Reproduce “ToolACE-DEV: Self-Improving Tool Learning via Decomposition and EVolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xu Huang, Weiwen Liu, Xingshan Zeng, Yuefeng Huang, Xinlong Hao, Yuxian Wang, Yirong Zeng, Chuhan Wu and 3 more
ToolACE-DEV introduces a self-improving framework that enables lightweight Large Language Models to autonomously enhance their tool-learning capabilities through decomposition and iterative self-evolution. This approach allows LLaMA3.1-8B-Instruct to achieve state-of-the-art performance on the BFCL leaderboard, outperforming many larger and proprietary models, while significantly reducing reliance on expensive external data synthesis.
Key insights
- Lightweight LLMs can achieve state-of-the-art tool-using performance by autonomously generating and learning from their own data, significantly reducing the need for expensive external advanced models.
- Decomposing the complex task of tool learning into explicit tool documentation adaptation, query-aware tool generation, and tool invocation tasks is crucial for enabling effective self-improvement.
- The self-evolution process, particularly with self-consistency decoding, effectively produces high-quality training data that continuously enhances model performance, especially on challenging, live-like scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.