Enhanced tool invocation method through multi-model collaboration

Abstract This paper proposes a multi-model collaborative framework to address the format non-compliance and limited generalization observed in tool-invocation tasks when Large Language Models are directly fine-tuned. Our study reveals that the success of tool invocation is highly sensitive to the standardization of output formats. However, conventional fine-tuning methods often cause models to rigidly adhere to specific formats, which in turn increases the risk of parsing failures. To overcome this limitation, we design a division-of-labor approach: the large model focuses on reasoning and tool selection, generating preliminary invocation outputs, while the small model is dedicated to format normalization, converting these outputs into structured, API-compliant requests. Experiments conducted using the Berkeley Function Call Evaluation benchmark demonstrate that format standardization plays a more critical role in invocation success than semantic correctness. Furthermore, perturbation experiments, such as reordering parameters or altering bracket types, show that even minor deviations significantly degrade performance. In contrast, our collaborative framework mitigates these issues by introducing a post-processing formatting step, ultimately improving accuracy. This work offers new insights into tool learning and the interplay between reasoning and format control in LLM-based tool invocation.