Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

ToolACE: Winning the Points of LLM Function Calling

Created by
  • Haebom

Author

Weiwen Liu, Xu Huang, Xingshan Zeng, Liu, Enhong Chen

Outline

In this paper, we present ToolACE, a high-quality and diverse training data generation pipeline for improving the function-calling capability of large-scale language models. To overcome the limitations of existing synthetic data generation methods, ToolACE constructs an API pool containing 26,507 diverse APIs through a self-evolving synthesis process, and generates conversational data through multiple agent interactions and formalized thought processes. Rule-based and model-based validation systems ensure data accuracy, and we show that an 8 billion-parameter model trained with the generated data achieves state-of-the-art performance on the Berkeley Function-Calling Leaderboard. The model and some of the data are publicly available.

Takeaways, Limitations

•
Takeaways:
◦
We present ToolACE, a novel data generation pipeline that contributes to improving the function call performance of large-scale language models.
◦
Solve the problem of insufficient coverage and accuracy of existing synthetic data __T117226_____.
◦
Generating diverse and complex data through self-evolving synthetic processes and multi-agent interactions.
◦
Ensuring high data accuracy through rule-based and model-based verification systems.
◦
Achieving state-of-the-art performance even on models with relatively small parameters (8 billion).
◦
Promoting research sharing and advancement through disclosure of generated data and models.
•
Limitations:
◦
Further validation of the generalization performance of data generated by ToolACE is needed.
◦
Lack of clear discussion of the differences and limitations from real data.
◦
Potential loss of generality due to dependency on specific API pools.
◦
Applicability review for a wider range of function call scenarios is needed.
👍
Made with Slashpage