arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2602.03219cs.AI
收藏

超越数量:代码代理的轨迹多样性扩展

Beyond Quantity: Trajectory Diversity Scaling for Code Agents

SIAT, CAS(中国科学院软件研究所);Alibaba Group(阿里巴巴集团);UNSW Sydney(新南威尔士大学悉尼分校);SUAT(上海大学)
Guhong Chen, Chenghao Sun, Cheng Fu, Qiyao Wang, Zhihong Huang, Chaopeng Wei, Guangxu Chen, Feiteng Fang, Ahmadreza Argha, Bing Zhao, Xander Xu, Qi Han, Hamid A… 展开作者

Guhong Chen, Chenghao Sun, Cheng Fu, Qiyao Wang, Zhihong Huang, Chaopeng Wei, Guangxu Chen, Feiteng Fang, Ahmadreza Argha, Bing Zhao, Xander Xu, Qi Han, Hamid Alinejad-Rokny, Qiang Qu, Binhua Li, Shiwen Ni, Min Yang, Hu Wei, Yongbin Li

AI总结:

TDScaling通过提升轨迹多样性而非单纯增加数据量,优化代码代理训练的性能-成本平衡,提升工具使用泛化和编码能力。

摘要:

随着代码大语言模型(LLMs)通过模型上下文协议(MCP)演变为工具交互代理,其泛化能力越来越受到低质量合成数据和数量扩展的边际效益递减的限制。此外,以数量为中心的扩展存在一个早期瓶颈,导致轨迹数据的利用不足。我们提出了TDScaling,一种基于轨迹多样性的数据合成框架,用于代码代理,通过多样性而非原始体积来提升性能。在固定训练预算下,增加轨迹多样性带来的收益大于增加更多轨迹,从而改善代理训练的性能-成本权衡。TDScaling集成了四项创新:(1)业务集群机制,捕捉真实服务的逻辑依赖;(2)蓝图驱动的多代理范式,强制轨迹一致性;(3)自适应进化机制,利用领域熵、推理模式熵和累积动作复杂度引导合成向长尾场景发展,以防止模式崩溃;(4)沙盒代码工具,缓解内在编码能力的灾难性遗忘。在通用工具使用基准(BFCL,tau^2-Bench)和代码代理任务(RebenchT,CodeCI,BIRD)上的实验表明,TDScaling实现了双赢:既提高了工具使用泛化能力,又提升了固有编码能力。我们计划在发表时发布完整的代码库和合成数据集(包括30000+工具集群)。

英文摘要:

As code large language models (LLMs) evolve into tool-interactive agents via the Model Context Protocol (MCP), their generalization is increasingly limited by low-quality synthetic data and the diminishing returns of quantity scaling. Moreover, quantity-centric scaling exhibits an early bottleneck that underutilizes trajectory data. We propose TDScaling, a Trajectory Diversity Scaling-based data synthesis framework for code agents that scales performance through diversity rather than raw volume. Under a fixed training budget, increasing trajectory diversity yields larger gains than adding more trajectories, improving the performance-cost trade-off for agent training. TDScaling integrates four innovations: (1) a Business Cluster mechanism that captures real-service logical dependencies; (2) a blueprint-driven multi-agent paradigm that enforces trajectory coherence; (3) an adaptive evolution mechanism that steers synthesis toward long-tail scenarios using Domain Entropy, Reasoning Mode Entropy, and Cumulative Action Complexity to prevent mode collapse; and (4) a sandboxed code tool that mitigates catastrophic forgetting of intrinsic coding capabilities. Experiments on general tool-use benchmarks (BFCL, tau^2-Bench) and code agent tasks (RebenchT, CodeCI, BIRD) demonstrate a win-win outcome: TDScaling improves both tool-use generalization and inherent coding proficiency. We plan to release the full codebase and the synthesized dataset (including 30,000+ tool clusters) upon publication.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑