AI Paper Atlas

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

NeurIPS 2023 · Large language models › Agents, code and math · approach: Execution-based code verification · 18 citations (OpenAlex)

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 changedWhereMeasuredEffect
learn toolken embeddings instead of in-context ReAct tool use with 13 toolsLLaMA-33B · FuncQA one-hopaccuracy: 0.57 -> 0.73Improved
learn toolken embeddings instead of ReActLLaMA-33B · FuncQA multi-hopaccuracy: 0.06 -> 0.15Improved
toolkens instead of ReAct with only 4 basic arithmetic toolsLLaMA-33B · GSM8K-XLaccuracy: 0.32 -> 0.33No clear effect
toolkens instead of grounded-decoding in-context learningLLaMA-13B · VirtualHomeSuccess: 0.38 -> 0.68Improved
add tool mode (tool-specific demonstrations) to ReAct (ablation)ReAct · FuncQA one-hopaccuracy: 0.57 -> 0.60Improved
add toolken embeddings on top of ReAct + tool mode (ablation)ReAct + Tool mode · FuncQA one-hopaccuracy: 0.60 -> 0.73Improved
fine-tune with LoRA instead of training toolken embeddingsLLaMA-7B · FuncQA one-hopaccuracy: 0.55 -> 0.62; 2 min -> 40 min trainingMixed
increase supervised training examples per tool from 10 to 40 (ablation)ToolkenGPT (sup) · KAMEL 30 toolsaccuracy: 0.56 -> 0.95Improved
increase synthetic training examples per tool from 10 to 40 (ablation)ToolkenGPT (syn) · KAMEL 30 toolsaccuracy: 0.36 -> 0.52Improved
use supervised instead of synthetic demonstration data at 40 examples (ablation)ToolkenGPT · KAMEL 30 toolsaccuracy: 0.52 -> 0.95Improved
in-context learning with more than 30 toolsICL baseline · KAMELaccuracyHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy