IC-986Most LLMs lack tool usage awareness, with only ChatGPT exceeding 70% F1 in zero-shot evaluation

Yue Huang, Jiawen Shi, Yuan Li, Chenrui Fan, Siyuan Wu, Qihui Zhang, Yixin Liu, Pan Zhou, Yao Wan, Neil Zhenqiang Gong, Lichao Sun

SourceMetaTool Benchmark for Large Language Models: Deciding Whether to Use Tools and Which to Use

The paper evaluates whether eight LLMs can recognise when a user query requires an external tool versus one they can answer directly. Under zero-shot prompting, only ChatGPT achieves both accuracy and F1 above 70% (74.85% acc, 77.81 F1). Llama2-13b scores an F1 of just 11.53% and Vicuna-7b 13.43%. Five-shot prompting improves some models substantially (Llama2-13b F1 rises by 42.79 points, Vicuna-7b by 42.28), yet even the best non-ChatGPT model (Vicuna-33b) reaches only 69.12 F1. The authors conclude that most LLMs still lack sufficient tool usage awareness to function as reliable agents.

Evidence
correlational
Key metric
Zero-shot F1: ChatGPT 77.81, ChatGLM2 48.88, Llama2-7b 65.03, Llama2-13b 11.53, Vicuna-7b 13.43, Vicuna-13b 42.2, Vicuna-33b 53.49, Koala-13b 54.04. Five-shot F1: ChatGPT 81.87, Llama2-13b 54.32 (Δ 42.79), Vicuna-7b 55.71 (Δ 42.28).
Caveat
The test set is a manually curated subset of 515 positive samples from TOOLE plus 515 negative samples from public instruction datasets; the authors note that the boundary between 'needs a tool' and 'does not' can be ambiguous.
Model
ChatGPT, ChatGLM2, Llama 2 / Llama 2 base, Vicuna, Koala
Concepts
Failure mode
Datasets
Self-Instruct [eval], CommonsenseQA [eval], LIMA [eval]
Related work
API-Bank [compared-to], ToolLLM [compared-to], GPT4Tools [compared-to]
Related findings
IC-987, IC-988, IC-989
Extraction
automatic-extraction