PaperCritic: Training AI Reviewers to Assess the Soundness of Papers
Abstract
An AI reviewer trained by reinforcement learning to jointly maximize precision and recall of soundness review points outperforms frontier models and other training methods and transfers to other review domains.
As AI conference submissions grow rapidly, AI reviewers are being considered as an alternative. However, simply imitating human reviewing is problematic because reviewing serves multiple purposes, including evaluating papers' soundness, interestingness, and clarity, the last two of which are subjective. In this work, we focus solely on evaluating soundness, the most important and objective criterion, and define the learning objective as jointly maximizing (1) precision, based on the soundness and substantiveness of review points, and (2) recall, the coverage of high-quality reference review points. Under this problem formulation, we construct a quality-labeled review point dataset, PaperSound-23K, using 1K ICLR submissions and 23K human-written and LLM-generated review points. Using this dataset, we train an AI reviewer model, PaperCritic, through reinforcement learning with the F1 score of precision and recall as the reward. Our experiments show that PaperCritic outperforms frontier models and models trained with other AI reviewer training methods on an in-domain test set. Surprisingly, learning to review ICLR submissions also transfers to reviewing Nature-family papers and even to out-of-domain reviewing tasks (e.g., code review, legal contract review, and feedback on story writing).
Community
As AI conference submissions grow rapidly, AI reviewers are being considered as an alternative. However, simply imitating human reviewing is problematic because reviewing serves multiple purposes, including evaluating papers' soundness, interestingness, and clarity, the last two of which are subjective. In this work, we focus solely on evaluating soundness, the most important and objective criterion, and define the learning objective as jointly maximizing (1) precision, based on the soundness and substantiveness of review points, and (2) recall, the coverage of high-quality reference review points. Under this problem formulation, we construct a quality-labeled review point dataset, PaperSound-23K, using 1K ICLR submissions and 23K human-written and LLM-generated review points. Using this dataset, we train an AI reviewer model, PaperCritic, through reinforcement learning with the F1 score of precision and recall as the reward. Our experiments show that PaperCritic outperforms frontier models and models trained with other AI reviewer training methods on an in-domain test set. Surprisingly, learning to review ICLR submissions also transfers to reviewing Nature-family papers and even to out-of-domain reviewing tasks (e.g., code review, legal contract review, and feedback on story writing).
One control worth requiring before an automated reviewer is trusted, from a measurement we ran. We scored 29 confident answers from a 1B model on medical questions with a published answer-scoring method, four ways. Against the true key: endorsed 29 of 29. Against a deliberately swapped wrong key: endorsed 29 of 29. Paraphrased, same confidence: 29 of 29. With the confidence markers removed: 0 of 29.
We found it at 0% — on the scorer's first fixture, seven sentences, before it had scored any of our data. Much better observed before the data is scored than after.
Changing the truth moved none of its decisions. Changing the wording moved all of them.
That is the failure a soundness reviewer must not have. A judge that follows confidence will pass fluent unsound work and flag hesitant sound work, and ordinary accuracy scores will not reveal it — the judge agrees with the key whenever the answer is confident, right or wrong. The swap exposes it in one step: re-score against a wrong key, and a judge of content changes its verdict while a judge of confidence does not. If this paper's evaluation does not already include a key-swap control, this measurement suggests it is the first one to add.
Get this paper in your agent:
hf papers read 2611.17834 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper