Research map / Large language models / Agents, code and math
ToolkenGPT: Augmenting Frozen Language Models with Massive Tools via Tool Embeddings
Shibo Hao, Tianyang Liu, Zhen Wang, Zhiting Hu
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 |
|---|---|---|---|
| learn toolken embeddings instead of in-context ReAct tool use with 13 tools | LLaMA-33B · FuncQA one-hop | accuracy: 0.57 -> 0.73 | Improved |
| learn toolken embeddings instead of ReAct | LLaMA-33B · FuncQA multi-hop | accuracy: 0.06 -> 0.15 | Improved |
| toolkens instead of ReAct with only 4 basic arithmetic tools | LLaMA-33B · GSM8K-XL | accuracy: 0.32 -> 0.33 | No clear effect |
| toolkens instead of grounded-decoding in-context learning | LLaMA-13B · VirtualHome | Success: 0.38 -> 0.68 | Improved |
| add tool mode (tool-specific demonstrations) to ReAct (ablation) | ReAct · FuncQA one-hop | accuracy: 0.57 -> 0.60 | Improved |
| add toolken embeddings on top of ReAct + tool mode (ablation) | ReAct + Tool mode · FuncQA one-hop | accuracy: 0.60 -> 0.73 | Improved |
| fine-tune with LoRA instead of training toolken embeddings | LLaMA-7B · FuncQA one-hop | accuracy: 0.55 -> 0.62; 2 min -> 40 min training | Mixed |
| increase supervised training examples per tool from 10 to 40 (ablation) | ToolkenGPT (sup) · KAMEL 30 tools | accuracy: 0.56 -> 0.95 | Improved |
| increase synthetic training examples per tool from 10 to 40 (ablation) | ToolkenGPT (syn) · KAMEL 30 tools | accuracy: 0.36 -> 0.52 | Improved |
| use supervised instead of synthetic demonstration data at 40 examples (ablation) | ToolkenGPT · KAMEL 30 tools | accuracy: 0.52 -> 0.95 | Improved |
| in-context learning with more than 30 tools | ICL baseline · KAMEL | accuracy | Hurt |
Closest papers by meaning
- EmbedLLM: Learning Compact Representations of Large Language Models (ICLR 2025 · similarity 0.90)
- Large Language Models as Tool Makers (ICLR 2024 · similarity 0.90)
- ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs (ICLR 2024 · similarity 0.89)
- Learning Evolving Tools for Large Language Models (ICLR 2025 · similarity 0.89)
- Joint Optimization of Tool Creation and Use for Large Language Model Agents (NeurIPS 2026 · similarity 0.89)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.88)
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.
