Augment the Cache, Replace the Activation: A 2x2 Study of Internal-State Sharing
作者: Adam Deng, Shicheng Liu, Yada Zhu
分类: cs.LG
发布日期: 2026-10-15
备注: 10 pages, 2 figures
💡 一句话要点
提出共享内部状态机制以优化多智能体系统通信
🎯 匹配领域: 支柱一：机器人控制 (Robot Control)
关键词: 多智能体系统 内部状态共享 通信优化 键值缓存 激活机制 智能代理 信息传递 系统效率
📋 核心要点
- 现有多智能体系统中，代理通过文本通信，导致内部状态的丢失和重建成本。
- 本研究提出共享内部状态的机制，代理将状态发布到共享池中，其他代理通过单一机制读取，简化了通信架构。
- 实验表明，增强KV缓存的方式能够在保持准确度的同时，减少通信成本，激活则直接写入接收者，提升了效率。
📝 摘要（中文）
多智能体系统中，基于冻结语言模型的代理通过文本进行通信，这种方式会丢失发送者已计算的内部状态，接收者需要重新构建。现有的内部通道方法将问题框架设定为信息转移，而本研究将其重新定义为共享：代理将内部状态发布到一个池中，其他代理通过单一机制读取该池。研究发现，内部状态的载体与融合机制是相互关联的，接收者的计算可消耗性是关键属性。通过增强键值缓存（KV cache）而非替换，接收者能够在保持自身缓存的同时交叉关注其他代理的缓存，准确性与直接读取源文档相当。激活则需直接写入接收者，且无需桥接、适配和训练，表现优于与接收者自身状态并行暴露的方式。
🔬 方法详解
问题定义：本研究旨在解决多智能体系统中代理间通信时内部状态丢失的问题。现有方法需要为每对代理建立独立的转移桥接，增加了系统复杂性和通信成本。
核心思路：论文提出通过共享内部状态的方式，代理将其状态发布到一个共享池中，其他代理通过单一机制读取该池，从而避免了为每个新代理建立新的桥接。
技术框架：整体架构包括状态发布模块、共享池和读取机制。每个代理在发布其内部状态后，其他代理可以通过统一的读取机制访问这些状态，简化了通信流程。
关键创新：最重要的创新在于将内部状态的载体与融合机制的选择结合起来，提出了增强KV缓存而非替换的策略，从而提高了接收者的计算效率。
关键设计：在设计中，接收者的KV缓存被增强以交叉关注其他代理的状态，同时保持自身缓存，且只需训练一个小的查询侧适配器，便可达到与直接读取源文档相当的准确性。
🖼️ 关键图片
📊 实验亮点
实验结果显示，增强KV缓存的策略使得接收者在保持自身缓存的同时，能够准确交叉关注其他代理的状态，准确性与直接读取源文档相当。此外，激活直接写入接收者的方式，表现优于传统的状态并行暴露，提升了整体系统性能。
🎯 应用场景
该研究的潜在应用领域包括多智能体协作系统、智能助手和复杂任务分配等场景。通过优化代理间的状态共享机制，可以显著提高系统的通信效率和响应速度，推动智能体技术在实际应用中的落地与发展。
📄 摘要（原文）
Multi-agent systems built from frozen language models communicate through text, which discards the internal state the sender already computed and makes the receiver pay to rebuild it. Work on internal channels avoids that cost, framing the problem as transfer: agent A sends information to agent B. Every agent pair that must communicate needs its own transfer bridge, and every agent added to a system needs a new bridge fitted against each previous agent. We reframe this problem as sharing: an agent publishes its internal state once into a pool, and every peer reads that pool through a single mechanism fitted to no sender in particular, so adding an agent adds a publisher and no new bridge. Whereas the carrier of the internal state and the fusion mechanism are usually chosen independently in the literature, our main finding is that these two choices are coupled; the deciding factor is where the carrier enters the receiver's computation. Consumability is a property of the receiving computation rather than of the representation. We find that the key-value (KV) cache must be augmented rather than swapped in: the receiver cross-attends over a peer's cache while keeping its own, and with only a small query-side adapter trained, matches the accuracy it gets from reading the source document. Activations invert this. They are best written straight into the receiver, with no bridge, no adapter and no training, which beats exposing them alongside the receiver's own, though that advantage rests on knowing where in the receiver the state belongs. The rule is augment the KV cache, replace the activation, established under one protocol with a shared control ladder. One adapter, trained once, also serves publishers it never trained on, and a pool of several at once, but only those whose internal state lies close to the receiver's own.