DeepSearch-World: AI agents that self-evolve without teacher models
A 9B-parameter model rivals larger agents by learning from its own experience
Training AI agents to use tools effectively over multiple steps is notoriously hard. Supervised fine-tuning relies on fixed trajectories from a teacher model, while reinforcement learning struggles with sparse rewards. A new paper from Xinyu Geng and nine co-authors tackles this with DeepSearch-Evolve, a self-distillation framework built on DeepSearch-World—a deterministic, verifiable environment with reproducible search and page-reading tools.
The environment contains 420K multi-hop QA tasks constructed from entity-level random walks, supporting key cognitive behaviors like progress verification, grounded reflection, and failure recovery. DeepSearch-Evolve iteratively performs trajectory generation, filtering, data mixing, and fine-tuning to train increasingly stronger agents. Remarkably, their 9B-parameter model (DeepSearch-World-9B) achieves competitive performance without any distillation from more capable models: 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA. The team will release the environment, training pool, validation set, model, and code to accelerate research on self-improving deep search agents.
- 420K multi-hop questions created from entity-level random walks for training
- Self-distillation framework iteratively improves agents without teacher models
- 9B-parameter model scores 31.2% on BrowseComp, 61.5% on GAIA, 93.4% on HotpotQA
Why It Matters
Self-evolving web agents could reduce reliance on costly human-annotated training data and enable scalable autonomous research.