When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation
Training LLM reviewers on model-generated reviews compresses rating distributions and diversity, a failure the authors name scientific-judgment collapse and mitigate with TrustReviewer.
The authors fine-tune Llama 3.1 8B as a reviewer on official ICLR reviews from 2018-2023, then train successor models on ICLR 2024 data with varied mixtures of official and synthetic reviews. Synthetic reviews compress rating distributions and reduce same-paper and corpus-level semantic diversity, a pattern termed scientific-judgment collapse. They introduce TrustReviewer, an open-source reviewer system combining curated single-stage training with test-time paired activation steering to preserve judgment diversity.
- Recursive AI peer review training compresses ratings and reduces semantic diversity
- Experiments use Llama 3.1 8B fine-tuned on ICLR 2018-2024 reviews
- TrustReviewer mitigates collapse via curated training and activation steering
Full article216 words · extracted from huggingface.co · click to collapse
Large language models (LLMs) increasingly participate in scientific evaluation, both as automated reviewers and as assistants to human reviewers. As model-generated reviews enter public data and future training corpora, AI peer review can become recursive: later reviewers learn from judgments produced by earlier models. We study one step of this feedback loop in a controlled setting. Starting from Llama 3.1 8B, we first fine-tune a reviewer on official ICLR reviews from 2018--2023 and then train four successor models on ICLR 2024 data with systematically varied mixtures of official and model-generated reviews. Our study shows that introducing synthetic reviews compresses rating distributions and reduces both same-paper and corpus-level semantic diversity. We call this pattern scientific-judgment collapse. To mitigate this failure mode, we introduce TrustReviewer, an open-source LLM-based system for generating peer reviews of AI and machine learning papers. TrustReviewer intervenes at two complementary stages. For training-time prevention, we train the core reviewer in a single stage on a curated corpus designed to reduce low-quality and semantically degenerate supervision. For test-time correction, paired activation steering aims to further mitigate residual tendencies toward collapsed judgments without further training or additional expert annotation. Together, these results characterize a concrete risk of recursive reviewer training and provide practical interventions for preserving judgment diversity and improving recommendation alignment in AI-assisted scientific evaluation.
Text extracted automatically; images, tables and formatting may be missing. Original: https://huggingface.co/papers/2609.20942