API Mapping using Self Reflection in Large Language Models

ACL ARR 2025 February Submission6619 Authors

Abstract: API sequencing has received significant attention, leading to the development of datasets and solutions aimed at generating correct sequences of API calls for complex tasks. However, little work has been done on the task of API mapping—a novel task that involves identifying and linking functionally equivalent endpoints across different tools to accomplish tasks that require accessing multiple functionalities, much like performing joins in database querying. In this work, we focus on mapping APIs within the ToolBench dataset. By leveraging rich annotation resources and a self-reflection mechanism to iteratively refine mapping decisions, our approach identifies common fields and functional overlaps between API endpoints, thereby enabling the integration of multiple tools for multi-faceted tasks. Unlike existing systems that rely on the function calling capabilities of frontier models and require meticulously organized API data, our method demonstrates that effective API mapping can be achieved with smaller open source models and more flexible data organization, thereby providing a more accessible and cost-efficient solution for real-world applications.

Paper Type: Short

Research Area: Resources and Evaluation

Research Area Keywords: API mapping, Large Language Models, Self-Reflection

Contribution Types: Model analysis & interpretability, Data resources

Languages Studied: English

Submission Number: 6619