跳到正文
返回论文列表
cs.MA提交于 已译

揭示分歧胜过解决分歧:面向多智能体辩论的来源加权记忆的门控消融研究

Surfacing Disagreement Beats Resolving It: A Gated Ablation Study of Provenance-Weighted Memory for Multi-Agent Debate

Ziaullah Khan · Abdul Rehman Khalid · Shrestha Nischal Lal · Nheng Vanchhay · Md Ariful Islam Mozumder · Hee-Cheol Kim

中文摘要

多智能体辩论的共享记忆通常被设计为消除不一致:检索返回单一胜出事实,分歧在任何智能体看到之前就消失。研究检验这一做法是否正确。在 DebateGraph 中,四个后端共享同一契约(读/写/快照),因此在固定预算下可独立移除智能体记忆文献中的五种机制,每种机制均以预先声明的阈值作为门控条件。在 39 次运行中,仅有一种机制存活:保留矛盾并将排序后的冲突集传给辩论,而非将其坍缩为单一胜出者,可使 F1 提升 +0.340(95% CI [+0.267, +0.414]),并将攻击成功率从 122 个投毒项中的 22 个降至 7 个(McNemar 精确检验 p=0.0003),且成本相同。实现该效果需修复测试框架中的瓶颈:冲突集仅基于检索到的边构建,而预先声明的上限导致四分之三的植入矛盾无法被检索,首次测量该机制时仅得 +0.084。该效应在两个检索基底上可复现,并在 HotpotQA 文档上呈方向性一致。其他机制均未存活,其中两种比惰性更差:一旦矛盾进入提示,写入信任门控与路由器的来源特征分别损失准确率(F1 为 -0.075 和 -0.096)。门控在结构上失效:投毒对象始终是图中已存在的实体,而金标准中该比例为 53%,因此基于接地性的评分将谎言排在真实首次提及之上。研究发布测试框架、攻击集与日志,并主张静默调和记忆是值得放弃的假设。

关键要点

  1. 01共享记忆通常消除分歧,研究检验保留矛盾是否更优。
  2. 02保留冲突集使 F1 提升 +0.340,攻击成功率从 22/122 降至 7/122。
  3. 03修复检索瓶颈后效应可复现,首次测量仅 +0.084。
  4. 04写入信任门控与来源特征反而损失 F1(-0.075 和 -0.096)。
  5. 05门控因投毒对象均为图中已有实体而结构性失效。

解读

尚无解读。

原始英文摘要

arXiv:2611.22461v1 Announce Type: cross Abstract: Shared memory for multi-agent debate is usually built to resolve inconsistency: retrieval returns one winning fact and the disagreement disappears before any agent sees it. We ask whether that is correct. In DebateGraph, four backends share one contract (read/write/snapshot), so five mechanisms from the agent-memory literature can be removed independently under a fixed budget, each gated on a pre-declared threshold. Across 39 runs one survives: preserving contradictions and passing the ranked conflict set to the debate, rather than collapsing it to a winner, improves F1 by +0.340 (95% CI [+0.267, +0.414]) and cuts attack success from 22 to 7 of 122 poisoned items (McNemar exact p=0.0003), at equal cost. Reaching it meant fixing a bottleneck in our harness: conflict sets are built only over retrieved edges, and our pre-declared cap left three quarters of planted contradictions unretrievable, first measuring the mechanism at +0.084. The effect replicates on two retrieval substrates and directionally on HotpotQA documents. No other mechanism survives, and two are worse than inert: once contradictions reach the prompt, a write trust-gate and the router's provenance features each cost accuracy (-0.075 and -0.096 F1). The gate fails structurally: poisoned objects are always entities the graph already contains, against 53% of gold, so groundedness scores the lie above a truthful first mention. We release the harness, attack set, and logs, and argue that silently reconciling memory is the assumption worth abandoning.

同方向论文 · cs.MA

查看全部 →