Hugging Face Daily PapersSy-Tuyen Ho, Minghui Liu, Furong Huang1 min readpaperadvanced
When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation
Summary
The authors show that training LLM reviewers on synthetic reviews leads to a compression of rating distributions and loss of semantic diversity, a phenomenon they call scientific-judgment collapse. They mitigate it with TrustReviewer, which uses curated training data and activation steering to preserve judgment diversity.
- Synthetic reviews bias subsequent reviewer models, narrowing rating spreads and reducing semantic variety.
- Training a reviewer on a curated corpus (TrustReviewer) prevents low-quality, degenerate supervision.
- Paired activation steering at inference time further corrects residual collapse without extra training.
- Empirical results on ICLR data demonstrate restored diversity and better alignment of recommendations.
Anyone building LLM-based peer-review tools should care, as data contamination can degrade evaluation quality over time.
7/10
