MMLU-Pro: Advanced Multitask Knowledge and Reasoning Evaluation

An advanced benchmark that tests both broad knowledge and reasoning capabilities across many subjects, featuring challenging questions and multiple-choice answers with increased difficulty and complexity.

Overview

MMLU-Pro is a more robust and challenging massive multi-task understanding dataset tailored to more rigorously benchmark large language models’ capabilities. This dataset contains 12K complex questions across 14 disciplines (including ‘Other’).

There are three major differences compared to original MMLU:

1. MMLU-Pro increases the number of options from 4 to 10, making the evaluation more realistic and challenging.
2. In this dataset, the creators increase the problem difficulty and integrate more reasoning-focused problems.
3. The benchmark is made more robust by increasing the number of distractor options.

Usage

Installation

Install with pip install inspect-evals, or uv sync from a checkout of this repository.

Running evaluations

uv run inspect eval inspect_evals/mmlu_pro --model openai/gpt-5-nano

You can also import tasks as normal Python objects and run them from python:

from inspect_ai import eval
from inspect_evals.mmlu_pro import mmlu_pro
eval(mmlu_pro)

Drop uv run if you manage dependencies yourself. Log viewing (inspect view) and default-model setup are documented in the Inspect Evals README.

Options

You can control a variety of options from the command line. For example:

uv run inspect eval inspect_evals/mmlu_pro --limit 10 --sample-shuffle
uv run inspect eval inspect_evals/mmlu_pro --max-connections 10
uv run inspect eval inspect_evals/mmlu_pro --temperature 0.5

See uv run inspect eval --help for all available options.

Parameters

mmlu_pro

• subjects (str | list[str]): List of subjects to evaluate on. (default: [])
• fewshot (int): Number of few shot examples to use. (default: 0)
• shuffle (bool): Whether we shuffle the dataset for evaluation (default: True)

Dataset

Here is an example from the dataset:

Question: Approximately how far away is the Andromeda Galaxy?

Options:
A) 5 million light years
B) 2.5 million light years  
C) 2.1 million light years
D) 1.9 million light years
E) 3.2 million light years
F) 4 million light years
G) 1.7 million light years
H) 3.5 million light years
I) 1.2 million light years
J) 2.8 million light years

The model is tasked to answer the question and choose the appropriate option.

Scoring

The prompts are based on EleutherAI’s lm-evaluation-harness and MultipleChoiceTemplate.SINGLE_ANSWER. The in-built choice scorer is used for evaluation. A simple accuracy is calculated over the datapoints.

Changelog

[2-A] - 2026-02-16

• Migrate version to new scheme. See #907.

[1.0.1] - 2025-12-18

• Adds backoff policy for functions that connect to huggingface servers.