Research map / Large language models / Language, safety and interpretability
ToolWeaver: Weaving Collaborative Semantics for Scalable Tool Use in Large Language Models
Bowen Fang, Wen Ye, Yunyue Su, Jinghao Zhang, Qiang Liu, Yesheng Liu, Xin Sun, Shu Wu, Jiabing Yang, Baole Wei, Liang Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace atomic one-token-per-tool IDs with collaborative-aware hierarchical codes | generative tool retrieval (Llama-3-8B) · ToolBench retrieval I3 (multi-domain) | NDCG@1: 81.00 -> 88.00 | Improved |
| replace atomic IDs with ToolWeaver codes for end-to-end tool use | ToolGen-style agent (Llama-3-8B) · ToolBench end-to-end I3 | SoPR: 36.34 -> 52.19 | Improved |
| replace atomic IDs with ToolWeaver codes | ToolGen-style agent (Llama-3-8B) · ToolBench end-to-end I2 | SoPR: 45.13 -> 44.03 | Hurt |
| replace atomic IDs with ToolWeaver codes | ToolGen-style agent (Llama-3-8B) · ToolBench end-to-end I1 | SoPR: 52.97 -> 53.17 | No clear effect |
| replace atomic IDs with ToolWeaver codes | ToolGen-style agent (Llama-3-8B) · ToolBench retrieval I3 (in-domain) | NDCG@1: 87.00 -> 86.00 | No clear effect |
| add collaborative guidance (graph Laplacian regularization) to the tokenizer (ablation) | ToolWeaver tokenization · ToolBench retrieval (I1-I3) | NDCG@k | Improved |
| add semantic initialization from a text encoder (ablation) | ToolWeaver tokenization · ToolBench retrieval (I1-I3) | NDCG@k: over 20 NDCG points | Improved |
| increase collaborative weight lambda from 0.01 to 1 (ablation) | ToolWeaver tokenization · ToolBench retrieval (avg NDCG) | NDCG@k | Improved |
| increase collaborative weight lambda from 1 to 10 (ablation) | ToolWeaver tokenization · ToolBench retrieval (avg NDCG) | NDCG@k | Hurt |
| use hierarchical or semantic naming codes instead of atomic tokens (ablation) | ToolGen-style agent (Llama-3-8B) · ToolBench end-to-end I3 | SoPR: 36.34 -> 17.49 (hierarchical), 16.39 (semantic) | Hurt |
| use linear atomic vocabulary expansion (ToolGen) rather than logarithmic code expansion | Llama-3-8B · WikiText-2 | perplexity: 25.36 -> 104.54 | Hurt |
| add ToolWeaver codes to Llama-3-8B | Llama-3-8B · CNN/DailyMail summarization | BERTScore F1: 85.35 -> 85.07 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ToolGen — ToolGen: Unified Tool Retrieval and Calling via Generation (ICLR 2025) · better on 12 of 15 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ToolGen: Unified Tool Retrieval and Calling via Generation (ICLR 2025 · similarity 0.92)
- Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.89)
- From Exploration to Mastery: Enabling LLMs to Master Tools via Self-Driven Interactions (ICLR 2025 · similarity 0.88)
- ToolSearcher: Optimizing Tool Selection at Scale via Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
- Learning Evolving Tools for Large Language Models (ICLR 2025 · similarity 0.87)
- ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
