TMLR Group

Co-Reward is a self-supervised reinforcement learning method for LLM reasoning, which leverages contrastive agreement between original and rephrased q

Can Language Models Perform Robust Reasoning in Chain-of-thought Prompting with Noisy Rationales?