Reproduce “Argus: Evidence Assembly for Scalable 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
Zhen Zhang, Liangcai Su, Zhuo Chen, Xiang Lin, Haotian Xu, Simon Shaolei Du, Kaiyu Yang, Bo An and 2 more
MiroMind AI's Argus system introduces a cooperative multi-agent architecture for deep research, leveraging a dynamic evidence graph to enable compositional evidence assembly and targeted information seeking. The system achieved state-of-the-art accuracy on five benchmarks, demonstrating log-linear performance scaling and a 1,200:1 context compression ratio for the orchestrating agent.
Key insights
- Reframing deep research as compositional evidence assembly on a shared graph enables more efficient, less redundant information gathering compared to independent parallel rollouts.
- A structured evidence graph allows the orchestrating agent's context to remain compact by providing a summarized view, effectively decoupling its reasoning from the vast amount of raw search output.
- Training a Navigator with a contrastive reinforcement learning reward successfully teaches the agent to identify weaknesses in gathered evidence and dispatch targeted verification queries.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.