Agents & Tool Use

BFCL

BFCL tests whether models can answer a user request by calling the right function with the right arguments, including requests that no available function can satisfy and ones spread over several conversational turns.

4,133items

67subjects

Apache-2.0license

agents_and_tool_usedomain

textmodality

item-level responses released

Saturation status: No

Full data on Hugging Face Back to the gallery

Paper and credit

2025

The Berkeley Function Calling Leaderboard (BFCL): From Tool Use to Agentic Evaluation of Large Language Models (opens in a new tab)

Original source

Authors

Shishir G Patil, Huanzhi Mao, Fanjia Yan, Charlie Cheng-Jie Ji +3 moreShow all

01Shishir G Patil
02Huanzhi Mao
03Fanjia Yan
04Charlie Cheng-Jie Ji
05Vishnu Suresh
06Ion Stoica
07Joseph E. Gonzalez

Citation

BibTeX citationShow

BibTeX

1@InProceedings{pmlr-v267-patil25a,
2  title={The Berkeley Function Calling Leaderboard ({BFCL}): From Tool Use to Agentic Evaluation of Large Language Models},
3  author={Patil, Shishir G and Mao, Huanzhi and Yan, Fanjia and Ji, Charlie Cheng-Jie and Suresh, Vishnu and Stoica, Ion and Gonzalez, Joseph E.},
4  booktitle={Proceedings of the 42nd International Conference on Machine Learning},
5  pages={48371--48392},
6  year={2025},
7  editor={Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry},
8  volume={267},
9  series={Proceedings of Machine Learning Research},
10  month={13--19 Jul},
11  publisher={PMLR},
12  pdf={https://raw.githubusercontent.com/mlresearch/v267/main/assets/patil25a/patil25a.pdf},
13  url={https://proceedings.mlr.press/v267/patil25a.html},
14}

Response matrix

Rasch analysis p = σ(θ − z)

441,086 responses, 80/20 split over cells · 93 subjects · 4,133 items

AUC train

0.908

AUC test

0.902

Loading session strips…

Correct (1)Incorrect (0)Unobserved

Scale: 1 = correct · 0 = incorrect