Back to feed
arXiv cs.CL
arXiv cs.CL
7/10/2026
DeepSearch-World: Self-Distillation for Deep Search Agents in a Verifiable Environment

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

Short summary

A self-distillation framework (DeepSearch-Evolve) enables web agents to improve autonomously in a verifiable environment containing 420K multi-hop QA tasks. The resulting DeepSearch-World-9B model achieves competitive performance (93.4% on HotpotQA, 61.5% on GAIA, 31.2% on BrowseComp) without requiring distillation from more capable models. Full code, trained model, environment, and dataset are being released to enable scalable self-evolution research.

  • •Self-distillation framework enables web agents to learn from their own experience without teacher models
  • •420K multi-hop QA dataset in deterministic environment with progress verification and failure recovery
  • •Competitive benchmark results achieved (93.4% HotpotQA) with full code, model, and data release

Generated with AI, which can make mistakes.

#ai-agents#research-breakthrough#open-source
Read full article at arXiv cs.CL

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

SEAD: Competence-Aware On-Policy Distillation via Entropy-Guided Supervision
arXiv cs.CL
Recursive Self-Evolving Agents via Held-Out Selection
arXiv CS.AI
Cost-Effective Agent Harnesses for Abstract Reasoning and Generalization on ARC-AGI-1
arXiv CS.AI
Better Starts, Better Ends: Bootstrapped Iterative Self-Reasoning Distillation for Compressed Reasoning
arXiv cs.CL