ReCrit: Transition-Aware Reinforcement Learning for Scientific Critic Reasoning

In three linesReCrit is a reinforcement learning framework improving LLM handling of user criticism in scientific reasoning. It decomposes behaviors into four quadrants (Correction, Sycophancy, Robustness, Boundary) using transition-aware rewards. On ChemBench, TRQA, and EarthSE, ReCrit improves accuracy from 38.15% to 51.49% on Qwen3.5-4B.

Summary generated by Claude — human-verified