Researchers propose DeepSearch-World and DeepSearch-Evolve for self-improving web agents in verifiable environments

Researchers from multiple institutions propose DeepSearch-World, a deterministic and verifiable environment designed to support training of web agents with reproducible search and page-reading tools. The environment is constructed to enable key agentic cognitive behaviors—progress verification, grounded reflection, and failure recovery—necessary for self-evolving systems.

The paper introduces DeepSearch-Evolve, a self-distillation framework that iteratively performs trajectory generation, filtering, data mixing, and fine-tuning to train stronger agents without relying on distillation from more capable teacher models. This addresses challenges in both supervised fine-tuning, which depends on fixed teacher-distilled trajectories, and sparse-reward reinforcement learning, which provides weak supervision for long-horizon interactions.

DeepSearch-World contains 420K multi-hop QA tasks constructed from entity-level random walks, providing a large-scale benchmark for training and evaluating agents in verifiable settings. The authors report that a 9B-parameter model trained with DeepSearch-Evolve achieves competitive performance on three benchmarks: 31.2% on BrowseComp, 61.5% on GAIA, and 93.4% on HotpotQA.

The authors state they will release the environment, the 420K training pool, validation set, model, and code to support further research on self-improving deep search agents. The work is presented as a step toward scalable self-evolution in long-horizon web agent tasks without external teacher models.