多智能体代码评判何时有据可依？两项无标签度量与拒答机制
原标题：When Is a Multi-Agent Code Judge Actually Grounded? Two Label-Free Measurements, and a Judge That Declines to Guess
AI 摘要
研究团队将已发布的验证框架 MARCH 原封不动地运行在代码评判基准 CodeJudgeBench 和 HumanEvalFix 上，用 qwen3-8b 与 qwen3-14b 作为评判模型，共约 2000 次比较。结果显示该多智能体流程从不报错，却在 78% 至 95% 的比较中判定两个方案同样好，准确率仅 4.4%，远低于同一模型直接评判的 43.7%。作者提出两项无需标签的度量：76.6% 的比较中系统对两个方案提出相同问题，检查器对 80.7% 的待验证主张表示同意；据此在运行检查器前过滤掉提问相同的比较，可将准确率从 20.7% 提升至 36.9%，但仍不及直接评判的 73.4%。
正文节选
When Is a Multi-Agent Code Judge Actually Grounded? Two Label-Free Measurements, and a Judge That Declines to Guess Software teams increasingly let one language model write code and a second decide whether that code is correct enough to merge [7, 1]. The second model acts as a safety check, and it has a failure mode that accuracy alone does not reveal. When it lacks any basis for a decision, it does not report that fact. It returns a confident verdict with reasoning attached, and that verdict is