符毓
(2025-09-30 23:42):
#paper doi: 10.48550/arXiv.2509.13311, 2025, Towards General Agentic Intelligence via Environment Scaling.
以往训练这类“代理智能”的主要瓶颈在于缺乏高质量、大规模、多样化的交互数据。人工标注成本极高,而单纯用模型生成的数据又往往不够真实或难以验证。这篇由阿里巴巴通义实验室团队发表的论文(通过环境扩展迈向通用代理智能)提出了一条全新的路径:通过程序化、自动化地构建海量、异构、可验证的模拟环境,让语言模型能在其中自主交互、收集经验、学习成长。基于该方法训练的AgentScaler模型系列,仅用数十亿参数就在多项权威测试中达到了与万亿级模型或闭源商业系统媲美的性能,为高效、轻量级代理智能的发展打开了新的可能性。
arXiv,
2025-09-16T17:57:20Z.
DOI: 10.48550/arXiv.2509.13311
Towards General Agentic Intelligence via Environment Scaling
翻译
Abstract:
Advanced agentic intelligence is a prerequisite for deploying Large LanguageModels in practical, real-world applications. Diverse real-world APIs demandprecise, robust function-calling intelligence, which needs agents to developthese capabilities through interaction in varied environments. The breadth offunction-calling competence is closely tied to the diversity of environments inwhich agents are trained. In this work, we scale up environments as a steptowards advancing general agentic intelligence. This gives rise to two centralchallenges: (i) how to scale environments in a principled manner, and (ii) howto effectively train agentic capabilities from experiences derived throughinteractions with these environments. To address these, we design a scalableframework that automatically constructs heterogeneous environments that arefully simulated, systematically broadening the space of function-callingscenarios. We further adapt a two-phase agent fine-tuning strategy: firstendowing agents with fundamental agentic capabilities, then specializing themfor domain-specific contexts. Extensive experiments on agentic benchmarks,tau-bench, tau2-Bench, and ACEBench, demonstrate that our trained model,AgentScaler, significantly enhances the function-calling capability of models.
翻译
Related Links: