A study on classification based concurrent API calls and optimal model combination for tool augmented LLMs for AI agent

TL;DR

This study improves AI agent performance by using multiple tools and models together, boosting accuracy and reducing errors.

Contribution

Proposes concurrent API calls and optimal model combination for tool-augmented LLMs.

Findings

Concurrent tool calls improve accuracy by 4.4–9.3% compared to single-tool approaches.

Optimal model combination reduces response errors by up to 9%.

Multi-step reasoning benefits from using both enhanced and existing LLM models.

Abstract

AI Agents have evolved to not only recommend content but also facilitate information retrieval and task processing. Developing AI Agents using general-purpose LLM models necessitates integration with external tools, leading to tool-augmented LLM studies. Despite the availability of multiple tools for the same purpose, existing research has not fully leveraged this diversity. This study categorizes external tools by type and proposes a method to simultaneously call tools of the same type. This allows for the utilization of diverse external tools in LLM inference, thereby achieving a higher accuracy compared to when only a single tool for one task is used. Experimental results show an accuracy improvement of 4.4–9.3% over existing studies. Furthermore, when utilizing tool-augmented LLM, a multi-step reasoning approach that divides the process into stages such as planning and tool invocation is widely employed. With the rapid advancement of LLMs, enhanced models continue to emerge. Considering the trade-offs between performance and cost in models, it is crucial to find an optimal combination of models in each stage of tool augmented LLM. In this study, we propose a novel method for efficiently utilizing both enhanced LLM models and existing models, which reduces response errors by up to 9%.

The online version contains supplementary material available at 10.1038/s41598-025-06469-w.

Figures