Reproduce “Talk to Right Specialists: Iterative Routing in Multi-agent Systems for Question Answering” 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
Feijie Wu, Zitao Li, Fei Wei, Yaliang Li, Bolin Ding, Jing Gao
Researchers from Purdue University and Alibaba developed RIRS, a training-free framework for multi-agent question answering that routes queries to specialized RAG agents based on embedding-space knowledge summaries and iteratively refines questions for complex queries. This approach improved accuracy by up to 15% over existing learning-based routing methods in distributed RAG environments, while preserving knowledge sovereignty.
Key insights
- Centroid-based knowledge summarization provides an accurate and privacy-preserving method for characterizing agent capabilities without sharing raw data or requiring router retraining.
- Iterative query refinement and response composition are crucial for effectively addressing complex, multi-hop questions that require information from multiple specialized RAG agents.
- A robust server-side LLM controller for response evaluation and question simplification is critical for mitigating error propagation and improving overall accuracy, especially when using weaker agent-side LLMs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.