This post from zhichai.net discusses the paper "Review Arcade: On the Human Alignment and Gameability of LLM Reviews" (arXiv:2605.28897v1) by Hans Ole Hatzel, Sebastian Steindl, and Jan Strich (cs.AI, cs.MA, May 2026), which examines how LLM-based peer review aligns with human reviewers — and how easily it can be gamed.
Background
LLM-assisted reviewing is moving from experiment to practice: ACL Rolling Review piloted LLM-assisted review in 2025, while ICLR and NeurIPS are exploring LLM screening. The drivers are reviewer shortages, demand for consistency, and speed. The paper asks a neglected question: if LLMs enter the review pipeline, how will the submission ecosystem adapt?
Research Questions
1. How well do LLM reviews align with human reviews? 2. How much do LLM reviews vary across models and prompts? 3. Can authors systematically inflate LLM review scores through iterative revision?
Method
The study uses 984 real ARR 2025 submissions, each with three human reviews, LLM-generated reviews (multiple models and prompts), and full text/metadata.
Key Findings
- Limited but variable alignment. LLM reviews correlate with human reviews only moderately, and alignment depends heavily on the model (GPT-4 vs Claude vs Llama) and the prompt (e.g., strict vs friendly instructions).
- Gameability works. Simulating a draft → LLM review → revise → re-review loop (up to 5 rounds), 35% of papers gained statistically significant overall score improvements — with no change in actual scientific contribution.
- Systematic reviewer biases explain the effect:
- *Surface-structure bias*: clearer headings, subheadings, and bullet lists improve scores even without substantive changes.
- *Self-consistency bias*: papers written in LLM-like academic style are favored.
- *Specificity bias*: added technical detail (even redundant) raises scores.
- *Responsiveness bias*: canned acknowledgments of limitations ("we admit X, but...") earn higher ratings.
- Convergence trap: iterative revision converges toward the LLM's preferred style rather than genuine quality — and if many authors iterate, field-wide writing style converges on LLM bias.
- A review arms race: authors use LLM optimizers, reviewers deploy anti-gaming tools, and resources shift from better science to pleasing/detecting manipulation.
- Hidden quality decline: surface polish rises while substance stagnates; innovative, non-conforming work may be penalized; stylistic homogenization increases.
- Fairness threats: unequal access to LLM tools, bias against non-native English writers, and asymmetry in knowledge of how to game LLM reviewers.
- Hans Ole Hatzel, Sebastian Steindl, Jan Strich (2026). *Review Arcade: On the Human Alignment and Gameability of LLM Reviews*. arXiv:2605.28897v1.
- ACL Rolling Review: https://aclrollingreview.org/
- Liang, P., et al. (2023). *Holistic Evaluation of Language Models*. NeurIPS.
- Liu, Y., et al. (2024). *LLM Reviewers' Unconscious Bias: A Large-Scale Study*. arXiv.
- Stelmakh, I., et al. (2024). *Peer Review as a Multi-Player Game: A Mechanism Design Approach*. arXiv.
Systemic Risks
Proposed Mitigations
1. Hybrid review: LLM output as auxiliary only, clearly labeled AI-generated, humans hold final say. 2. Adversarial design: multiple diverse LLM reviewers; deliberately varied prompt styles to prevent prompt-specific optimization. 3. Dynamic rubrics: periodically rotate scoring criteria; occasional blind human checks for consistency. 4. Author education: clarify that LLM review comments are advisory, not final.
Takeaway
The paper's value lies not in showing that LLM review can be gamed (arguably inevitable) but in quantifying it (35% significant gains), revealing the mechanism (iterative convergence on reviewer bias), and grounding the systemic risks in a large real-data experiment. The post's conclusion: LLM review is not inherently dangerous, but treating it as an opaque black box hands judgment of scientific quality to a manipulable system. Responsible adoption requires understanding its biases first.