确定性 LLM 多智能体审议中的低裕度分支点审计研究
Low-Margin Branch Points in Deterministic LLM Deliberation
AI 导读
一项 arXiv 研究对五智能体委员会在贪心解码下的确定性 LLM 审议进行白盒审计，前瞻选取低裕度与高裕度 token 位置，注入范数相对残差位移，并测量十个生成 token 内的首次转写分歧。结果显示低裕度位置比高裕度对照更分支敏感，匹配的降裕度梯度方向分支频率显著更高。研究声明这是受控局部分支易感性审计，不估计部署频率或决策正确性。
正文 · 原文
Abstract:Deterministic multi-agent LLM deliberation can exactly re- play a fixed transcript, yet exact replay alone does not test local resistance to controlled internal interventions. Because each response becomes context for later agents, a changed greedy token can redirect a transcript-mediated execution. We present a white-box audit of this local failure mode in a five-agent committee under greedy decoding. We prospectively select low- and high-margin token positions, require exact no-op re- play, inject norm-relative residual displacements, and measure first transcript divergence within ten generated tokens. Low- margin positions are more branch-sensitive than high-margin controls, while matched margin-reducing gradient directions branch substantially more often than margin-increasing con- trols. We then continue selected realized branches through the committee to distinguish textual persistence from preference and final-decision changes. The contribution is a controlled audit for local branch susceptibility in this specified protocol, not an estimate of deployment frequency, decision correct- ness, or a phenomenon unique to multi-agent systems.
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2611.00546 [cs.AI] |
| (or arXiv:2611.00546v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2611.00546 arXiv-issued DOI via DataCite |
Submission history
From: Hajime Shimao [view email]
[v1]
Sat, 31 Oct 2026 02:01:39 UTC (108 KB)
[v2]
Mon, 9 Nov 2026 22:12:06 UTC (108 KB)
来源：arXiv cs.AI · arxiv.org