- [2026.07] 📊 The 2025 Future-Year benchmark and evaluation results have been refreshed (904 pairs).
- [2026.07] 🤗 The SciJudge-2506 series has been updated on Hugging Face. [Collection]
- [2026.03] 🎮 Online Demo is now available! [Demo]
- [2026.03] 📄 Paper available on arXiv. [arXiv]
- [2026.03] 🤗 Models released on Hugging Face. [Collection]
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 the 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.
RLCF consists of three stages:
- Construct community preference Citations are converted into pairwise preference signals by matching papers within the same field and publication period.
- Preference modeling with Scientific Judge We train a generative reward model with GRPO that reasons over a pair of paper abstracts and predicts which one is more likely to have higher impact.
- Preference alignment with Scientific Thinker We use Scientific Judge as a reward model and optimize a policy model with comparison-based GRPO to generate higher-impact research ideas.
- A generative reward model that reasons over paired paper abstracts and predicts which has higher potential impact.
- Trained with GRPO on 720K field- and time-matched citation-based preference pairs.
- Serves both as an evaluator of research ideas and as the reward model for Scientific Thinker training.
- A scientific ideation policy trained with Scientific Judge as the reward model.
- Takes a paper title and abstract as input and proposes follow-up research ideas with higher potential impact aligned with community preference.
- Optimized with comparison-based GRPO for open-ended idea generation.
- 720,341 preference pairs and 1,440,682 pair-level paper records; papers may recur across pairs, so this is not a unique-paper count.
- Built from arXiv papers across Computer Science, Mathematics, Physics, and Other scientific fields.
- Evaluated in-domain and across temporal OOD (904 pairs from papers published in 2025), metric OOD (ICLR peer review and Altmetric attention), field-transfer, and controlled-comparison settings, with bioRxiv as an additional biology evaluation.
Our paper shows that scientific taste can be learned and transferred:
- Scientific judgement scales with both data size and model size.
- Scientific Judge-Qwen3-30B reaches 82.7% in-domain accuracy, surpassing all listed strong LLM baselines, including GPT-5.4 Thinking at 81.6%.
- Learned judgement transfers to future-year papers: on the refreshed 904-pair test set from papers published in 2025, Qwen3-4B improves from 64.7% to 80.9% (+16.2 points), and Qwen3-30B-A3B improves from 71.7% to 83.1% (+11.4 points).
- Learned judgement generalizes across fields and community metrics, including bioRxiv biology transfer, ICLR peer-review preferences, and Altmetric attention, while its gains persist under author/institution and topic controls.
- Scientific Thinker achieves a 54.2% average win rate against three strong LLM baselines in both in-domain and out-of-domain settings, compared with 30.3% and 27.8% for its base policy, respectively.
If you find our work helpful, please consider citing:
@article{tong2026ai,
title={AI Can Learn Scientific Taste},
author={Tong, Jingqi and Li, Mingzhe and Li, Hangcheng and Yang, Yongzhuo and Mou, Yurong and Ma, Weijie and Xi, Zhiheng and Chen, Hongji and Liu, Xiaoran and Cheng, Qinyuan and others},
journal={arXiv preprint arXiv:2603.14473},
year={2026}
}This project is licensed under the Apache License 2.0. See LICENSE for details.