文章背景与核心概要
现代大语言模型（LLM）智能体在处理复杂任务时，往往受制于串行的“动作-观察”（action-observation）交互循环。每一次工具调用、环境状态转换和观察结果反馈都会带来显著的系统延迟。为打破这一瓶颈，本文作者提出了推测性宏指令提交（Speculative Macro Commit, SMC）机制，专为双层智能体系统设计：包含生成权威轨迹的大型决策模型（Actor），以及在隔离环境快照上持续预测和执行未来动作链的快速草稿模型（Drafter）。
SMC 通过从训练轨迹中挖掘重复出现的复合动作骨架，并将其存储在宏指令库中。在运行时，系统将这些模板与草稿模型预测的动作链进行匹配。当权威模型的下一个工具调用与草稿的第一步动作相匹配时，SMC 便可直接将剩余预执行的草稿步骤及其观察结果“提交”到官方轨迹中。实验表明，在 \(\tau^2\)-Bench（电信子集）和 AppWorld 等基准测试中，SMC 在保持智能体准确率的同时，大幅降低了运行延迟，为多步骤推测执行在智能体领域的应用提供了高效实用的解决方案。
Speculative Macro Commit for Faster Tool-Using Agents
Speculative Macro Commit for Faster Tool-Using Agents
Summary
Summary
Tool-using Large Language Model (LLM) agents are often bottlenecked by serial action–observation turns, where each tool call, environment transition, and observation creates significant latency. To address this, the authors introduce Speculative Macro Commit (SMC), a runtime mechanism designed for a two-tier agent system: 1. Authoritative Actor Model: A large model that produces the official trajectory. 2. Speculative Drafter Model: A faster model that continuously predicts and executes future action chains on an isolated environment snapshot.
SMC mines recurring multi-action skeletons from training traces and stores them in a macro library. At runtime, these are matched against action chains predicted by the drafter. When the actor's next tool call matches the first drafted action, SMC commits the remaining pre-executed draft steps—along with their observations—directly to the official trajectory.
Key Results
Key Results
- \(\tau^2\)-Bench (Telecom Subset): Using Qwen3.5-27B INT4 (actor) and Qwen3.5-4B (drafter), SMC matches sequential agent accuracy while reducing latency by 10.23% over the Speculative Actions (SA) baseline and 18.59% over standard sequential execution.
- AppWorld: SMC reduces wall-clock time by 7.7% over the SA baseline and 44.9% over sequential execution, accompanied by only a minor reduction in task completion.
Paper Metadata
Paper Metadata
- arXiv ID: arXiv:2609.03236 [cs.AI]
- Authors: Zeyu Liu, Souvik Kundu, Peter A. Beerel
- Submitted: September 3, 2026
- Primary Subject: Artificial Intelligence (
cs.AI)- Secondary Subjects: Multiagent Systems (
cs.MA)- Comments: Accepted in MLSP 2026
- Official Code Repository: GitHub - speculative-macro-commit
Abstract
工具使用型大语言模型（LLM）智能体不仅在模型推理上消耗实际运行时间（wall-clock time），还在串行的“动作-观察”交互轮次中耗费大量时间，其中每一次工具调用、环境状态转换和观察结果都会延迟后续的决策。为此，我们引入了推测性宏指令提交（Speculative Macro Commit, SMC），这是一种针对双层智能体系统的运行时机制：一个大型的权威行动者模型（Actor）生成官方轨迹，而一个更快的推测性草稿模型（Drafter）则在隔离的环境快照上持续预测并执行未来的动作链。SMC 从训练轨迹中挖掘重复的多动作骨架，并将其存储在宏指令库中，用于在运行时与草稿模型预测的动作链进行匹配。当行动者的下一个工具调用与草稿的第一步动作匹配时，SMC 会将剩余的预执行草稿步骤及其观察结果直接提交到官方轨迹中。使用 Qwen3.5-27B INT4 作为权威行动者模型、Qwen3.5-4B 作为推测草稿模型，在 \(\tau^2\)-Bench 电信子集上，SMC 在保持串行智能体总体准确率的同时，相比推测动作（SA）基线降低了 10.23% 的延迟，相比串行执行降低了 18.59% 的延迟。在 AppWorld 上，SMC 的实际运行时间比 SA 基线减少了 7.7%，比串行执行减少了 44.9%，同时任务完成率仅有微小的下降。总体而言，SMC 提供了一种实用方法，能够复用多步推测执行，并进一步超越单步推测动作，从而减少智能体延迟。
Tool-using LLM agents spend wall-clock time not only on model inference but also in serial action--observation turns, where each tool call, environment transition, and observation can delay subsequent decisions. We introduce Speculative Macro Commit (SMC), a runtime mechanism for a two-tier agent system: a large authoritative actor model produces the official trajectory, while a faster speculative drafter model continuously predicts and executes future action chains on an isolated environment snapshot. SMC mines recurring multi-action skeletons from training traces and stores them in a macro library used to match against action chains predicted by the drafter at runtime. When the actor's next tool call matches the first drafted action, SMC commits the remaining pre-executed draft steps, together with their observations, to the official trajectory. Using Qwen3.5-27B INT4 as the authoritative actor model and Qwen3.5-4B as the speculative drafter model, SMC matches the sequential agent's overall accuracy while reducing latency by 10.23\% over the Speculative Actions (SA) baseline and 18.59\% over sequential execution on the \(\tau^2\)-Bench Telecom subset. On AppWorld, SMC reduces wall time by 7.7\% over SA baseline and 44.9\% over sequential execution, with a small reduction in task completion. Overall, SMC provides a practical way to reuse multi-step speculative execution and reduce agent latency beyond single-step speculative actions.