ToolScope: Enhancing LLM Agent Tool Use through Tool Merging and Context-Aware Filtering

ACL ARR 2025 May Submission4216 Authors

Abstract: Large language model (LLM) agents rely on external tools to solve complex tasks. However, real-world toolsets often include tools with overlapping names and descriptions, leading to ambiguity in tool selection and degradation in reasoning performance. Furthermore, the limited input context available to LLM agents constrains their ability to effectively utilize information from a large number of tools. To address these challenges, we propose ToolScope, a two-part workflow which contains: (1) ToolScopeMerger, an automated tool-merging framework that reduces semantic redundancy on tools and offers observability to users to refine their toolset. (2) ToolScopeFilter, which mitigates the limitations imposed by LLMs’ context length by retrieving the top-k relevant tools for a given query. This selective filtering reduces input length, enabling more efficient tool usage without compromising selection accuracy. Experimental results on open-source benchmarks demonstrate that ToolScope significantly improves tool selection accuracy, achieving a 34.5% improvement on Seal-Tools and a 18.3% improvement on BFCL compared to baseline methods.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: LLM/AI agents, retrieval, knowledge graphs, word embeddings

Contribution Types: NLP engineering experiment

Languages Studied: English, Code

Submission Number: 4216