对抗式审查：结构化分歧提升多智能体代码审查质量
2 AI agents set to check each other's work will usually end up agreeing, whether…
Agent 协作中的共识偏差是常见痛点，这篇论文给出了具体的结构化干预方案和量化效果，对构建可靠 Agent 工作流有直接参考价值。
2 AI agents set to check each other's work will usually end up agreeing, whether or not the code is right.
两个互相检查对方工作的 AI 代理通常会达成一致，无论代码是否正确。
So the thing to add is not another reviewer but a rule about what an objection has to contain before either side is allowed to drop it.
因此需要添加的不是另一个审查者，而是一条规则：规定异议在允许任何一方放弃之前必须包含什么内容。
This paper shows a reviewer plus a critic beating much larger review teams at writing code, then failing at reviewing code until that rule is in place.
这篇论文展示了一个审查者加一个批评者组合在编写代码方面击败了规模大得多的审查团队，但在该规则到位之前，它们在审查代码时却表现不佳。
The structure is small: the code stays frozen while a reviewer writes a review and a critic audits it, and only the settled review goes back for edits.
其结构很小：代码保持冻结状态，由审查者撰写审查意见，批评者对其进行审计，只有定稿的审查意见才会返回进行修改。
On LiveCodeBench it reaches 87% with 3 agents, against 82% for a 5-agent version.
在 LiveCodeBench 上，使用 3 个代理达到了 87% 的准确率，而 5 个代理的版本仅为 82%。
On real pull-request review it lands last, at 0.457 F1. One prompt change fixes that: the critic must state whether its objection cites code or is only a hunch, and the reviewer has to answer with code either way, which takes it to 0.533 and the top of the set.
在实际的拉取请求（pull-request）审查中，它排在最后，F1 分数为 0.457。只需更改一个提示词即可解决此问题：批评者必须说明其异议是否引用了代码，还是仅凭直觉；审查者无论如何都必须用代码作答，这使得 F1 分数提升至 0.533，并达到该集合的最高水平。
So the second agent only helps when agreement has to be paid for with code evidence; without that rule it mostly ratifies the first.
因此，第二个代理仅在同意必须以代码证据来换取时才有帮助；没有这条规则时，它主要只是对第一个代理的意见予以认可。
– arxiv. org/abs/2608.18167
– arxiv.org/abs/2608.18167
Title: "Adversarial Review: Structured Disagreement for Grounded Agentic Code Review"
标题：《对抗性审查：用于有根据的代理代码审查的结构化分歧》
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力