ToolWeaver: Weaving Collaborative Semantics for Scalable Tool Use in Large Language Models
Published in International Conference on Learning Representations (ICLR 2026), 2026
Mapping each tool to a unique new token causes the vocabulary to explode and leaves every tool semantically isolated. ToolWeaver instead encodes tools into hierarchical sequences generated by a tokenization process that weaves together a tool’s intrinsic semantics with its co-usage patterns, and aligns the LLM to generate these codes. Evaluated with nearly 47,000 tools, it significantly outperforms state-of-the-art methods.
Recommended citation: Bowen Fang, Wen Ye, Yunyue Su, Jinghao Zhang, Qiang Liu, Yesheng Liu, Xin Sun, Shu Wu, Jiabing Yang, Baole Wei, Liang Wang. (2026). "ToolWeaver: Weaving Collaborative Semantics for Scalable Tool Use in Large Language Models." International Conference on Learning Representations (ICLR), pp. 90654-90684.
Download Paper