Reproduce “Multi-Field Tool Retrieval” 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

Yichen Tang, Weihang Su, Yiqun Liu, Qingyao Ai

A new Multi-Field Tool Retrieval (MFTR) framework enhances large language model capabilities by enabling more precise tool selection from diverse repositories. The framework standardizes tool documentation and aligns user queries to multiple tool utility fields, resulting in consistent performance gains of up to 47.98% in Recall@10 on benchmarks like Gorilla.

Key insights

  • Tool retrieval fundamentally differs from traditional ad-hoc text retrieval and requires a multi-aspect modeling approach to accurately capture tool utility.
  • Standardizing heterogeneous tool documentation and performing structured, tool-aware query rewriting are crucial for bridging the semantic and granularity gaps between user intent and tool specifications.
  • Dynamically weighting the importance of different functional fields and incorporating an adaptive penalty for missing parameters is essential for enhancing both relevance and functional correctness in tool selection.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.