Abstract

WhatMulti-category function-calling benchmark: the core Simple, Parallel, Multi-Turn, Live and Relevance-Detection categories, plus v4 agentic evaluation (web search, memory, format sensitivity).

WhoGorilla team, UC Berkeley (Yan, Patil, Chen, et al.). Current release: V4 (ICML 2025).

ScoresPer-model rankings live on the official leaderboard (linked); we do not reprint a score table.

Leaderboardgorilla.cs.berkeley.edu/leaderboard

Section II.viii Agent Benchmarks|Reviewed 2026

BFCL v4: The Berkeley Function-Calling Leaderboard, Explained

The benchmark every tool-using assistant team checks before shipping, now extended to agentic tool use.

Construction

BFCL evaluates function-calling: given a user query and a candidate set of function definitions (name, description, parameters, return type), the model must emit a structured tool call. The Berkeley Gorilla team curates test cases spread across categories that escalate in difficulty, and re-runs the board as new models appear.

v1 (Feb 2024) tested only Simple AST grading: one function in scope, exact-match on the emitted call. v2 (Aug 2024) added live function execution (the call actually runs against a real API), multi-step (the model must call several functions in sequence), and parallel (one turn, multiple calls). v3 (Dec 2024) added Relevance Detection (the model must refuse to call any tool when none is relevant) and Multi-Turn state tracking across long dialogues. v4, introduced at ICML 2025, adds a holistic agentic layer on top of those function-calling categories: agentic web search (multi-hop retrieval with error recovery), agent memory management, and format sensitivity (how much a model's answer shifts when the tool schema is re-formatted). The public board carried V4 as of its 12 April 2026 update.

Per-Model Scores

We do not reprint a per-model BFCL score table. A responsible cross-model table would need each cell taken from the official Berkeley leaderboard under a disclosed harness, and that board is JavaScript-rendered and lags the current frontier generation, while vendor-reported function-calling numbers come from harnesses we cannot inspect. Any table we copied here would be stale or unverifiable. For live rankings use the official BFCL leaderboard, and start from the homepage task picker to find the benchmark that matches your use case.

III

Reading BFCL Scores

Always inspect the per-category breakdown, not just the overall. A model that scores 85% overall by acing Simple and tanking Multi-Turn is a different proposition for production than one that scores 80% by being merely competent across categories. The overall number averages over varying difficulty; the category-level table is where you find which failure mode dominates.

Tool-use benchmarks compared →Tau-Bench retail and airline →OpenAI Agents SDK on BFCL →

Reader Questions

Q.01What is BFCL?+

BFCL (Berkeley Function-Calling Leaderboard) is the standard benchmark for evaluating function-calling, the LLM capability where the model decides which tool to invoke and with what arguments. Maintained by the Gorilla team at UC Berkeley, it has grown from v1 (simple AST-grading) through v2 (enterprise and live data), v3 (multi-turn and multi-step) to the current v4, introduced at ICML 2025, which adds holistic agentic evaluation. The board carried V4 as of its 12 April 2026 update.

Q.02What are the BFCL categories?+

The core function-calling categories carry through every version: Simple (one function, AST match), Multiple (pick one from a candidate set), Parallel (call several functions in one turn), Parallel-Multiple (combine the previous two), Multi-Turn (track tool state across turns), Live (real APIs invoked, not just AST graded), and Relevance Detection (refuse if no tool is relevant). V4 adds three agentic categories on top: web search (multi-hop retrieval with error recovery), memory management, and format sensitivity (how much the answer changes when the tool schema is re-formatted). The headline number is the unweighted accuracy across categories.

Q.03Why does function-calling need a dedicated benchmark?+

Function-calling failure modes do not show up in MMLU or HumanEval. A model can write correct Python but invent function signatures, hallucinate parameter names, fail to handle ambiguous user intent, or call a function when the user did not ask for a tool. BFCL surfaces these failure modes directly and is the only benchmark most function-calling teams trust as a regression signal.

Q.04Where can I find current per-model BFCL scores?+

On the official Berkeley leaderboard at gorilla.cs.berkeley.edu/leaderboard, which the Gorilla team re-runs as new models appear. We do not reprint a per-model BFCL score table here: the public board is JavaScript-rendered and lags the current frontier generation, and vendor-reported function-calling numbers are produced under harnesses we cannot inspect, so any table we copied would be stale or unverifiable. Read the live board, and when comparing models check the per-category breakdown rather than the single overall accuracy.

Q.05Is BFCL gameable?+

Partially. AST grading rewards exact match on function names and arguments, which means models trained on similar synthetic data (a common practice) score higher on Simple and Multiple categories than their underlying tool-use capability deserves. Live and Multi-Turn categories are more robust because they require correct tool execution and state tracking, not just lexical match.

Sources

[1] BFCL leaderboard: gorilla.cs.berkeley.edu/leaderboard
[2] BFCL v3 (multi-turn) announcement: gorilla.cs.berkeley.edu/blogs/13_bfcl_v3
[3] BFCL v4 agentic evaluation (web search / memory / format sensitivity): blogs/15_bfcl_v4_web_search, 16_bfcl_v4_memory, 17_bfcl_v4_prompt_variation
[4] Gorilla repository: github.com/ShishirPatil/gorilla