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

The paper introduces DeepSearch-Evolve, a self-distillation framework that improves web agents through iterative trajectory generation, filtering, data mixing, and fine-tuning. Its DeepSearch-World environment provides deterministic, reproducible search and page-reading tools, together with 420K multi-hop QA tasks generated from entity-level random walks. The resulting 9B agent reportedly reaches 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA, without distillation from stronger models. The authors say they will release the environment, training pool, validation set, model, and code.