RLCF: AI Trained on 700K Citation-Matched Paper Pairs to Judge Research Quality — Outperforms GPT-5.2 on Idea Evaluation

arXiv / Fudan University / OpenMOSS·medium signal

Fudan University / OpenMOSS proposes Reinforcement Learning from Community Feedback (RLCF): a 'Scientific Judge' model trained on 700,000 field- and time-matched high vs. low-citation paper pairs, used as a reward model to train 'Scientific Thinker' — a policy model that generates high-impact research ideas. The Scientific Judge outperforms GPT-5.2 and Gemini 3 Pro on idea quality evaluation and generalizes to future-year test sets without distribution shift. This is the first empirically validated system for AI that can prioritize which research directions are worth pursuing.