The Pulse
One AI training step cut review diversity by 11%
A University of Maryland study finds that adding synthetic peer reviews to AI training compresses ratings and reduces semantic diversity. The researchers introduce TrustReviewer, an open-source system designed to preserve broader scientific
AI.info Team ·
Introducing synthetic reviews into the training of an AI peer reviewer reduced the semantic diversity of reviews for the same paper by approximately 11% after a single recursive training step, according to a University of Maryland study published on arXiv on September 17, 2026. The researchers also measured an approximately 5% contraction in the broader semantic spread of the review corpus.
The paper, “When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation,” examines a feedback loop that could form when model-written reviews become public, enter later training datasets and shape the judgments of successor systems. Sy-Tuyen Ho, Minghui Liu and Furong Huang call the resulting pattern “scientific-judgment collapse.”
The authors stress that their experiment does not show that AI systems already dominate scientific peer review. Instead, it isolates what happens when a known amount of model-generated supervision is added to an otherwise controlled training process.
Four training mixtures, one narrowing pattern
The experiment starts with Meta’s Llama 3.1 8B Instruct model. The researchers fine-tune it on official ICLR reviews from 2018 through 2023, then use that reviewer to generate synthetic reviews for ICLR 2024 papers.
Four successor models receive different mixtures of official and generated reviews. Each paper contributes three reviews to the training data: zero, one, two or three synthetic reviews, corresponding to synthetic exposure levels of 0%, 33%, 66% and 100%.
The models share the same initialization, filtering process and training settings. The planned difference is the proportion of reviews explicitly generated by the earlier model, allowing the researchers to compare the effect of synthetic exposure after one recursive step.
Ratings become more concentrated
The study evaluates the models on 2,000 held-out papers drawn from ICLR submissions from 2018 through 2025. The researchers extract each generated review’s overall recommendation on a 1-to-10 scale and measure the mean, standard deviation and entropy of the resulting rating distributions.
As synthetic reviews enter the training mixture, the rating distribution becomes narrower. The standard deviation falls from 1.63 for the official-review-only successor model to 1.44 when synthetic reviews make up 33% of the mixture; the value remains below the baseline at higher exposure levels.
Mean ratings do not move in one direction. They rise from 5.30 at 0% synthetic exposure to 5.85 at 33%, then fall to 5.70 at 100%. The authors interpret the result as a reduction in the range of judgments rather than a consistent move toward either harsher or more generous recommendations.
Reviews of the same paper grow more alike
The researchers also compare the language of multiple reviews addressing the same submission. They convert reviews into semantic embeddings and calculate the average pairwise distance between them, using a lower distance as a proxy for greater similarity.
That distance declines from approximately 0.159 for the official-review-only model to 0.153, 0.147 and 0.142 as synthetic exposure rises from 33% to 66% and then 100%. The change from 0% to 100% represents an approximately 11% reduction in same-paper semantic diversity.
The broader corpus shows a similar, though smaller, contraction. Semantic spread around the corpus centroid decreases from approximately 0.609 at 0% synthetic exposure to 0.579 at 100%, a reduction of about 5%.
Those measurements do not establish that the reviews identify fewer valid scientific problems. The paper uses embedding distance as a proxy for diversity, and the authors say the method does not directly test factual accuracy, critique quality or substantive coverage.
TrustReviewer avoids the second training loop
Ho, Liu and Huang propose TrustReviewer as an open-source system for generating reviews of artificial intelligence and machine learning papers. Its training process uses a curated collection of 112,743 paper-review examples drawn from official ICLR material published from 2018 through 2025, totaling approximately 1.9 billion tokens.
The curation process removes malformed, duplicated, excessively short, repetitive and overlength examples. The model is then trained in a single stage rather than being trained on reviews produced by an earlier reviewer.
The system adds a second intervention at generation time. The researchers create paired activation directions by comparing official and model-generated reviews of the same papers, then add the resulting steering vector to the model’s final decoder layer while it writes a review.
On the held-out evaluation set, TrustReviewer matches at least one official rating for 75.40% of papers. OpenReviewer reaches 73.10%, Qwen3.6-35B-A3B reaches 61.85% and the unmodified Llama 3.1 8B model reaches 33.93%.
TrustReviewer’s rating entropy is 2.18, compared with 2.38 for the official review reference. Activation steering raises exact-match performance from 73.85% to 75.40% against the same checkpoint without steering, a gain of 1.55 percentage points.
A narrow experiment with a broader warning
The paper’s findings apply to one recursive step, one base-model family and one scientific domain. The authors do not establish that the effect compounds over multiple generations, transfers to other fields or appears at other model scales.
They also caution that official ICLR reviews are not ground truth. Such reviews can contain errors, disagreement and unobserved AI assistance, which is why the study uses “official” rather than “human” to describe the reference data.
The concrete risk identified by the experiment is narrower: if generated judgments become training material for later reviewers, the next system may express a smaller range of ratings and less varied reasoning. TrustReviewer’s results suggest that filtering the training corpus and applying a modest inference-time intervention can preserve more diversity, but the study does not show that broader diversity alone produces better scientific decisions.