Reproduce “AgentDisCo: Towards Disentanglement and Collaboration in Open-ended Deep Research Agents” 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
Jiarui Jin, Zexuan Yan, Shijian Wang, Wenxiang Jiao, Yuan Lu
Researchers at Xiaohongshu Inc. developed AgentDisCo, a deep research agent framework that disentangles information exploration and exploitation into adversarial yet collaborative agents, enhancing iterative refinement. The system, featuring a meta-optimization harness for self-evolving strategies and a new lifestyle-oriented GALA benchmark, demonstrated consistent performance gains across multiple benchmarks and improved citation accuracy to 89.06%.
Key insights
- Disentangling information exploration and exploitation into distinct, adversarially-collaborative agents can significantly improve the flexibility, stability, and incremental refinement capabilities of deep research systems.
- Automated meta-optimization through a self-evolving harness can systematically discover and refine agentic design strategies, leading to more robust and controllable agent execution without extensive human intervention.
- Developing benchmarks derived from authentic, lifestyle-oriented user data, such as the GALA benchmark, is essential for advancing deep research agents that are genuinely applicable and performant for real-world information-seeking behaviors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.