As LLM-assisted writing becomes routine in academic research, a parallel shift is occurring on the other side of the submission portal: reviewers are increasingly using the same tools to generate or augment their reviews. This paper, available at arXiv:2605.28897, asks a pointed question: if both authors and reviewers are relying on LLMs, what does that mean for the integrity of peer review? The authors study this on 984 real submissions from the 2025 ACL Rolling Review (ARR), covering multiple models, prompts, and an adversarial editing pipeline. The result is one of the more grounded empirical treatments of a problem the community has been circling around theoretically for some time.
Key Contributions
The paper makes three concrete contributions. First, it provides the largest empirical evaluation of LLM reviews against human reviews on actual ARR submissions, rather than synthetic or historical data. Second, it introduces and evaluates an Iterative Submission Improvement (ISI) pipeline, in which an LLM both reviews a paper and rewrites it to address its own criticisms, repeated up to ten times. Third, it offers a taxonomy of edit types, ranging from superficial rephrasing to outright fabrication of evidence, grounding the adversarial framing in prior literature on paper laundering.
The three research questions are cleanly separated: whether LLM reviews align with human ones (validity), whether they are consistent across models and prompts (stability), and whether the review process can be gamed through automated iterative editing (gameability). This structure makes the paper easy to follow and the findings easy to interpret independently.
Methodology
The experimental setup is deliberately conservative. Rather than using agentic pipelines or retrieval-augmented systems that might inflate performance, the authors evaluate off-the-shelf models as a proxy for how reviewers actually use these tools. Five prompts of increasing specificity are tested, from a minimal instruction to a full ARR-specific prompt with a senior reviewer persona. Six models are evaluated, including open-weight options (Qwen-3.6-35B, Gemma-3-27B, Llama-3.3-70B) and closed-weight ones (GPT-5.4-mini, GPT-5.4). Each model-prompt combination generates three reviews per paper, enabling stability analysis.
Alignment is measured using mean absolute error (MAE) and Pearson correlation against human Overall scores, with a best-match variant that compares against the closest human reviewer rather than the mean. A separate LLM judge evaluates semantic overlap between LLM-identified strengths and weaknesses and those in the human review, providing a recall-style content metric beyond scores alone.
For the ISI experiments, three editing prompts are used: a constrained prompt targeting surface-level changes, a default prompt based on prior work by Baumann et al., and an adversarial prompt that explicitly permits fabrication of evidence. The baseline for comparison is simply re-running the review on the unedited paper ten times, which controls for score variance from stochasticity.
Results
The alignment findings are sobering. Human-human Pearson correlation on the Overall score reaches 0.31 on the combined split, while the best LLM model (GPT-5.4 with the best-matching prompt) achieves 0.28. That sounds close, but the variance across prompts and models is substantial. Across all prompts, GPT-5.4 drops to 0.18, and Llama-3.3-70B averages just 0.10. The MAE picture is similarly mixed: most models sit around 0.70 to 0.97 points on a rating scale, compared to a naive midpoint baseline of 0.64. In other words, some models barely outperform predicting the same score for every paper.
Key findings from the alignment analysis:
- LLM-human alignment is at best comparable to human-human agreement, but this holds only for the best model-prompt combination.
- Performance degrades significantly on rejected papers, where LLM reviews show near-zero correlation with human scores in some configurations.
- Prompt choice matters substantially, sometimes more than model choice, which has practical implications for any official deployment.
On gameability, the ISI pipeline produces statistically significant score increases for up to 35% of papers under constrained editing conditions. The adversarial condition, somewhat surprisingly, does not outperform constrained editing and in some cases performs worse. The authors attribute this to model guardrails resisting fabrication and to the possibility that fabricated content introduces internal inconsistencies that the reviewing LLM penalises. The default prompt, replicating Baumann et al.'s approach, shows weaker effects than the constrained version, which the authors argue reflects the importance of structured editing guidance rather than open-ended revision instructions.
Implications and Limitations
The Goodhart's law framing is apt and worth taking seriously. If authors begin optimising submissions for LLM reviewers, and if LLM reviews become more prevalent in official processes, then scores that currently show modest alignment with human judgment will degrade further as a signal of actual paper quality. The ISI results demonstrate this is not a hypothetical: automated rewriting can push papers past score thresholds without necessarily improving scientific content.
There are genuine limitations worth flagging. The study does not test cross-model transfer for the adversarial editing pipeline. A paper optimised to score well under GPT-5.4 may not fool a human reviewer or a different LLM, and this generalisation question is left open. The acceptance/rejection ground truth is also imperfect: the NeurIPS experiment cited in the paper found roughly 50% disagreement between independent committees, so human scores are a noisy gold standard to begin with. The authors acknowledge this, but it complicates interpretation of the alignment numbers.
The study also evaluates models in isolation rather than in the kind of multi-reviewer aggregation that real conferences use. A single LLM review is not how peer review works; the question of whether LLM scores, averaged across multiple runs or models, converge more reliably with human consensus is left for future work.
What this paper does well is resist the temptation to frame LLM reviews as either clearly adequate or clearly inadequate. The honest answer from the data is that they are inconsistent, prompt-sensitive, and gameable in specific conditions. That is a more useful characterisation than a binary verdict, and it gives the community concrete parameters to work with when designing any official integration. The call for caution is appropriate, and the released code makes the findings reproducible and extendable.