When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation

When AI Reviews Train AI Reviewers. Scientific-Judgment Collapse and Mitigation. By Sy-Tuyen Ho and two others, from the University of Maryland, College Park. Published on the seventeenth of September. No code repository is linked to it.

Large language models increasingly take part in scientific evaluation, both as automated reviewers and as assistants to human ones. As model-generated reviews enter public data and future training corpora, AI peer review can turn recursive: later reviewers learn from judgments made by earlier models. The authors study one step of that loop and give the result a name, scientific- judgment collapse. When synthetic reviews are introduced into training, rating distributions compress, and semantic diversity falls, both among reviews of the same paper and across the corpus as a whole. Reviewers trained partly on machine reviews end up rating more alike and writing more alike.

The setting is controlled. Starting from Llama three point one with eight billion parameters, the authors first fine-tune a reviewer on official ICLR reviews from two thousand eighteen through two thousand twenty-three. That reviewer plays the part of the earlier model in the loop. They then train four successor models on ICLR two thousand twenty-four data, systematically varying the mixture of official reviews and reviews generated by a model. Comparing the successors against one another is what isolates the effect of the synthetic share.

To mitigate the failure, the authors introduce TrustReviewer, an open-source system built on a language model that generates peer reviews of AI and machine learning papers. It intervenes at two complementary stages. At training time, for prevention, the core reviewer is trained in a single stage on a curated corpus designed to reduce low-quality and semantically degenerate supervision. At test time, for correction, paired activation steering aims to further mitigate residual tendencies toward collapsed judgments, with no further training and no additional expert annotation.

The claim is bounded by its design. It concerns one step of the feedback loop, observed in a controlled setting that starts from one base model and draws on reviews from one venue. The mitigation is offered as practical intervention, aimed at preserving the diversity of judgment and improving how well recommendations align, in scientific evaluation assisted by AI. What the study characterizes is a concrete risk of recursive reviewer training, identified at the scale of that single step.

Nothing dramatic breaks when a reviewer learns from machine reviews. The ratings bunch together and the reasoning grows more uniform, which is the kind of drift that is easy to miss one generation at a time. The authors name it, measure it, and offer ways to push back at both ends: in the training data, and at the moment a review is written.