Tool-use benchmark
Berkeley Function Calling Leaderboard
Selecting tools and supplying the right arguments, across increasingly involved interactions.
Key information
- Focus
- Function calling and multi-turn agentic tool use.
- Measure
- Category accuracy; V4 overall accuracy is an unweighted category average.
- Version / scope
- BFCL V4 · evaluator f7cf735
- Update pattern
- The publisher says periodic updates. The reviewed page still declares 12 April 2026; we do not label that as a recent evaluation.
- Publisher update
- 12 April 2026
- Editorial verification
- 15 October 2026
In plain English
What makes it different?
BFCL examines how well models choose and invoke functions, including more involved interactions in newer versions. It provides a focused view of tool use. It is not a complete test of whether an agent will deliver a useful outcome in every real workflow.
Sources and attribution are listed below.
What it measures
Function calling and multi-turn agentic tool use.
Strengths
- Makes tool-use failures easier to isolate.
- Publishes versioned methods and a reproducibility checkpoint.
Limitations
- An overall average can hide weak performance in one interaction type.
- Tool schemas, prompt formatting and the model's calling interface influence the score.
How the evaluation works
Different categories check function selection, arguments and interactions, using evaluation methods appropriate to the task.
The V4 page reports a category average and identifies a reproducible evaluator checkpoint. Its native function-calling and text-prompt routes are distinguished.
Configuration matters
Keep FC (native tool calling) distinct from Prompt mode, and compare the same BFCL generation, subcategories and evaluator. A row's model licence is not a licence for the table.
Primary methods: BFCL V4 methodology ↗
Results at the publisher
Scores stay with the original source
We have not documented the permissions needed to reproduce this leaderboard here. You can still explore its methods and visit the publisher for results.
Versions & freshness
Verified 15 October 2026The official V4 page declares 12 April 2026 and evaluator f7cf735 / bfcl-eval 2025.12.17. Our October review does not change those source dates.
Official V4 page, method links and repository accessible. The publisher says periodic updates. The reviewed page still declares 12 April 2026; we do not label that as a recent evaluation.
Our review date is not an evaluation date. Individual run dates are not consistently available; they remain unknown here. No automatic result collection is enabled.
Publication & reuse
The repository's software licence does not by itself settle hosted table, response dataset or submission-reproduction rights. A scoped data grant remains to be documented.
Gorilla's Apache 2.0 software licence is recorded separately from result-data permissions.
View the scope of the reuse decision
- Result fields
- No result fields cleared
- Public result display
- not enabled
- Storage
- not enabled
- Historical copies
- not enabled
- Updates
- not enabled
- Commercial reuse
- not enabled
- Derived ratings
- No scoped grant documented; not enabled.
- Attribution
- Credit the named publisher and link the primary sources. Attribution alone does not grant reproduction rights.
This decision does not clear mixed-source scientific recommendations. Publisher content, software, datasets and model licences are separate. No partnership is implied.
Primary sources & attribution
This guide is original scor.es commentary about Gorilla / Berkeley Function Calling Leaderboard research team, UC Berkeley's work. The links below support the identity, methods, versions and reuse review.
- Official leaderboard and reproducibility note ↗Checked 15 October 2026
- BFCL V4 methodology ↗Checked 15 October 2026
- Official evaluation repository ↗Checked 15 October 2026
- Gorilla repository licence ↗Checked 15 October 2026
Editorial edition 2026-10-15.1 · reuse review directory-rights-1. Re-check due 14 November 2026.