Thought Blocks：多智能体非自回归潜在通信
原标题：Thought Blocks: Non-Autoregressive Latent Communication for Multi-Agent Systems via Trajectory-to-State Compilation
AI 摘要
该论文提出 Thought Blocks，一种面向 LLM 多智能体系统的非自回归潜在通信方法。作者区分了消息表示与消息构造，提出轨迹到状态编译（trajectory-to-state compilation），用一次性构造的固定大小潜在 KV 续写记忆替代自回归潜在轨迹生成。实验显示，相比文本通信，文本 token 减少 45.7%、自回归前向步数减少 84.8%、端到端延迟平均降低 8.6 倍；相比自回归潜在通信，前向步数再减少 35.7%，准确率差异很小。
正文节选
Thought Blocks: Non-Autoregressive Latent Communication for Multi-Agent Systems via Trajectory-to-State Compilation Abstract LLM-based multi-agent systems typically communicate through text, incurring repeated generation and re-encoding costs. Although latent communication removes explicit verbalization, constructing latent messages may still require autoregressive latent trajectories. We therefore distinguish message representation from message construction and ask whether agents can directly c