Paper summary: AI Can Learn Scientific Taste
Reinforcement Learning from Community Feedback trains a "Scientific Judge" on signals like citations, and a "Scientific Thinker" that proposes research ideas; the judge generalises to future years and unseen fields.
Problem
Scientific taste is the ability to judge and propose ideas with long-term impact. It’s concentrated among experienced researchers in narrow fields. Can AI learn it?
Method
Reinforcement Learning from Community Feedback (RLCF):
- Scientific Judge learns from community feedback such as citations
- Scientific Thinker learns to propose research ideas with high potential impact
Key results
From the abstract: the Judge beats strong LLM baselines, and its judgement generalises to future-year papers, other community metrics and unseen fields. The Thinker proposes ideas with higher potential impact than baselines.
Limitations
From the paper:
- scientific taste may involve more than judgement and ideation, such as experimental feasibility and diversity of ideas
- citations are an imperfect signal: some important papers are cited late, and citation data carries biases that controlled evaluations can’t fully remove
- the field categories are coarse
Sources
- AI Can Learn Scientific Taste (arXiv:2603.14473) — arXiv
- Hugging Face Daily Papers, 2026-03-17 — Hugging Face
- AI Can Learn Scientific Taste (arXiv HTML full text) — arXiv
- Code repository (linked in the paper) — GitHub
This is a summary written from the linked public sources. Figures and numbers are as reported there.