AI Can Learn Scientific Taste
This work addresses the underexplored challenge of improving AI's scientific taste for advancing toward human-level AI scientists, representing a novel but incremental step in AI research.
The paper tackled the problem of enhancing AI's scientific taste, defined as the capacity to judge and propose high-impact research ideas, by proposing Reinforcement Learning from Community Feedback (RLCF) with Scientific Judge and Scientific Thinker models. Experiments showed Scientific Judge outperformed SOTA LLMs like GPT-5.2 and generalized to future-year tests and unseen fields, while Scientific Thinker proposed ideas with higher potential impact than baselines.
Great scientists have strong judgement and foresight, closely tied to what we call scientific taste. Here, we use the term to refer to the capacity to judge and propose research ideas with high potential impact. However, most relative research focuses on improving an AI scientist's executive capability, while enhancing an AI's scientific taste remains underexplored. In this work, we propose Reinforcement Learning from Community Feedback (RLCF), a training paradigm that uses large-scale community signals as supervision, and formulate scientific taste learning as a preference modeling and alignment problem. For preference modeling, we train Scientific Judge on 700K field- and time-matched pairs of high- vs. low-citation papers to judge ideas. For preference alignment, using Scientific Judge as a reward model, we train a policy model, Scientific Thinker, to propose research ideas with high potential impact. Experiments show Scientific Judge outperforms SOTA LLMs (e.g., GPT-5.2, Gemini 3 Pro) and generalizes to future-year test, unseen fields, and peer-review preference. Furthermore, Scientific Thinker proposes research ideas with higher potential impact than baselines. Our findings show that AI can learn scientific taste, marking a key step toward reaching human-level AI scientists.