Research Papers 论文研究 6h ago Updated 1h ago 更新于 1小时前 45

Speculative Macro Commit for Faster Tool-Using Agents 用于加速工具使用智能体的投机宏提交

Speculative Macro Commit (SMC) is a runtime mechanism for two-tier agent systems that uses a fast speculative drafter model to predict and pre-execute future action chains on isolated environment snapshots while a larger authoritative actor model produces the official trajectory SMC mines recurring multi-action skeletons from training traces into a macro library, enabling runtime matching against drafter-predicted action chains to commit pre-executed steps when the actor's next tool call aligns 提出Speculative Macro Commit (SMC)机制,通过两级agent系统(大模型actor + 小模型draft)加速工具使用型LLM agent的推理过程 从训练轨迹中挖掘多动作骨架构建宏库,运行时匹配draft预测的动作链,实现多步推测性执行而非单步推测 在τ²-Bench Telecom子集上,SMC相比顺序执行降低18.59%延迟,相比SA基线降低10.23%延迟,且保持整体准确率 在AppWorld基准上,SMC相比顺序执行减少44.9%墙钟时间,相比SA基线减少7.7%,任务完成率仅有微小下降 使用Qwen3.5-27B INT4作为权威actor模型,Qwen

58
Hot 热度
72
Quality 质量
65
Impact 影响力

Analysis 深度分析

TL;DR

  • Speculative Macro Commit (SMC) is a runtime mechanism for two-tier agent systems that uses a fast speculative drafter model to predict and pre-execute future action chains on isolated environment snapshots while a larger authoritative actor model produces the official trajectory
  • SMC mines recurring multi-action skeletons from training traces into a macro library, enabling runtime matching against drafter-predicted action chains to commit pre-executed steps when the actor's next tool call aligns with the first drafted action
  • Evaluated with Qwen3.5-27B INT4 as the actor and Qwen3.5-4B as the drafter, SMC matches sequential agent accuracy while reducing latency by 10.23% over Speculative Actions baseline and 18.59% over sequential execution on tau^2-Bench Telecom subset
  • On AppWorld, SMC achieves 7.7% wall-time reduction over SA baseline and 44.9% over sequential execution with only a small reduction in task completion rate
  • SMC extends speculative execution beyond single-step actions to multi-step macro-level speculation, offering a practical latency-reduction strategy for tool-using LLM agents

Why It Matters

This work addresses a critical bottleneck in LLM agent deployment: the serial action-observation loop that dominates wall-clock time in tool-using agents. By introducing macro-level speculative execution, SMC demonstrates that multi-step action chains can be predicted and pre-committed, significantly improving throughput without sacrificing accuracy. For AI practitioners building production agent systems, this represents a scalable optimization path that leverages existing two-tier model architectures without requiring architectural overhauls.

Technical Details

  • SMC employs a two-tier architecture: a large authoritative actor model (Qwen3.5-27B INT4) generates the official trajectory, while a smaller, faster speculative drafter model (Qwen3.5-4B) continuously predicts and executes future action chains on isolated environment snapshots
  • A macro library is constructed by mining recurring multi-action skeletons from training traces, which are then matched against drafter-predicted action chains at runtime to enable batch commitment of pre-executed steps
  • Commit logic triggers when the actor's next tool call matches the first drafted action, at which point the remaining pre-executed draft steps and their observations are committed to the official trajectory
  • Benchmarks include tau^2-Bench Telecom subset and AppWorld, with latency measurements showing 10.23% improvement over Speculative Actions (SA) baseline on tau^2-Bench and 7.7% over SA on AppWorld
  • The approach trades a small reduction in task completion rate for substantial latency gains, with 44.9% wall-time reduction over pure sequential execution on AppWorld

Industry Insight

  • Multi-step speculative execution represents a significant advancement over single-step speculation, suggesting that future agent optimizations should focus on pattern mining from historical trajectories to build reusable macro libraries
  • The two-tier actor-drafter architecture with quantized models (INT4) demonstrates that production-grade latency improvements are achievable without requiring expensive full-precision model deployments
  • The small accuracy-latency tradeoff observed on AppWorld indicates that SMC is production-viable for latency-sensitive applications, and practitioners should consider deploying SMC in scenarios where response time is critical and minor task completion reductions are acceptable

TL;DR

  • 提出Speculative Macro Commit (SMC)机制,通过两级agent系统(大模型actor + 小模型draft)加速工具使用型LLM agent的推理过程
  • 从训练轨迹中挖掘多动作骨架构建宏库,运行时匹配draft预测的动作链,实现多步推测性执行而非单步推测
  • 在τ²-Bench Telecom子集上,SMC相比顺序执行降低18.59%延迟,相比SA基线降低10.23%延迟,且保持整体准确率
  • 在AppWorld基准上,SMC相比顺序执行减少44.9%墙钟时间,相比SA基线减少7.7%,任务完成率仅有微小下降
  • 使用Qwen3.5-27B INT4作为权威actor模型,Qwen3.5-4B作为推测性draft模型,代码已公开

为什么值得看

本文针对工具使用型LLM agent的核心瓶颈——串行action-observation循环导致的延迟问题,提出了多步推测性执行的新思路。SMC通过宏库匹配机制实现高效的动作链预执行,为agent系统的实时性优化提供了可落地的技术方案,对构建低延迟、高吞吐的agent应用具有重要参考价值。

技术解析

  • 两级架构设计:SMC采用权威actor模型(大模型)与推测性draft模型(小模型)分离的架构。Actor模型生成官方轨迹,而draft模型在隔离的环境快照上持续预测并执行未来动作链,实现并行推测。
  • 宏库匹配机制:从训练轨迹中挖掘 recurring multi-action skeletons( recurring多动作骨架),存储于宏库中。运行时,draft模型预测的动作链与宏库进行匹配,当actor的下一个工具调用与draft的第一个动作匹配时,提交预执行的剩余步骤及其观测结果。
  • 模型配置与实验设置:使用Qwen3.5-27B INT4作为权威actor模型,Qwen3.5-4B作为推测性draft模型。在τ²-Bench Telecom子集和AppWorld两个基准上进行评估,对比基线包括Speculative Actions (SA)和顺序执行。
  • 性能提升显著:在τ²-Bench Telecom上,SMC相比SA基线延迟降低10.23%,相比顺序执行降低18.59%;在AppWorld上,墙钟时间相比SA基线降低7.7%,相比顺序执行降低44.9%,准确率仅有微小下降。

行业启示

  • 多步推测是agent加速的关键方向:当前agent系统多采用单步推测(如SA方法),SMC证明多步动作链的推测性执行能带来更大延迟优化空间,未来研究可进一步探索更长的动作链推测策略。
  • 宏库/模板复用可提升agent效率:从历史轨迹中挖掘可复用的动作模式并构建宏库,是一种低成本高效率的优化手段,适合在特定领域(如电信、应用操作)中部署专用agent。
  • 大小模型协同架构的实用价值:SMC验证了"大模型决策+小模型推测"的两级架构在agent系统中的可行性,为资源受限场景下的agent部署提供了参考范式。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent LLM 大模型 Inference 推理 Research 科学研究