Navalia：可验证任务合成与长程后训练提升协作智能体
原标题：Navalia: Advancing Cowork Agents through Verifiable Task Synthesis and Long-Horizon Post-Training
AI 摘要
该论文提出 Navalia 框架，用于训练能自主规划、调用工具并交付可验证结果的 Cowork Agent。它从企业财报构建文档与事实的双视图表示，由 LLM 语义筛选证据并构建带真实数据依赖的可执行计算链，再通过确定性执行、独立 LLM 审查和程序化检查验证题目、证据、计算过程与参考答案的一致性。训练采用冷启动 SFT 加强化学习的两阶段流程，基于 Qwen3.6-35B-A3B 在五个基准上均优于基座模型，部分评测接近领先开源与闭源模型。
正文节选
Advancing Cowork Agents through Verifiable Task Synthesis and Long-Horizon Post-Training numbers,square,comma,sortcompress \fancyhead 1 Introduction Agents based on large language models (LLMs) [Wang et al.(2024)Wang, Ma, Feng, Zhang, Yang, Zhang, Chen, Tang, Chen, Lin, et al.] are evolving from conversational assistants [Bai et al.(2022)Bai, Jones, Ndousse, Askell, Chen, DasSarma, Drain, Fort, Ganguli, Henighan, et al., Touvron et al.(2023)Touvron, Martin, Stone, Albert, Almahairi, Babaei, Bash