arXiv cs.CL· Yirui Zhang, Yash Raj Shrestha·· 3 天前AI 评分36
研究：LLM 审查很少胜过更强模型，验证与能力迁移需区分
Review Rarely Beats the Stronger Model: Separating Verification from Capability Transfer in LLM Pipelines
AI 导读
对 12 个开源模型（3B 至 72B）组成的 48 对组合测试显示，在数学、知识、代码和推理任务中，审查仅在 96 个审查方向中的 9 个胜过更强模型，且能力差距越大表现越差。原因是锚定效应：审查者大多跟随生成者，丢弃了自己的正确答案。审查成本是单独使用更强模型的 2.6 至 2.8 倍，而同样两次调用的选择器比审查准确率高 2.7 至 4.2 分且更便宜。
正文 · 原文
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 anchoring: 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.
| Comments: | 22 pages, 7 figures, 12 tables. Code: this https URL |
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.15897 [cs.CL] |
| (or arXiv:2610.15897v2 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.15897 arXiv-issued DOI via DataCite |
Submission history
From: Yirui Zhang [view email]
[v1]
Sun, 11 Oct 2026 20:26:36 UTC (288 KB)
[v2]
Wed, 14 Oct 2026 11:38:52 UTC (291 KB)
来源：arXiv cs.CL · arxiv.org