IC-988LLMs show large gaps in multi-tool selection and over-rely on the number of tools specified in the prompt

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

In the multi-tool selection task, models must choose the correct pair of tools from a candidate list. ChatGPT achieves 88.28% 2/2 CSR while ChatGLM2 manages only 20.20%, a gap of nearly 70 points. A common error is omitting a tool: Vicuna-33b selects only one correct tool in 48.49% of cases under the multi-choice prompt. When the prompt is changed to force exactly two selections (one-choice), Vicuna-33b's CSR jumps to 91.15% and Vicuna-7b improves by over 20 points, indicating that these models rely on the explicitly stated count rather than reasoning about which tools are actually needed.

Evidence
correlational
Key metric
2/2 CSR (multi-choice): ChatGPT 88.28, ChatGLM2 20.20, Llama2-7b 35.69, Llama2-13b 81.49, Vicuna-7b 44.06, Vicuna-13b 83.70, Vicuna-33b 48.69, Koala-13b 39.03. One-choice CSR: ChatGPT 88.53, ChatGLM2 23.34, Llama2-7b 57.34, Llama2-13b 77.87, Vicuna-7b 64.34, Vicuna-13b 78.47, Vicuna-33b 91.15, Koala-13b 25.10. Vicuna-33b 1/1 CSR: 48.49%.
Caveat
The multi-tool test set is small (497 queries from TOOLE), and the authors note that the 15 most similar tools to the ground-truth pair are excluded from the candidate list to avoid excessive difficulty.
Model
ChatGPT, ChatGLM2, Llama 2 / Llama 2 base, Vicuna, Koala
Concepts
Failure mode
Related work
ToolLLM [compared-to]
Related findings
IC-986, IC-987, IC-989
Extraction
automatic-extraction