ModelRefs / BFCL v3 Leaderboard — AI Model Scores
BFCL v3 Leaderboard — AI Model Scores
Berkeley Function-Calling Leaderboard, parallel + multi-turn tool use. Current leaders, methodology, and citation sources for BFCL v3.
Overview
Berkeley Function-Calling Leaderboard, parallel + multi-turn tool use.
How it is measured: Aggregate accuracy across AST, executable, multi-turn splits.
What this benchmark measures
- function selection
- argument construction
- multi-turn and multi-step tool use
Relevant to:
- tool-calling evaluation
- agent workflow screening
Failure modes it exercises:
- wrong function
- invalid arguments
- state-transition failure
- irrelevant tool use
Method and its limits
Generate and execute function calls across single-turn, parallel, multiple, and multi-turn/multi-step categories, then grade structure, execution, response, and state outcomes according to the pinned BFCL v3 harness.
- Category mix, handler implementation, model API behavior, retries, and harness revision affect results.
- Controlled functions do not cover a deployment's permissions, irreversible actions, private data, or business-specific error costs.
Dataset
- Dataset
- Berkeley Function Calling Leaderboard v3
- Type
- executable function-calling and multi-turn tool-use tasks
- Freshness
- unknown
BFCL v3 adds multi-turn and multi-step evaluation to prior single-turn categories; the changelog records corrections and metric changes that make revision pinning essential.
How to read this score
Inspect category-level failures and use the exact harness revision; aggregate accuracy does not establish safe end-to-end agency.
Similarity to real tasks: Medium — Executable functions and multi-turn state improve task realism, but the benchmark still uses controlled schemas and environments.
Data contamination risk: Unknown — Public tasks and schemas exist; no model-specific training-overlap assessment is registered.
Benchmark gaming risk: Unknown — Model-specific handlers, prompt formatting, retries, and benchmark-targeted tuning can influence results.
What you still need to test yourself
- Test the deployment's real tools, permissions, schemas, data, retries, side effects, and human approvals.
- Measure task success, unsafe-action prevention, recovery, latency, and cost under representative failures.
This benchmark supports decisions about:
- Compare tool-calling systems under the same BFCL v3 categories, handlers, and revision.
- Identify routing, argument, execution, and multi-turn failure modes for deeper testing.
Limitations
- BFCL v3 is a tool-calling benchmark, not proof of safe autonomous workflow completion.
- Leaderboard comparisons can drift as APIs, handlers, ground truth, and metrics change.
Dataset freshness is not fully confirmed from available source coverage. Pin the repository commit, data revision, categories, handlers, and grading protocol.
Sources
- Berkeley Function Calling Leaderboard UC Berkeley Gorilla project · accessed 2026-06-29
- BFCL v3: Multi-Turn and Multi-Step Function Calling Evaluation UC Berkeley Gorilla project · accessed 2026-06-29
- BFCL changelog UC Berkeley Gorilla project · accessed 2026-06-29
How this benchmark is scored
| Category | agents |
|---|---|
| Maximum score | 100 % accuracy |
| Direction | Higher is better |
Primary source: https://gorilla.cs.berkeley.edu/leaderboard.html
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to BFCL v3 Leaderboard — AI Model Scores.