Paper 2603.14473
AI Can Learn Scientific Taste
- Published
- Mar 2026
- Research lab
- Independent
- Citations
- 4
- GitHub
- 428 stars
01 In brief
Summary
The paper introduces Reinforcement Learning from Community Feedback (RLCF), a paradigm that uses citation-based community signals to train AI models for scientific taste—the ability to judge and propose high-impact research ideas.
They built SciJudgeBench with 720K field- and time-matched paper abstract pairs, trained Scientific Judge via GRPO to predict citation preferences, and then used it as a reward model to train Scientific Thinker via Comparison-Based GRPO for ideation.
The 30B Scientific Judge achieved 82.7% accuracy on SciJudgeBench, surpassing strong baselines like GPT-5.4 Thinking, and generalized to future-year papers, unseen fields, and other metrics (ICLR review scores, Altmetric).
Scientific Thinker outperformed base policies with win rates up to 81.5% in-domain and 83.0% out-of-domain, and beat strong LLM baselines in ideation.
Results suggest AI can learn scientific taste from community feedback, aiding scientific discovery.
Limitations include imperfect citation signals and reliance on LLM evaluators for ideation quality.
02 From the paper
Abstract
Scientific discovery depends on expert judgement and foresight, which we call scientific taste: the ability to judge and propose research ideas with potential for long-term scientific impact. Whether AI can learn this ability remains an open question. Here we provide evidence that artificial intelligence can learn judgement and ideation. We introduce Reinforcement Learning from Community Feedback (RLCF), a training paradigm that uses large-scale signals from scientific community as supervision. We first train Scientific Judge on field- and time-matched pairs of high- vs. low-citation papers to judge ideas. We then train a Scientific Thinker, to propose research ideas with high potential impact. Experiments show that the 30B Scientific Judge variant outperforms strong LLM baselines (e.g., GPT-5.4 Thinking), while Scientific Judge generalizes across future-year papers, unseen fields, and other community metrics. Furthermore, Scientific Thinker proposes research ideas with higher potential impact than baselines. These results suggest that AI can learn scientific taste, marking an important step towards AI systems that could help accelerate scientific discovery.