Knowledge Bases and Evaluation Organizations: Berkeley BFCL
Knowledge Bases and Evaluation Organizations · Berkeley BFCL. Dedicated leaderboard for tool use and function calling. This source has 7 entries admitted under this category's review policy; each links directly to the original.
Authority tiers in this category: A2–B. Admission policy: All core resources are included. Each secondary source is admitted according to its own database, methodology, versions and leaderboard structure; there is no cross-organization scoring scheme.
Berkeley Function Calling Leaderboard V4
Evaluates function and tool calling, parallel and multi-turn interaction, retrieval, memory and format sensitivity, and publishes cost and latency.
Evidence use: Supports comparisons of tool-use capability under a specified BFCL version and submission conditions.
Limitations: The overall score is a weighted combination of several tasks. The live API category changes with external services.
Related concepts: AI Agent · Model Evaluation and Benchmarks
Link last checked: 2026-09-24
Berkeley Function-Calling Leaderboard Methodology
Introduces function selection, parallel calls, relevance detection, multilingual functions and AST- and execution-based scoring.
Evidence use: Supports explaining BFCL's task composition, metrics and evaluation logic.
Limitations: The leaderboard composition described in earlier articles has been extended in newer versions. Read together with the current version's changelog.
Related concepts: Model Evaluation and Benchmarks
Link last checked: 2026-09-24
BFCL V4：Agentic Evaluation
Extends BFCL from function calling to agentic capabilities such as web search, memory and format sensitivity, and discloses how the new score is composed.
Evidence use: Supports describing BFCL V4's agent tasks, scale and weights.
Limitations: The article focuses on the web search part. Full conclusions require reading the leaderboard and other V4 documents together.
Related concepts: AI Agent · Model Evaluation and Benchmarks
Link last checked: 2026-09-24
BFCL V2：Live and Irrelevance Detection
Adds live functions, irrelevant-function detection and a design that mitigates data contamination.
Evidence use: Supports explaining BFCL's expansion from static function calling to live evaluation.
Limitations: Each generation has different task composition and weights. Live APIs and tool environments make results time-sensitive.
Related concepts: Model Evaluation and Benchmarks
Link last checked: 2026-09-24
BFCL V3：Multi-Turn and Multi-Step
Adds multi-turn, multi-step and state-dependent tool calling tasks.
Evidence use: Supports evaluating agents' tool planning and execution over continuous interaction.
Limitations: Each generation has different task composition and weights. Live APIs and tool environments make results time-sensitive.
Related concepts: AI Agent
Link last checked: 2026-09-24
BFCL V4：Memory Evaluation
Evaluates agents' memory: retrieving, updating and using information across turns.
Evidence use: Supports explaining BFCL V4's memory tasks and scoring.
Limitations: Each generation has different task composition and weights. Live APIs and tool environments make results time-sensitive.
Related concepts: AI Agent
Link last checked: 2026-09-24
BFCL V4：Prompt and Format Sensitivity
Measures how changes to tool definitions, prompts and output formats affect function calling performance.
Evidence use: Supports judging how robust tool-use capability is to formatting.
Limitations: Each generation has different task composition and weights. Live APIs and tool environments make results time-sensitive.
Related concepts: Model Evaluation and Benchmarks
Link last checked: 2026-09-24