AI News AI资讯 4h ago Updated 1h ago 更新于 1小时前 52

Stanford Researchers Introduce TRACE: A Capability-Targeted Agentic Training System That Turns Recurrent Agent Failures Into Synthetic RL Environment 斯坦福研究人员推出TRACE:一种能力定向代理训练系统,将反复出现的代理失败转化为合成强化学习环境

TRACE is an open-source agentic training system that converts recurrent agent failures into targeted training environments by diagnosing missing reusable capabilities. The system employs a four-step pipeline: contrastive capability analysis, targeted environment synthesis, capability adapter training via GRPO, and Mixture-of-Experts (MoE) composition with token-level routing. TRACE identifies specific deficits such as structured data reasoning, multi-step task completion, precondition verificati TRACE系统通过对比能力分析诊断智能体缺失的可复用技能,将重复性失败转化为针对性训练信号。 采用四步自动化流水线:对比能力分析、定向环境合成、LoRA适配器训练及MoE组合路由。 在Qwen3-30B-A3B上,TRACE使τ²-Bench通过率提升15.3分,SWE-bench Verified提升15分,显著优于基线。 该方法样本效率高,使用不到四分之一rollout数据即超越GRPO和GEPA等最强外部基准。 开源发布(MIT许可证),解决了传统RL/SFT奖励稀疏且无法精准定位技能缺陷的问题。

72
Hot 热度
78
Quality 质量
75
Impact 影响力

Analysis 深度分析

TL;DR

  • TRACE is an open-source agentic training system that converts recurrent agent failures into targeted training environments by diagnosing missing reusable capabilities.
  • The system employs a four-step pipeline: contrastive capability analysis, targeted environment synthesis, capability adapter training via GRPO, and Mixture-of-Experts (MoE) composition with token-level routing.
  • TRACE identifies specific deficits such as structured data reasoning, multi-step task completion, precondition verification, and tool calling precision, isolating them for focused training.
  • On Qwen3-30B-A3B, TRACE achieved a 48.2% pass rate on τ²-Bench and 41.0% Pass@1 on SWE-bench Verified, significantly outperforming baselines like GEPA and SWE-RL.
  • The approach is highly sample-efficient, requiring less than one-fourth of the rollouts needed by standard GRPO or prompt optimization methods while delivering superior accuracy.

Why It Matters

This research addresses a critical inefficiency in current agentic AI development: the wasteful allocation of compute resources on broad, untargeted training data. By diagnosing specific capability gaps and creating verifiable, synthetic environments for each, TRACE allows practitioners to train models precisely where they fail, leading to faster convergence and better performance. For the industry, this offers a scalable, automated pathway to improve agent reliability without relying heavily on expensive human labeling or massive, generic datasets.

Technical Details

  • Contrastive Capability Analysis: The pipeline analyzes agent rollouts, splitting them into success and failure sets. It labels trajectory-capability pairs as NA, PRESENT, or LACKING, retaining only capabilities with a contrastive gap (δ ≥ 0.20) and coverage (ρ ≥ 0.10) to ensure the identified deficits are concentrated in failures.
  • Targeted Environment Synthesis: For each retained capability deficit, a generation agent creates a synthetic environment that isolates that specific skill while preserving original tool schemas. Task instances are procedurally generated from random seeds, allowing for algorithmic verification without human labels or LLM judges.
  • Capability Adapter Training: Each capability is assigned a dedicated LoRA adapter trained using Group Relative Policy Optimization (GRPO). The base model remains frozen, and rewards are normalized within groups sharing the same seed to isolate the policy's contribution effectively.
  • MoE Composition: The trained adapters are composed into a Mixture-of-Experts model. Lightweight token-level gates are trained to route each token top-1 to a single capability expert during inference, enabling the model to switch experts mid-trajectory based on immediate needs.
  • Benchmark Performance: Evaluated on Qwen3-30B-A3B and Qwen3.6-27B across τ²-Bench, SWE-bench Verified, and ToolSandBox, TRACE demonstrated significant improvements over base models and existing baselines like GEPA and SWE-RL, particularly in customer service and software engineering tasks.

Industry Insight

  • Shift from Broad to Targeted Training: AI teams should move away from blanket fine-tuning strategies and adopt diagnostic pipelines that identify specific capability bottlenecks before investing in extensive training runs.
  • Cost Efficiency in Agentic Systems: By leveraging synthetic, algorithmically verifiable environments, organizations can reduce dependency on costly human-in-the-loop annotation processes, making continuous agent improvement more economically viable.
  • Modular Expertise via MoE: Implementing Mixture-of-Experts architectures with token-level routing allows for dynamic skill activation, potentially improving model robustness and reducing inference latency by activating only relevant capabilities for each token.

TL;DR

  • TRACE系统通过对比能力分析诊断智能体缺失的可复用技能,将重复性失败转化为针对性训练信号。
  • 采用四步自动化流水线:对比能力分析、定向环境合成、LoRA适配器训练及MoE组合路由。
  • 在Qwen3-30B-A3B上,TRACE使τ²-Bench通过率提升15.3分,SWE-bench Verified提升15分,显著优于基线。
  • 该方法样本效率高,使用不到四分之一rollout数据即超越GRPO和GEPA等最强外部基准。
  • 开源发布(MIT许可证),解决了传统RL/SFT奖励稀疏且无法精准定位技能缺陷的问题。

为什么值得看

TRACE提出了一种从“盲目试错”转向“精准补强”的智能体训练范式,揭示了智能体失败往往源于少数可复用的能力缺失而非随机错误。对于AI从业者而言,这提供了一种高效利用计算资源、通过模块化专家组合提升复杂任务表现的具体技术路径。

技术解析

  • 对比能力分析:将轨迹分为成功/失败集,标记能力为NA/PRESENT/LACKING。仅保留对比差距δ≥0.20且覆盖率ρ≥0.10的能力,确保训练信号集中在导致失败的关键缺陷上。
  • 定向环境合成:为每个保留的能力生成一个独立的合成环境,保留目标工具模式但隔离单一能力。通过算法化生成和验证提供无需人工标注或LLM裁判的密集奖励信号。
  • 能力适配器训练:为每个能力训练一个独立的LoRA适配器,使用GRPO算法进行优化。GRPO按共享种子分组rollout并在组内归一化奖励,以隔离策略贡献,基础模型保持冻结。
  • MoE组合与路由:将多个LoRA适配器组合成混合专家(MoE)模型。训练轻量级的token级门控网络,推理时每个token通过top-1路由到单个能力适配器,允许智能体在轨迹中动态切换专家。
  • 基准测试结果:在τ²-Bench和SWE-bench Verified上,TRACE在Qwen3-30B-A3B和Qwen3.6-27B上均取得显著性能提升,特别是在结构化数据推理、多步骤任务完成等特定能力上表现突出。

行业启示

  • 从通用优化转向能力模块化:智能体开发应关注识别和修复具体的“能力缺口”,而非仅仅优化整体提示或进行泛化的强化学习,模块化专家组合可能成为提升复杂任务可靠性的主流架构。
  • 自动化数据工程的价值:TRACE证明了通过算法自动生成可验证的训练环境和奖励信号,可以大幅降低对昂贵人工标注的依赖,提高训练的数据效率和可扩展性。
  • 计算资源的精准投放:在智能体训练中,区分“已有能力”和“缺失能力”至关重要。TRACE的方法表明,将计算资源集中在少数高频失败原因上,能以更少的样本获得更高的性能增益,这对降低大模型训练成本具有战略意义。

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 Open Source 开源 Research 科学研究