Speculative Macro Commit for Faster Tool-Using Agents

推测性宏提交：加速工具使用型智能体

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. 工具使用型大模型智能体不仅在模型推理上消耗挂钟时间，还在串行的“动作-观察”循环中耗时，其中每一次工具调用、环境转换和观察都会延迟后续的决策。

We introduce \textbf{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. 我们引入了“推测性宏提交”（Speculative Macro Commit, SMC），这是一种用于双层智能体系统的运行时机制：由一个大型权威执行模型生成官方轨迹，同时由一个更快的推测性草拟模型在隔离的环境快照上持续预测并执行未来的动作链。

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. SMC 从训练轨迹中挖掘重复出现的多动作骨架，并将它们存储在宏库中，用于在运行时与草拟模型预测的动作链进行匹配。

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. 当执行模型的下一次工具调用与草拟的第一个动作匹配时，SMC 会将剩余的预执行草拟步骤及其观察结果提交到官方轨迹中。

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. 使用 Qwen3.5-27B INT4 作为权威执行模型，Qwen3.5-4B 作为推测性草拟模型，SMC 在 $\tau^2$-Bench 电信子集上的表现与串行智能体保持了相同的整体准确率，同时相比“推测性动作”（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。

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. 在 AppWorld 测试中，SMC 相比 SA 基线减少了 7.7% 的挂钟时间，相比串行执行减少了 44.9%，任务完成率仅有小幅下降。

Overall, SMC provides a practical way to reuse multi-step speculative execution and reduce agent latency beyond single-step speculative actions. Our code is publicly available here. 总的来说，SMC 提供了一种实用的方法来复用多步推测执行，并超越了单步推测动作，进一步降低了智能体的延迟。我们的代码已公开在此处。