MiroThinker：通过模型、上下文与交互式扩展推动开源研究代理性能边界
计算与语言
2026-04-22 v3
摘要
我们提出MiroThinker v1.0，这是一个开源研究代理，旨在推进工具增强推理与信息获取能力。不同于以往仅通过放大模型规模或上下文长度来提升性能，MiroThinker从模型层面探索交互式扩展，系统训练模型处理更深刻且更频繁的代理-环境交互，这是第三维度的性能提升。不同于LLM测试时扩展，这类方法在孤立环境中运行，随推理链长度增长可能导致性能下降，交互式扩展则利用环境反馈和外部信息获取纠正错误、优化轨迹。通过强化学习，模型实现了高效的交互式扩展：在256K上下文窗口下，可执行每项任务最高600次工具调用，实现持续的多轮推理和复杂的实际研究工作流程。在四个代表性基准测试——GAIA、HLE、BrowseComp和BrowseComp-ZH上，72B规模版本分别实现了81.9%、37.7%、47.1%和55.6%的准确率，超越了之前的开源代理，接近GPT-5-high等商业级模型。我们的分析表明，MiroThinker在交互式扩展方面始终受益于该方法：研究性能随模型在更深入、更频繁的代理-环境交互中的表现而可预测地提升，显示示交互深度呈现类似模型规模和上下文长度的扩展行为。这一发现确立了交互式扩展作为构建下一代开源研究代理的第三关键维度，补充了模型容量与上下文窗口的作用。
引用
@article{arxiv.2511.11793,
title = {MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling},
author = {MiroMind Team and Song Bai and Lidong Bing and Carson Chen and Guanzheng Chen and Yuntao Chen and Zhe Chen and Ziyi Chen and Jifeng Dai and Xuan Dong and Wenhan Dou and Yue Deng and Yunjie Fu and Junqi Ge and Chenxia Han and Tammy Huang and Zhenhang Huang and Jerry Jiao and Shilei Jiang and Tianyu Jiao and Xiaoqi Jian and Lei Lei and Ruilin Li and Gen Luo and Tiantong Li and Xiang Lin and Ziyuan Liu and Zhiqi Li and Jie Ni and Qiang Ren and Pax Sun and Shiqian Su and Chenxin Tao and Bin Wang and Wenhai Wang and Haonan Wang and James Wang and Jin Wang and Jojo Wang and Letian Wang and Shizun Wang and Weizhi Wang and Zixuan Wang and Jinfan Xu and Sen Xing and Chenyu Yang and Hai Ye and Jiaheng Yu and Yue Yu and Muyan Zhong and Tianchen Zhao and Xizhou Zhu and Yanpeng Zhou and Yifan Zhang and Zhi Zhu},
journal= {arXiv preprint arXiv:2511.11793},
year = {2026}
}
备注
Technical Report