IC-987When the correct tool is absent from the candidate list, most LLMs hallucinate a tool rather than returning 'none'

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 reliability sub-task, the ground-truth tool and its k most similar tools are removed from the candidate list, so the correct answer is 'none'. Under zero-shot prompting, the CSR of most LLMs is below 3%: Llama2-7b 0.90%, Vicuna-7b 1.50%, Koala-13b 1.70%, Llama2-13b 2.31%, Vicuna-13b 2.51%, Vicuna-33b 2.81%. Only ChatGPT (50.35%) and ChatGLM2 (6.63%) perform above this floor. Even with five-shot prompting, most models remain below 6% CSR. The authors attribute this to hallucination (fabricating non-existent tools) and sycophancy (avoiding a 'none' answer to please the user).

Evidence
correlational
Key metric
Zero-shot CSR (reliability): ChatGLM2 6.63, ChatGPT 50.35, Llama2-7b 0.90, Llama2-13b 2.31, Vicuna-7b 1.50, Vicuna-13b 2.51, Vicuna-33b 2.81, Koala-13b 1.70. Five-shot CSR: ChatGLM2 15.68, ChatGPT 78.49, Llama2-7b 2.51, Llama2-13b 5.93, Vicuna-7b 1.81, Vicuna-13b 3.42, Vicuna-33b 3.11, Koala-13b 5.83.
Caveat
The task removes the correct tool and its k most similar tools, making the remaining candidates deliberately dissimilar; real-world tool lists may contain closer distractors.
Model
ChatGPT, ChatGLM2, Llama 2 / Llama 2 base, Vicuna, Koala
Concepts
Failure mode
Related work
API-Bank [compared-to], ToolLLM [compared-to]
Related findings
IC-986, IC-988, IC-989
Extraction
automatic-extraction