The AI arena is free today

Open Superagent
AI OrganizationMaintains 26 benchmarks

UC Berkeley: AI benchmarks & research

University of California, Berkeley. Home of the Gorilla / Berkeley Function Calling Leaderboard, LiveCodeBench, CyberGym, ExploitGym, EgoSchema, and the Hendrycks MMLU and MATH datasets.

Maintained by UC Berkeley

26 benchmarks
  • AMC_2022_23

    American Mathematics Competition problems from the 2022-23 academic year, consisting of multiple-choice mathematics competition problems designed for high school students. These problems require advanced mathematical reasoning, problem-solving strategies, and mathematical knowledge covering topics like algebra, geometry, number theory, and combinatorics. The benchmark is derived from the official AMC competitions sponsored by the Mathematical Association of America.

    mathreasoningCounts toward the LLM Stats Score
    Leader
    Mistral Large 3 (675B Base)52.0%
    Models scored
    6
  • BFCL

    The Berkeley Function Calling Leaderboard (BFCL) is the first comprehensive and executable function call evaluation dedicated to assessing Large Language Models' ability to invoke functions. It evaluates serial and parallel function calls across multiple programming languages (Python, Java, JavaScript, REST API) using a novel Abstract Syntax Tree (AST) evaluation method. The benchmark consists of over 2,000 question-function-answer pairs covering diverse application domains and complex use cases including multiple function calls, parallel function calls, and multi-turn interactions.

    reasoninggeneraltool callingCounts toward the LLM Stats Score
    Leader
    Llama 3.1 405B Instruct88.5%
    Models scored
    11
  • BFCL v2

    Berkeley Function Calling Leaderboard (BFCL) v2 is a comprehensive benchmark for evaluating large language models' function calling capabilities. It features 2,251 question-function-answer pairs with enterprise and OSS-contributed functions, addressing data contamination and bias through live, user-contributed scenarios. The benchmark evaluates AST accuracy, executable accuracy, irrelevance detection, and relevance detection across multiple programming languages (Python, Java, JavaScript) and includes complex real-world function calling scenarios with multi-lingual prompts.

    reasoninggeneraltool callingCounts toward the LLM Stats Score
    Leader
    Llama 3.3 70B Instruct77.3%
    Models scored
    5
  • BFCL-v3

    Berkeley Function Calling Leaderboard v3 (BFCL-v3) is an advanced benchmark that evaluates large language models' function calling capabilities through multi-turn and multi-step interactions. It introduces extended conversational exchanges where models must retain contextual information across turns and execute multiple internal function calls for complex user requests. The benchmark includes 1000 test cases across domains like vehicle control, trading bots, travel booking, and file system management, using state-based evaluation to verify both system state changes and execution path correctness.

    reasoningstructured outputfinancegeneralagentstool callingCounts toward the LLM Stats Score
    Leader
    GLM-4.577.8%
    Models scored
    20
  • BFCL_v3_MultiTurn

    Berkeley Function Calling Leaderboard (BFCL) V3 MultiTurn benchmark that evaluates large language models' ability to handle multi-turn and multi-step function calling scenarios. The benchmark introduces complex interactions requiring models to manage sequential function calls, handle conversational context across multiple turns, and make dynamic decisions about when and how to use available functions. BFCL V3 uses state-based evaluation by verifying the actual state of API systems after function execution, providing more realistic assessment of function calling capabilities in agentic applications.

    reasoninggeneraltool callingCounts toward the LLM Stats Score
    Leader
    MiniMax M2.576.8%
    Models scored
    2
  • BFCL-V4

    Berkeley Function Calling Leaderboard V4 (BFCL-V4) evaluates LLMs on their ability to accurately call functions and APIs, including simple, multiple, parallel, and nested function calls across diverse programming scenarios.

    agentstool callingCounts toward the LLM Stats Score
    Leader
    Qwen3.7 Max75.0%
    Models scored
    24
  • CyberGym

    CyberGym is a benchmark for evaluating AI agents on cybersecurity tasks, testing their ability to identify vulnerabilities, perform security analysis, and complete security-related challenges in a controlled environment.

    safetyagentscodeCounts toward the LLM Stats Score
    Leader
    MiMo-V2.6-Flash95.1%
    Models scored
    21
  • EgoSchema

    A diagnostic benchmark for very long-form video language understanding consisting of over 5000 human curated multiple choice questions based on 3-minute video clips from Ego4D, covering a broad range of natural human activities and behaviors

    long contextreasoningvisionCounts toward the LLM Stats Score
    Leader
    Qwen2-VL-72B-Instruct77.9%
    Models scored
    9
  • ExploitGym

    ExploitGym is a large-scale, realistic benchmark built from real-world vulnerabilities across userspace programs, Google's V8 engine, and the Linux kernel. Given a vulnerability and a proof-of-vulnerability input, agents must craft a working end-to-end exploit that achieves unauthorized code execution.

    safetyagentscodeCounts toward the LLM Stats Score
    Leader
    Gemini 4 Carbon53.5%
    Models scored
    11
  • Gorilla Benchmark API Bench

    APIBench, a comprehensive dataset of over 11,000 instruction-API pairs from HuggingFace, TorchHub, and TensorHub APIs for evaluating language models' ability to generate accurate API calls.

    reasoningcodetool callingCounts toward the LLM Stats Score
    Leader
    Llama 3.1 405B Instruct35.3%
    Models scored
    3
  • LiveCodeBench

    LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code. It continuously collects new problems from programming contests (LeetCode, AtCoder, CodeForces) and evaluates four different scenarios: code generation, self-repair, code execution, and test output prediction. Problems are annotated with release dates to enable evaluation on unseen problems released after a model's training cutoff.

    reasoninggeneralcodeCounts toward the LLM Stats Score
    Leader
    DeepSeek-V4-Pro-Max93.5%
    Models scored
    76
  • LiveCodeBench(01-09)

    LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code. It continuously collects new problems from programming contests (LeetCode, AtCoder, CodeForces) and evaluates four different scenarios: code generation, self-repair, code execution, and test output prediction. Problems are annotated with release dates to enable evaluation on unseen problems released after a model's training cutoff.

    reasoninggeneralCounts toward the LLM Stats Score
    Leader
    DeepSeek-V2.541.8%
    Models scored
    1
  • LiveCodeBench Pro

    LiveCodeBench Pro is an advanced evaluation benchmark for large language models for code that uses Elo ratings to rank models based on their performance on coding tasks. It evaluates models on real-world coding problems from programming contests (LeetCode, AtCoder, CodeForces) and provides a relative ranking system where higher Elo scores indicate superior performance.

    reasoninggeneralcodeCounts toward the LLM Stats Score
    Leader
    Gemini 3.1 Pro2887
    Models scored
    5
  • LiveCodeBench v5

    LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code. It continuously collects new problems from programming contests (LeetCode, AtCoder, CodeForces) and evaluates four different scenarios: code generation, self-repair, code execution, and test output prediction. Problems are annotated with release dates to enable evaluation on unseen problems released after a model's training cutoff.

    reasoninggeneralCounts toward the LLM Stats Score
    Leader
    Gemini 2.5 Pro75.6%
    Models scored
    9
  • LiveCodeBench v5 24.12-25.2

    LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code. It continuously collects new problems from programming contests (LeetCode, AtCoder, CodeForces) and evaluates four different scenarios: code generation, self-repair, code execution, and test output prediction. Problems are annotated with release dates to enable evaluation on unseen problems released after a model's training cutoff.

    reasoninggeneralCounts toward the LLM Stats Score
    Leader
    Kimi-k1.562.5%
    Models scored
    1
  • LiveCodeBench v6

    LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code. It continuously collects new problems from programming contests (LeetCode, AtCoder, CodeForces) and evaluates four different scenarios: code generation, self-repair, code execution, and test output prediction. Problems are annotated with release dates to enable evaluation on unseen problems released after a model's training cutoff.

    reasoninggeneralCounts toward the LLM Stats Score
    Leader
    Qwen3.8 Flash91.9%
    Models scored
    69
  • MATH

    MATH dataset contains 12,500 challenging competition mathematics problems from AMC 10, AMC 12, AIME, and other mathematics competitions. Each problem includes full step-by-step solutions and spans multiple difficulty levels (1-5) across seven mathematical subjects including Prealgebra, Algebra, Number Theory, Counting and Probability, Geometry, Intermediate Algebra, and Precalculus.

    mathreasoningCounts toward the LLM Stats Score
    Leader
    o3-mini97.9%
    Models scored
    71
  • MATH-500

    MATH-500 is a subset of the MATH dataset containing 500 challenging competition mathematics problems from AMC 10, AMC 12, AIME, and other mathematics competitions. Each problem includes full step-by-step solutions and spans multiple difficulty levels across seven mathematical subjects including Prealgebra, Algebra, Number Theory, Counting and Probability, Geometry, Intermediate Algebra, and Precalculus.

    mathreasoningCounts toward the LLM Stats Score
    Leader
    LongCat-Flash-Thinking99.2%
    Models scored
    33
  • MATH (CoT)

    MATH dataset contains 12,500 challenging competition mathematics problems from AMC 10, AMC 12, AIME, and other mathematics competitions. Each problem includes full step-by-step solutions and spans multiple difficulty levels (1-5) across seven mathematical subjects. This variant uses Chain-of-Thought prompting to encourage step-by-step reasoning.

    mathreasoningCounts toward the LLM Stats Score
    Leader
    Llama 3.1 70B Instruct68.0%
    Models scored
    6
  • MMLU

    Massive Multitask Language Understanding benchmark testing knowledge across 57 diverse subjects including STEM, humanities, social sciences, and professional domains

    languagelegalmathreasoningfinancegeneralhealthcareCounts toward the LLM Stats Score
    Leader
    GPT-592.5%
    Models scored
    104
  • MMLU-Base

    Base version of the Massive Multitask Language Understanding benchmark, evaluating language models across 57 tasks including elementary mathematics, US history, computer science, law, and other professional and academic subjects. Designed to comprehensively measure the breadth and depth of a model's academic and professional understanding.

    languagelegalmathreasoningfinancegeneralhealthcareCounts toward the LLM Stats Score
    Leader
    Qwen2.5-Coder 7B Instruct68.0%
    Models scored
    1
  • MMLU Chat

    Chat-format variant of the Massive Multitask Language Understanding benchmark, evaluating language models across 57 tasks including elementary mathematics, US history, computer science, law, and other professional and academic subjects. This version uses conversational prompting format for model evaluation.

    languagelegalmathreasoningfinancegeneralhealthcareCounts toward the LLM Stats Score
    Leader
    Llama 3.1 Nemotron 70B Instruct80.6%
    Models scored
    1
  • MMLU (CoT)

    Chain-of-Thought variant of the Massive Multitask Language Understanding benchmark, evaluating language models across 57 tasks including elementary mathematics, US history, computer science, law, and other professional and academic subjects. This version uses chain-of-thought prompting to elicit step-by-step reasoning.

    languagelegalmathreasoningfinancegeneralhealthcareCounts toward the LLM Stats Score
    Leader
    Llama 3.1 405B Instruct88.6%
    Models scored
    3
  • MMLU French

    French language variant of the Massive Multitask Language Understanding benchmark, evaluating language models across 57 tasks including elementary mathematics, US history, computer science, law, and other professional and academic subjects. This multilingual version tests model performance in French.

    languagelegalmathreasoningfinancegeneralhealthcareCounts toward the LLM Stats Score
    Leader
    Mistral Large 282.8%
    Models scored
    1
  • MMLU-STEM

    STEM-focused subset of the Massive Multitask Language Understanding benchmark, evaluating language models on science, technology, engineering, and mathematics topics including physics, chemistry, mathematics, and other technical subjects.

    mathphysicsreasoningchemistryCounts toward the LLM Stats Score
    Leader
    Qwen2.5 32B Instruct80.9%
    Models scored
    2
  • OpenAI MMLU

    MMLU (Massive Multitask Language Understanding) is a comprehensive benchmark that measures a text model's multitask accuracy across 57 diverse academic and professional subjects. The test covers elementary mathematics, US history, computer science, law, morality, business ethics, clinical knowledge, and many other domains spanning STEM, humanities, social sciences, and professional fields. To attain high accuracy, models must possess extensive world knowledge and problem-solving ability.

    legalmathphysicspsychologyreasoningfinancegeneralhealthcarechemistryeconomicsCounts toward the LLM Stats Score
    Leader
    Gemma 3n E4B Instructed35.6%
    Models scored
    2

In one line

UC Berkeley maintains AMC_2022_23 (counts toward the LLM Stats Score, 6 models scored), BFCL (counts toward the LLM Stats Score, 11 models scored), BFCL v2 (counts toward the LLM Stats Score, 5 models scored), BFCL-v3 (counts toward the LLM Stats Score, 20 models scored), BFCL_v3_MultiTurn (counts toward the LLM Stats Score, 2 models scored), BFCL-V4 (counts toward the LLM Stats Score, 24 models scored), CyberGym (counts toward the LLM Stats Score, 21 models scored), EgoSchema (counts toward the LLM Stats Score, 9 models scored), ExploitGym (counts toward the LLM Stats Score, 11 models scored), Gorilla Benchmark API Bench (counts toward the LLM Stats Score, 3 models scored), LiveCodeBench (counts toward the LLM Stats Score, 76 models scored), LiveCodeBench(01-09) (counts toward the LLM Stats Score, 1 model scored), LiveCodeBench Pro (counts toward the LLM Stats Score, 5 models scored), LiveCodeBench v5 (counts toward the LLM Stats Score, 9 models scored), LiveCodeBench v5 24.12-25.2 (counts toward the LLM Stats Score, 1 model scored), LiveCodeBench v6 (counts toward the LLM Stats Score, 69 models scored), MATH (counts toward the LLM Stats Score, 71 models scored), MATH-500 (counts toward the LLM Stats Score, 33 models scored), MATH (CoT) (counts toward the LLM Stats Score, 6 models scored), MMLU (counts toward the LLM Stats Score, 104 models scored), MMLU-Base (counts toward the LLM Stats Score, 1 model scored), MMLU Chat (counts toward the LLM Stats Score, 1 model scored), MMLU (CoT) (counts toward the LLM Stats Score, 3 models scored), MMLU French (counts toward the LLM Stats Score, 1 model scored), MMLU-STEM (counts toward the LLM Stats Score, 2 models scored) and OpenAI MMLU (counts toward the LLM Stats Score, 2 models scored). It ships no models of its own on LLM Stats.

FAQ

Common questions about UC Berkeley.

What is UC Berkeley?

University of California, Berkeley. Home of the Gorilla / Berkeley Function Calling Leaderboard, LiveCodeBench, CyberGym, ExploitGym, EgoSchema, and the Hendrycks MMLU and MATH datasets. On LLM Stats it maintains 26 benchmarks, led by MMLU, LiveCodeBench and MATH. It ships no models of its own; each score on its benchmarks carries its own provenance label on the model page.

Which UC Berkeley benchmarks count toward the LLM Stats Score?

MMLU, LiveCodeBench, MATH, LiveCodeBench v6, MATH-500, BFCL-V4 and 20 more count toward the LLM Stats Score. Only benchmarks admitted to the core set move the main leaderboard; the rest are shown for transparency.

Which models lead UC Berkeley's benchmarks?

GPT-5 leads MMLU at 92.5% (104 models scored), DeepSeek-V4-Pro-Max leads LiveCodeBench at 93.5% (76 models scored) and o3-mini leads MATH at 97.9% (71 models scored). Each benchmark page lists the full ranking and how the results were produced.

Where are UC Berkeley's official benchmark results?

UC Berkeley's official site is https://www.berkeley.edu. Each benchmark page on LLM Stats links to that benchmark's own paper or leaderboard, and shows the scores next to every other benchmark so models can be compared on one page.