Reproduce “DeepSearch-World: Self-Distillation for Deep Search Agents in a Verifiable Environment” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinyu Geng, Xuanhua He, Sixiang Chen, Yanjing Xiao, Fan Zhang, Shijue Huang, Haitao Mi, Zhenwen Liang and 2 more
This research introduces DeepSearch-World, a deterministic and verifiable offline environment, and DeepSearch-Evolve, an iterative self-distillation framework, enabling open-source deep search agents to improve their multi-hop reasoning and tool-use capabilities from self-generated, verified experience, achieving competitive performance on various benchmarks without relying on proprietary models.
Key insights
- A deterministic and verifiable environment with process-level supervision (entity-level verification, grounded reflection) is crucial for stable and scalable self-evolution in long-horizon agentic tasks.
- Scaffold process supervision, which internalizes planning, memory tracking, and error recovery into the agent's reasoning trace via ReAct-format conversion, enables the acquisition of complex cognitive behaviors.
- The diversity and scale of the data pool significantly impact sustained improvement and generalization, with a 420,000-instance QA pool leading to higher performance plateaus and more stable gains than a smaller one.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.