Towards Completeness-Oriented Tool Retrieval for Large Language Models

Anonymous

Abstract: Recently, enhancing the capabilities of Large Language Models (LLMs) through interaction with external tools has gathered widespread interest, where tool retrieval emerges as a crucial step. Existing tool retrieval approaches only focus on semantic matching. However, effective tool retrieval requires consideration of collaborative invocation among multiple tools rather than solely evaluating the utility of individual tools, which presents a challenge to existing tool retrieval methods. To address this, we propose a novel COllaborative Learning-based Tool Retrieval approach, COLT, which manages not only the semantic matching between user queries and tool descriptions but also takes into account the collaborative information of tools. Extensive experiments on both the open benchmark and the introduced ToolLens dataset show that COLT achieves superior performance. Notably, the performance of BERT-mini (11M) with our COLT framework outperforms BERT-large (340M), which has 30 times more parameters. Our codes and data are publicly available at https://anonymous.4open.science/r/COLT-4D13.

Paper Type: long

Research Area: Information Retrieval and Text Mining

Contribution Types: Publicly available software and/or pre-trained models, Data resources, Data analysis

Languages Studied: English