MMLU calibration (expected calibration error)
Models answer an MMLU subset and their answer probabilities are compared with accuracy. Scored as expected calibration error.
Reported in the GPT-4 Technical Report for the pre-trained and post-trained models.
- Made by
- OpenAI
- Released
- March 2023
- Kind
- Short factual questions
- Level
- University
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- Kept private
- Strongest model tested
- GPT-4 Mar 2023 #180 on the capability index
- Last in a system card
- March 2023
- Last model added
- March 2023
GPT-4 leads with 0.07
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
- 1 GPT-4 Mar 2023 0.07
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-4 Mar 2023 | 0.07 | Mar 2023 | Lab model card |
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Labs’ model cards
Labs report it only in system cards
We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark. Some system cards do, below.
One system card reports it
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Mar 14, 2023 | GPT-4 technical report | 0.074 | Calibration curve on an MMLU subset, post-trained (ppo) model ECE 0.074, pre-trained ECE 0.007 (Figure 8) | Hallucination |
More in Factuality
- SimpleVQA In use Models answer 2,025 short factual questions about images across 9 tasks and 9 topics. An LLM judge grades answers against fixed references.
Qwen3.7-Plus 81.7% - HallusionBench In use Models answer 1,129 yes or no questions about 346 images, built in pairs to expose language hallucination and visual illusion. Scored by accuracy.
Keye-VL-2.0-30B-A3B 73.5% - SimpleQA Verified In use 1,000 human-verified, de-duplicated short-answer factual questions measuring parametric knowledge.
Gemini 3.1 Pro 77.5% - Kimi Hallu Bench In use Models answer hard prompts drawn from practical work tasks. Fact-checking agents score the share of claims that are hallucinated. GPT-6 Astra 4.20%
- Internal adversarially-designed factuality evaluations In use Models answer adversarial questions in nine categories such as history, math and code. Scored on agreement with human ideal responses. GPT-6 Luna 7.60%
- AA-Omniscience v1.1 In use Models answer 3,000 knowledge questions in six domains without tools and may abstain. The index adds correct answers and subtracts wrong ones.
Inkling 2 Small 16.1