Review rarely improves stronger model answers in language AI tasks

Review Rarely Beats the Stronger Model: Separating Verification from Capability Transfer in LLM Pipelines

Computation and LanguageArtificial Intelligence

Summary

When one AI model checks another model's answers, it might seem like accuracy improves a lot. But this happens because the reviewer sometimes repeats answers it could have given on its own, not because reviewing really helps. The authors found that in nearly all cases, just using the stronger AI model alone is better than reviewing. Reviewing is also more expensive and often loses correct answers from the reviewer. They suggest that evaluations should compare to the stronger model alone, not only the improvement over the original answer.

What this means in practice

Authors

Yirui Zhang, Yash Raj Shrestha

Abstract

Qwen3-8B reviewing Llama-3.1-8B in math adds 33.2 points, yet Qwen3-8B alone is 6.4 points more accurate than the result. Reviewing another model's answer is usually judged by its gain over the generator. We show that this confuses verification, which adds accuracy beyond either model alone, with capability transfer, where a stronger reviewer supplies answers it could have given on its own. Measured against the stronger model alone, across 48 pairs of 12 open-weight models (3B to 72B) in math, knowledge, code and reasoning, review beats it in only 9 of 96 review directions, and does worse the larger the capability gap. The reason is that the reviewer mostly follows the generator, discarding its own correct answers. Review costs 2.6 to 2.8 times as much as the stronger model alone, while a selector using the same two calls is 2.7 to 4.2 points more accurate than review and cheaper. Studies of verification should report the stronger model alone, not only the gain over the generator. Code and per-question results: https://github.com/Kamilaabb/verification-or-capability-transfer
View PDFOpen arXiv