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.
This paper discusses tool learning, which has emerged as a promising approach to extend the capabilities of large-scale language models (LLMs). Existing tool learning approaches have primarily focused on data synthesis to fine-tune LLMs to effectively invoke tools, but have largely neglected methods to fully exploit the model's potential. This paper proposes ToolACE-R, a novel framework that incorporates both model-aware iterative learning and adaptive improvement. ToolACE-R features a model-aware iterative learning procedure that incrementally adjusts training samples based on the model's evolving capabilities to maximize its potential. Furthermore, it incorporates a self-improving training corpus, highlighting the LLM's ability to iteratively optimize tool invocation without external feedback. Furthermore, we introduce an adaptive self-improvement mechanism for efficient testing time extension, enabling the trained model to autonomously decide when to stop the iterative self-improvement process. Extensive experiments on multiple benchmark datasets demonstrate that ToolACE-R achieves competitive performance compared to advanced API-based models. Adaptive self-improvement can efficiently further enhance tool invocation performance. These results highlight the effectiveness and generalizability of ToolACE-R and suggest promising directions for more efficient and scalable tool learning.
Takeaways, Limitations
•
Takeaways:
◦
We demonstrate that the potential of LLM can be maximized through a model-aware iterative learning procedure that adjusts training samples according to the model's evolving capabilities.
◦
We demonstrate the effectiveness of a self-improving training corpus that optimizes LLM's tool invocation performance without external feedback.
◦
We propose that test time extension can be efficiently performed through an adaptive self-improvement mechanism.
◦
We experimentally demonstrate that ToolACE-R achieves competitive performance compared to advanced API-based models.
•
Limitations:
◦
This paper lacks a detailed description of the specific algorithms and implementation details of ToolACE-R.
◦
Further research is needed on generalization performance across different types of tools and tasks.
◦
Further research is needed to determine the optimal parameters of the adaptive self-improvement mechanism.
◦
Disclosure of code and data is required to ensure reproducibility of experimental results.
