Spending more compute at inference time is a standard way to make language models and agents more accurate. The usual recipe is to generate many candidate solutions and submit the one a verifier prefers. A new arXiv preprint (2610.23255) frames this as an allocation problem: how much compute should go into producing candidates versus checking them?

Success depends on two conditions: some candidate in the batch must be correct, and the verifier must be able to pick it. When verifiers are imperfect, the best allocation may change. The abstract introduces the problem but does not state results in the excerpt; the full paper presumably analyzes trade-offs among sampling, mixing, and verification.