,022

null

Speech Audio Corrector: using speech from non-target speakers for one-off correction of mispronunciations in grapheme-input text-to-speech

[PDF] using speech from non-target speakers for one-off correction of ...

https://www.research.ed.ac.uk/files/364801102/Speech_Audio_Corrector_FONG_DOA13062022_VOR.pdf

Missing: 04/08/2025

Dual Debiasing for Noisy In-Context Learning for Text Generation

2506.00418v1

yoo2022ground

\cite{yoo2022ground}

Ground-Truth Labels Matter: A Deeper Look into Input-Label Demonstrations

http://arxiv.org/abs/2205.12685v2

Despite recent explosion of interests in in-context learning, the underlying mechanism and the precise impact of the quality of demonstrations remain elusive. Intuitively, ground-truth labels should have as much impact in in-context learning (ICL) as supervised learning, but recent work reported that the input-label correspondence is significantly less important than previously thought. Intrigued by this counter-intuitive observation, we re-examine the importance of ground-truth labels in in-context learning. With the introduction of two novel metrics, namely Label-Correctness Sensitivity and Ground-truth Label Effect Ratio (GLER), we were able to conduct quantifiable analysis on the impact of ground-truth label demonstrations. Through extensive analyses, we find that the correct input-label mappings can have varying impacts on the downstream in-context learning performances, depending on the experimental configuration. Through additional studies, we identify key components, such as the verbosity of prompt templates and the language model size, as the controlling factor to achieve more noise-resilient ICL.

true

Yoo, Kang Min and Kim, Junyeob and Kim, Hyuhng Joon and Cho, Hyunsoo and Jo, Hwiyeol and Lee, Sang-Woo and Lee, Sang-goo and Kim, Taeuk

2,022

null

Ground-Truth Labels Matter: A Deeper Look into Input-Label Demonstrations

Ground-Truth Labels Matter: A Deeper Look into Input- ...

https://aclanthology.org/2022.emnlp-main.155.pdf

by KM Yoo · 2022 · Cited by 100 — We propose two new quantifiable metrics, sen- sitivity and GLER, to measure the impact of ground-truth label demonstrations on ICL. • We conduct

Dual Debiasing for Noisy In-Context Learning for Text Generation

2506.00418v1

o2023contrastive

\cite{o2023contrastive}

Contrastive Decoding Improves Reasoning in Large Language Models

http://arxiv.org/abs/2309.09117v2

We demonstrate that Contrastive Decoding -- a simple, computationally light, and training-free text generation method proposed by Li et al 2022 -- achieves large out-of-the-box improvements over greedy decoding on a variety of reasoning tasks. Originally shown to improve the perceived quality of long-form text generation, Contrastive Decoding searches for strings that maximize a weighted difference in likelihood between strong and weak models. We show that Contrastive Decoding leads LLaMA-65B to outperform LLaMA 2, GPT-3.5 and PaLM 2-L on the HellaSwag commonsense reasoning benchmark, and to outperform LLaMA 2, GPT-3.5 and PaLM-540B on the GSM8K math word reasoning benchmark, in addition to improvements on a collection of other tasks. Analysis suggests that Contrastive Decoding improves over existing methods by preventing some abstract reasoning errors, as well as by avoiding simpler modes such as copying sections of the input during chain-of-thought. Overall, Contrastive Decoding outperforms nucleus sampling for long-form generation and greedy decoding for reasoning tasks, making it a powerful general purpose method for generating text from language models.

true

O'Brien, Sean and Lewis, Mike

2,023

null

arXiv preprint arXiv:2309.09117

Contrastive Decoding Improves Reasoning in Large Language Models

http://arxiv.org/pdf/2309.09117v2

We demonstrate that Contrastive Decoding -- a simple, computationally light, and training-free text generation method proposed by Li et al 2022 -- achieves large out-of-the-box improvements over greedy decoding on a variety of reasoning tasks. Originally shown to improve the perceived quality of long-form text generation, Contrastive Decoding searches for strings that maximize a weighted difference in likelihood between strong and weak models. We show that Contrastive Decoding leads LLaMA-65B to outperform LLaMA 2, GPT-3.5 and PaLM 2-L on the HellaSwag commonsense reasoning benchmark, and to outperform LLaMA 2, GPT-3.5 and PaLM-540B on the GSM8K math word reasoning benchmark, in addition to improvements on a collection of other tasks. Analysis suggests that Contrastive Decoding improves over existing methods by preventing some abstract reasoning errors, as well as by avoiding simpler modes such as copying sections of the input during chain-of-thought. Overall, Contrastive Decoding outperforms nucleus sampling for long-form generation and greedy decoding for reasoning tasks, making it a powerful general purpose method for generating text from language models.

Dual Debiasing for Noisy In-Context Learning for Text Generation

2506.00418v1

li2023unified

\cite{li2023unified}

Unified Demonstration Retriever for In-Context Learning

http://arxiv.org/abs/2305.04320v2

In-context learning is a new learning paradigm where a language model conditions on a few input-output pairs (demonstrations) and a test input, and directly outputs the prediction. It has been shown highly dependent on the provided demonstrations and thus promotes the research of demonstration retrieval: given a test input, relevant examples are retrieved from the training set to serve as informative demonstrations for in-context learning. While previous works focus on training task-specific retrievers for several tasks separately, these methods are often hard to transfer and scale on various tasks, and separately trained retrievers incur a lot of parameter storage and deployment cost. In this paper, we propose Unified Demonstration Retriever (\textbf{UDR}), a single model to retrieve demonstrations for a wide range of tasks. To train UDR, we cast various tasks' training signals into a unified list-wise ranking formulation by language model's feedback. Then we propose a multi-task list-wise ranking training framework, with an iterative mining strategy to find high-quality candidates, which can help UDR fully incorporate various tasks' signals. Experiments on 30+ tasks across 13 task families and multiple data domains show that UDR significantly outperforms baselines. Further analyses show the effectiveness of each proposed component and UDR's strong ability in various scenarios including different LMs (1.3B - 175B), unseen datasets, varying demonstration quantities, etc.

true

Li, Xiaonan and Lv, Kai and Yan, Hang and Lin, Tianyang and Zhu, Wei and Ni, Yuan and Xie, Guotong and Wang, Xiaoling and Qiu, Xipeng

2,023

null

Unified Demonstration Retriever for In-Context Learning

https://aclanthology.org/2023.acl-long.256/

In this paper, we propose Unified Demonstration Retriever (UDR), a single model to retrieve demonstrations for a wide range of tasks.

Dual Debiasing for Noisy In-Context Learning for Text Generation

2506.00418v1

liucontext

\cite{liucontext}

In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering

http://arxiv.org/abs/2311.06668v3

Large language models (LLMs) demonstrate emergent in-context learning capabilities, where they adapt to new tasks based on example demonstrations. However, in-context learning has seen limited effectiveness in many settings, is difficult to quantitatively control and takes up context window space. To overcome these limitations, we propose an alternative approach that recasts in-context learning as in-context vectors (ICV). Using ICV has two steps. We first use a forward pass on demonstration examples to create the in-context vector from the latent embedding of the LLM. This vector captures essential information about the intended task. On a new query, instead of adding demonstrations to the prompt, we shift the latent states of the LLM using the ICV. The ICV approach has several benefits: 1) it enables the LLM to more effectively follow the demonstration examples; 2) it's easy to control by adjusting the magnitude of the ICV; 3) it reduces the length of the prompt by removing the in-context demonstrations; 4) ICV is computationally much more efficient than fine-tuning. We demonstrate that ICV achieves better performance compared to standard in-context learning and fine-tuning on diverse tasks including safety, style transfer, role-playing and formatting. Moreover, we show that we can flexibly teach LLM to simultaneously follow different types of instructions by simple vector arithmetics on the corresponding ICVs.

true

Liu, Sheng and Ye, Haotian and Xing, Lei and Zou, James Y

null

In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steering

Making In Context Learning More Effective and ...

https://consensus.app/papers/incontext-vectors-making-in-context-learning-more-zou-liu/20a28c8387155fa1ac876aad9841f1ee

Key takeaway: 'In-context vectors (ICV) improve in-context learning effectiveness, controllability, and computational efficiency in large