DeepSearch-World: A Revolutionary Self-Distillation Framework for Web Agents

DeepSearch-World, introduced by Xinyu Geng and colleagues, presents a self-distillation framework aimed at enhancing deep search agents. This innovative system was detailed in a paper submitted on July 8, 2026, catering to the challenges faced in training tool-use agents to learn from their own experiences.

Overview of DeepSearch-World Framework

The DeepSearch-World framework is designed to operate in a deterministic and verifiable environment, which is crucial for the reproduction of search tasks. It encompasses 420,000 multi-hop QA tasks derived from entity-level random walks, allowing web agents to exhibit key cognitive behaviors necessary for self-evolving.

Key features of this environment include:

• Progress verification
• Grounded reflection
• Failure recovery

Performance Insights and Achievements

Utilizing the DeepSearch-Evolve framework, agents undergo a rigorous training process that includes trajectory generation, data filtering, and fine-tuning. Remarkably, the DeepSearch-World-9B model has demonstrated competitive performance when compared to existing open-source agents, achieving:

• 31.2% on BrowseComp
• 61.5% on GAIA
• 93.4% on HotpotQA

These results highlight the scalability of self-evolution for long-horizon web agents within a verifiable environment.

Future Research Directions and Resources

The authors plan to release the DeepSearch-World environment along with the 420K training pool, validation set, model, and associated code. This initiative aims to facilitate further research into self-improving deep search agents, paving the way for advancements in computational language processing.

For those interested in exploring this framework, the full paper is available for viewing as a PDF.