DeepSearch-World: Self-Distillation for Deep Search Agents in a Verifiable Environment

DeepSearch-World：可验证环境下的深度搜索智能体自蒸馏框架

Abstract: Training tool-use agents to improve from their own experience remains challenging, as supervised fine-tuning relies on fixed teacher-distilled trajectories, while sparse-reward reinforcement learning provides weak supervision for long-horizon interactions.

摘要： 训练工具使用型智能体（tool-use agents）使其能够从自身经验中提升能力仍然是一项挑战，因为监督微调依赖于固定的教师模型蒸馏轨迹，而稀疏奖励的强化学习在长程交互中仅能提供微弱的监督信号。

We present DeepSearch-Evolve, a self-distillation framework for web agents built on DeepSearch-World, a deterministic and verifiable environment with reproducible search and page-reading tools.

我们提出了 DeepSearch-Evolve，这是一个专为网络智能体设计的自蒸馏框架，构建于 DeepSearch-World 之上。DeepSearch-World 是一个确定性且可验证的环境，配备了可复现的搜索和页面阅读工具。

DeepSearch-World contains 420K multi-hop QA tasks constructed from entity-level random walks and supports key agentic cognitive behaviors useful for self-evolving, including progress verification, grounded reflection, and failure recovery.

DeepSearch-World 包含 420,000 个多跳问答（multi-hop QA）任务，这些任务通过实体级随机游走构建而成，并支持对自我进化至关重要的关键智能体认知行为，包括进度验证、基于事实的反射（grounded reflection）以及故障恢复。

DeepSearch-Evolve iteratively performs trajectory generation, filtering, data mixing, and fine-tuning to train stronger agents.

DeepSearch-Evolve 通过迭代执行轨迹生成、过滤、数据混合和微调，旨在训练出更强大的智能体。

Without distillation from more capable models, DeepSearch-World-9B achieves competitive performance compared with open-source agents, reaching 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA, showing that verifiable environments enable scalable self-evolution for long-horizon web agents.

在没有从更强模型进行蒸馏的情况下，DeepSearch-World-9B 与开源智能体相比表现出了极具竞争力的性能，在 BrowseComp 上达到 31.2%，在 GAIA 上达到 61.5%，在 HotpotQA 上达到 93.4%。这表明可验证环境能够为长程网络智能体实现可扩展的自我进化。

We will release the environment, 420K training pool, validation set, model, and code to facilitate future research on self-improving deep search agents.

我们将发布该环境、420K 训练池、验证集、模型及代码，以促进未来关于自我提升型深度搜索智能体的研究。