Self-Assist: Deliberate Tool Selection by Large Language Models

Anonymous

Abstract: With the swift progress in tool-based learning, the number of tools available has also increased significantly. In comparison to the correct utilization of tools, the significance of precisely choosing the appropriate tool from an increasingly large selection is crucial. At present, depending solely on retrieval methods that use keywords and embeddings faces new challenges in the realm of tool selection. On one hand, it becomes difficult to distinguish between tools with similar functionalities. On the other hand, some queries require further reasoning to uncover the true tool needs, where direct matching with keywords or semantic embeddings does not yield the correct result. To address this issue, we introduce the Self-Assist method, which fully leverages the inherent knowledge and reasoning capabilities of large language models. Through a series of systematic steps, large language models actively engage in deliberate thought and select the most appropriate tool for a given query. In essence, our work champions a blend of LLMs and retrieval tools in a flexible, efficient, and universally compatible design, significantly bolstering retrieval outcomes. Evaluations on three datasets reveal superior performance over the previous approaches in retrieval accuracy and overall success.

Paper Type: long

Research Area: NLP Applications

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis, Theory

Languages Studied: English