Reproduce “FS-Researcher: Test-Time Scaling for Long-Horizon Research Tasks with File-System-Based 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
Chiwei Zhu, Benfeng Xu, Mingxuan Du, Shaohan Wang, Xiaorui Wang, Zhendong Mao, Yongdong Zhang
A dual-agent framework called FS-Researcher enables large language model agents to perform extensive deep research by utilizing a persistent file-system workspace for unlimited information storage and iterative refinement. This approach overcomes LLM context window limitations, achieving state-of-the-art performance on research benchmarks through effective test-time scaling.
Key insights
- Externalizing an agent's research state and evidence into a durable file-system workspace allows for unlimited information storage and on-demand access, effectively bypassing LLM context window limitations.
- A dual-agent architecture, separating information acquisition (Context Builder) from report composition (Report Writer), prevents premature synthesis and ensures dedicated computational resources for each phase.
- Implementing persistent control files and iterative refinement across multiple sessions enables robust task progression, self-correction, and direct correlation between allocated computational resources and output quality (test-time scaling).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.