跳转至

潜在通信何时真正发挥作用？多智能体大模型中中继KV缓存的因果审计

文章背景与核心概要

在多智能体大模型（LLM）系统中，智能体之间往往直接中继键值（KV）缓存而非纯文本，并将性能的提升归功于“潜在思想（latent thoughts）”的交互。然而，这种性能增益究竟是源于特定样本之间的有意义信息传递，还是仅仅由于缓存机制本身带来的正则化或上下文扩展效果？本文针对现有开源系统中的这一核心声明进行了严格的因果审计。

通过在不同实验范式下，将正常的KV缓存替换为错乱的（不匹配样本的）、清零的和矩匹配的随机对应物，作者评估了性能收益的真实来源。研究结果揭示了一个微妙的现实：虽然某些系统确实成功地通过潜在通道利用了私有信息，但其他系统则几乎没有表现出可测量的、依赖于特定样本配对的优势。这表明，大规模的缓存效应并不等同于有效的配对传输效应，基准测试的性能提升并不能直接证明潜在思想的成功传输，验证这一点需要进行本文所提出的错乱缓存审计。

📌 Summary

Multi-agent Large Language Model (LLM) systems often relay key-value (KV) caches instead of plain text, attributing performance improvements to the exchange of "latent thoughts." This paper presents a causal audit of this underlying claim in existing released systems. By replacing KV caches with deranged (mismatched-example), zeroed, and moment-matched random counterparts across different regimes, the authors evaluate whether the performance gains truly stem from meaningful example-specific transfer.

Their findings reveal a nuanced reality: while some systems successfully leverage private information via latent channels, others show little to no measurable advantage from example-specific pairings, demonstrating that large cache effects do not automatically equate to valid pairing effects.

📑 Metadata & Reference

🔬 Abstract

多智能体大语言模型（LLM）系统通过中继键值（KV）缓存来代替文本，并将性能的提升归功于所交换的“潜在思想”。这种归因本质上是对“究竟中继了哪个样本的缓存”这一决策的肯定，而不仅仅是承认中继了缓存这一事实。我们在现有的开源系统中对这一底层声明进行了因果审计。在由接收方是否需要发送方私有信息所定义的两种实验范式下，我们将缓存替换为了错乱的（不匹配样本的）、清零的和矩匹配的随机对应物。

在需要私有信息的范式下，测试结果达到了上限：在主要的骨干网络上，相关联的中继达到了 100% 的准确率，而与答案无关的中继仅为 23–25%，这一鲜明对比在三个模型家族、五个检查点以及散文文档问答（prose document-QA）场景中得到了复现。而在不需要私有信息的范式下，通过在 GSM8K 和 ARC-Challenge 上跨三个 Qwen3 尺度以及在 MedQA 的 8B 模型上采用经 Holm 校正的 TOST（两单侧检验），预注册的五随机种子协议在 2.8 个点（该数值锚定自被审计系统报告的增益）的界限内建立起了等效性（其中有一个单元格在边界内检测到了微弱优势）；而第二个模型家族则未检测到任何优势。

大规模的缓存效应并不一定构成配对效应。在某个自然测试单元中，将中继清零会导致 14.7 个点的性能损失，而不匹配的缓存则仅导致 0.4 个点。并且，“需求”本身也不足以说明问题：在相同的测试下，不同系统交付的通道表现各异，包括达到上限（LatentMAS 的原生中继）、部分有效（KVComm 的层子集）以及未检测到特定样本传输（C2C 的开源投影器）。基准测试的性能差（deltas）本身并不能确立潜在思想的传输；要确立这一点，必须进行我们所公开的错乱缓存审计。

Multi-agent LLM systems relay key-value caches instead of text and credit their gains to exchanged "latent thoughts." That credit is a claim about which example's cache is relayed, not merely that one is. We audit it causally in released systems. The cache is replaced with deranged (mismatched-example), zeroed, and moment-matched random counterparts, under two regimes defined by whether the receiver needs the sender's private information.

Where it does, the battery reads ceiling: 100% against 23–25% for answer-irrelevant relays on the primary backbone, a contrast replicated across three families, five checkpoints, and a prose document-QA surface. Where it does not, a pre-registered five-seed protocol establishes equivalence within 2.8 points, a margin anchored to the audited system's reported gain, under Holm-corrected TOST on GSM8K and ARC-Challenge across three Qwen3 scales and on MedQA at 8B (one cell shows a small detected advantage inside the margin); a second family shows no detected advantage.

A large cache effect need not be a pairing effect. In one natural cell, zeroing the relay costs 14.7 points; a mismatched cache, 0.4. Nor is need sufficient: under the same test, delivered channels span ceiling (LatentMAS's native relay), partial (KVComm's layer subset), and no detected example-specific transfer (C2C's released projector). Benchmark deltas do not by themselves establish latent-thought transmission; establishing it takes a mismatched-cache audit, which we release.