Review Arcade: On the Human Alignment and Gameability of LLM Reviews

Researchers evaluated LLM-generated peer reviews for scientific papers using ACL Rolling Review data, finding limited alignment between LLM and human reviews while discovering that authors can strategically game LLM feedback to improve paper scores by up to 35%. The study highlights emerging risks in automated academic review systems as both reviewers and authors increasingly leverage language models.

The adoption of large language models for peer review represents a significant shift in academic publishing infrastructure. This research exposes a critical vulnerability in systems that lack sufficient human oversight: when both reviewers and authors use LLMs, misalignment between model outputs and human judgment creates systematic biases. The findings demonstrate that LLM reviews vary substantially depending on prompts and model selection, suggesting no standardized evaluation framework exists yet.

The "gaming" phenomenon identified in the study reveals a fundamental challenge in automated evaluation systems. Authors iteratively revising work based on LLM feedback achieved statistically significant score improvements in up to 35% of cases, indicating that optimizing for machine-generated reviews differs from improving actual scientific quality. This creates a perverse incentive structure where authors chase algorithmic approval rather than pursuing rigorous research.

For the academic and AI communities, these findings signal that deploying LLMs in high-stakes evaluation contexts requires careful calibration and human validation. Major conferences piloting LLM reviews—including those using ACL Rolling Review—now face pressure to implement stronger safeguards. The research suggests that pure automation in peer review risks degrading academic quality if authors exploit predictable LLM patterns.

Looking forward, the field must balance efficiency gains from LLM assistance against integrity concerns. Hybrid models combining machine-generated initial reviews with mandatory human oversight appear necessary. The publication of reproducible code enables other researchers to audit and improve LLM review systems, potentially leading to more robust evaluation frameworks that resist gaming while maintaining alignment with expert judgment.

• →LLM reviews show limited alignment with human peer reviews, with consistency varying significantly across different prompts and models.
• →Authors can strategically revise papers based on LLM feedback to achieve statistically significant score improvements in up to 35% of submissions.
• →The "gaming" of LLM reviews creates incentive misalignment between optimizing for algorithmic approval and improving actual research quality.
• →Major academic conferences piloting LLM reviews need stronger human oversight mechanisms to prevent systematic bias and maintain integrity.
• →Reproducible research tools now enable auditing of LLM review systems and development of more robust evaluation frameworks.