AI Peer Review Under the Microscope: What a Major Audit Reveals About LLMs as Scientific Reviewers

The peer review system is under extraordinary strain. With global research output doubling roughly every nine years and editorial boards stretched thin across thousands of journals and conferences, the question of whether large language models can shoulder a meaningful portion of the review burden has moved from theoretical curiosity to operational reality. Thousands of conferences and journals are already experimenting with AI-assisted review pipelines. And yet, until recently, the empirical foundation for understanding what these systems actually do — what they catch, what they miss, and whether they treat all authors equally — has remained remarkably thin. A new preprint published on arXiv changes that calculus significantly, and its findings carry direct implications for anyone working at the intersection of AI peer review, scientific publishing, and research integrity.

The Audit: What Was Studied and Why It Matters

The study, authored by researchers examining two state-of-the-art multimodal large language models — Qwen2.5-VL-72B and Pixtral-Large-124B — is notable for the rigor of its experimental design. Rather than using archival or synthetic manuscripts, the researchers evaluated both models against 165 actual submissions to the 2026 International Conference on Learning Representations (ICLR 2026), a venue that postdates both models' training cutoffs. This detail is critical: it means neither model had been exposed to these papers during pretraining, eliminating the possibility that favorable scores reflect memorization rather than genuine critical assessment.

Manuscripts were presented with author identities blinded, allowing the researchers to probe two distinct dimensions simultaneously: first, whether the models could produce well-calibrated scores relative to human reviewers, and second, whether they demonstrated any systematic bias when author identity cues were present or absent. The choice of multimodal models — systems capable of processing figures, tables, and equations directly from PDF representations — also reflects a meaningful advance over text-only evaluations, since scientific manuscripts are inherently visual documents.

The scope is significant. One hundred and sixty-five papers across a competitive machine learning venue, two models representing different architectural and parameter-scale choices, and a methodological design that isolates both scoring calibration and identity effects. This is among the most comprehensive empirical audits of AI-powered peer review published to date.

Scoring Calibration: How Well Do LLMs Actually Judge Research Quality?

One of the central findings concerns scoring calibration — the degree to which an AI reviewer's numerical scores align with the consensus judgments of experienced human reviewers. This is not a trivial question. A system that consistently overscores mediocre work or underscores technically sound but unconventionally framed research is not merely inaccurate; it actively distorts the manuscript selection process.

The audit's findings on this dimension reveal a pattern that practitioners in automated manuscript analysis will recognize immediately: both models demonstrated meaningful correlation with human scores in aggregate, yet showed systematic miscalibration at the tails of the distribution. High-quality papers — those ultimately accepted by human reviewers — were not reliably distinguished from strong-but-ultimately-rejected submissions. This compression toward the middle of the scoring range, a phenomenon sometimes called range restriction in psychometrics, means that AI peer review tools operating on raw score outputs could inadvertently flatten the distinctions that matter most in competitive review contexts.

Moreover, the two models diverged noticeably from each other on a non-trivial subset of papers, suggesting that the variation is not merely noise but reflects genuine differences in how each model weights methodological rigor, novelty, and presentation quality. For researchers relying on AI research tools to obtain pre-submission feedback, this divergence is practically meaningful: a manuscript that scores well under one model's implicit evaluation rubric may score poorly under another's, with no clear ground truth to adjudicate between them.

Error Detection: Where LLMs Succeed and Where They Fall Short

Perhaps the most operationally significant findings concern error detection — the capacity of AI reviewers to identify specific technical flaws, logical inconsistencies, or methodological weaknesses rather than simply assigning holistic quality scores. Human peer review, at its best, functions as error detection: it catches underpowered statistical analyses, missing baselines, overgeneralized conclusions, and experimental designs that cannot support the claimed contributions.

The audit reveals a nuanced picture here. Both models demonstrated reasonable competence at identifying surface-level presentation issues — missing references, inconsistent notation, unclear figure captions — and at flagging claims that were not adequately supported by the presented experiments. This is consistent with what practitioners at platforms focused on AI manuscript review have observed: LLMs are generally effective at structural and rhetorical analysis.

Where performance degraded substantially was in the detection of deep methodological errors: flawed statistical reasoning embedded in technical prose, subtle data leakage issues in machine learning pipelines, and experimental confounds that require domain expertise to recognize. This is not surprising from a mechanistic standpoint — detecting that a particular choice of evaluation metric systematically favors a proposed method over baselines requires not just language understanding but a rich internalized model of experimental norms in a specific subfield. The audit's data suggests that neither 72-billion nor 124-billion parameter multimodal models have fully closed this gap.

For researchers using AI research tools as a first-pass quality check before submission, this implies a clear division of labor: AI systems can add substantial value in catching the errors that time pressure and familiarity blindness cause human authors to overlook, but they should not be treated as substitutes for domain-expert review on technical substance.

Author Identity Effects: The Question of Fairness in Automated Review

The third dimension examined — author identity effects — touches on one of the most sensitive fault lines in academic publishing. Human peer review has documented biases along lines of institutional prestige, perceived gender, and geographic origin. The question of whether AI peer review systems replicate, amplify, or mitigate these biases is not merely academic; it has direct implications for equity in scientific publishing.

The study's design, which blinded author identities in the primary condition, allows for controlled investigation of whether identity cues affect model behavior when they are present. The findings here warrant careful interpretation. Both models showed statistically detectable sensitivity to author-identity signals when these were available, though the magnitude and direction of effects varied. This sensitivity is troubling precisely because it suggests that AI reviewers are not simply evaluating scientific content in isolation — they are drawing on contextual signals in ways that parallel documented human biases, even if the mechanism differs fundamentally.

This finding has direct implications for how AI-powered peer review systems should be deployed. Blinding, which is already standard practice in double-blind review venues, should be treated as a non-negotiable requirement rather than an optional configuration when LLMs are part of the review pipeline. Platforms serious about research integrity — including tools like PeerReviewerAI, which focuses on manuscript-level analysis — need to build identity-blinding into the core of their processing architecture, not as an afterthought.

Implications for AI-Assisted Peer Review Platforms and Workflows

Taken together, the three findings — limited score calibration at the distribution tails, constrained deep error detection, and measurable identity sensitivity — paint a picture that is neither damning nor exculpatory. These systems are capable reviewers within a defined envelope of tasks, and operating outside that envelope produces outputs that range from unhelpful to actively misleading.

For developers and operators of AI peer review infrastructure, the audit provides a practical benchmark. Score outputs from current-generation LLMs should be treated as soft signals rather than authoritative evaluations. Uncertainty quantification — providing score ranges or confidence intervals rather than point estimates — would more honestly represent what these models can and cannot know. Ensemble approaches, combining multiple models as this study implicitly suggests by comparing two systems, may partially compensate for individual model blind spots.

The finding on error detection also motivates a design principle that the most thoughtful AI research tools are already moving toward: rather than asking a model to produce a single holistic review, decompose the review task into structured subtasks — claims verification, methodology assessment, statistical analysis review, and presentation quality — each of which can be separately validated and calibrated. This modular approach reduces the risk that strengths in one dimension mask failures in another.

For researchers preparing manuscripts, the practical implication is that AI-assisted pre-review tools remain most valuable when used as structured checklists rather than as oracles. A tool like PeerReviewerAI can systematically surface the structural and rhetorical weaknesses that are genuinely hard to self-diagnose after months of immersion in a project, and that value is real and documented. But researchers should approach AI-generated technical feedback on core methodological choices with appropriate skepticism, and should not use high AI scores as a proxy for submission readiness.

Practical Takeaways for Researchers Using AI Research Tools

The audit's findings translate into several concrete recommendations for researchers navigating the expanding landscape of AI-assisted manuscript preparation and review.

First, treat AI review outputs as structured feedback, not verdicts. The score compression documented in this study means that numerical outputs from current LLM reviewers carry less information than they appear to. Focus instead on the specific critiques and flagged weaknesses, which tend to be more reliably informative than aggregate scores.

Second, use multiple AI tools when possible. The divergence between Qwen2.5-VL-72B and Pixtral-Large-124B on a meaningful subset of papers suggests that any single AI reviewer introduces model-specific blind spots. Triangulating across multiple systems reduces the risk of missing important weaknesses that one model systematically underweights.

Third, never submit to a blinded venue without ensuring that author-identifying information has been removed from all manuscript layers — metadata, acknowledgments, self-citations, and figure file properties included. The identity sensitivity found in this study suggests that AI reviewers may be more susceptible to inadvertent unblinding than human reviewers, simply because they process all available text without the metacognitive awareness that would cause a human to flag a suspicious self-citation pattern.

Fourth, focus AI pre-review on the areas where it demonstrably adds value: structural coherence, citation completeness, figure and table clarity, and consistency between abstract claims and reported results. For deep methodological questions, prioritize expert human feedback.

The Road Ahead: Calibrating Expectations for AI Peer Review

The arXiv audit represents a maturing of the conversation around AI peer review — a move away from early-stage enthusiasm toward the kind of rigorous, empirical characterization that responsible deployment requires. The findings confirm that current multimodal LLMs occupy a specific and limited role in the review ecosystem: capable assistants for structured, surface-level analysis, with measurable limitations in deep technical assessment and non-trivial sensitivity to identity signals that demands careful system design.

This is, in many ways, exactly where the field should be at this stage of development. The appropriate response is not to retreat from AI research tools but to deploy them with precision — understanding their calibration properties, designing around their failure modes, and maintaining human expert judgment at the points in the process where it is genuinely irreplaceable. As models scale further, as training data increasingly incorporates structured scientific reasoning, and as evaluation frameworks like this audit accumulate into an empirical literature, the envelope of reliable AI peer review will expand. The imperative for researchers, publishers, and tool developers alike is to track that expansion carefully, updating deployment practices as the evidence warrants rather than ahead of it.