ModelRefs / EgoSchema Leaderboard
EgoSchema Leaderboard
Long-form first-person video question answering. Current leaders, methodology, and citation sources for EgoSchema.
Overview
Long-form first-person video question answering.
How it is measured: 5-way MC over 3-minute clips.
What this benchmark measures
- very long-form video understanding
- temporal reasoning over minutes of footage
Relevant to:
- video model screening
- long-form video comprehension
Failure modes it exercises:
- answering video questions from a single frame rather than the whole clip
Method and its limits
Over 5,000 human-curated multiple-choice questions, each over a three-minute video clip derived from Ego4D, answered by selecting one of five options, across more than 250 hours of real footage.
- Multiple choice permits elimination strategies that do not require understanding the full clip.
- Footage is egocentric Ego4D material, so third-person, broadcast and screen-recording video are not represented.
Dataset
- Dataset
- EgoSchema
- Type
- egocentric long-form video question answering
- Freshness
- aging
- Benchmark release
- 2023-08-17
Over 5,000 questions over three-minute clips spanning more than 250 hours, derived from Ego4D, as registered on the official page. The page states no release date; the freshness date recorded here is the primary paper's publication date. The primary paper establishes the remaining figures: EgoSchema is built on temporal certificate sets, a measure of the intrinsic temporal understanding length a task requires, on which it runs over 5.7x longer than the next closest video dataset; humans reach about 76% accuracy, while the models tested at publication scored under 33% against a 20% random baseline. The 76.2 registered on this page therefore sits at the human figure the paper reports.
How to read this score
Accuracy should be read against the five-option chance baseline of 20%, and against whether the model was given the full clip or sampled frames, which changes the task entirely.
Similarity to real tasks: Medium — Established from the official source: EgoSchema is over 5,000 human-curated multiple-choice questions, each over a three-minute egocentric clip drawn from Ego4D, across more than 250 hours of real video. Real first-person footage of extended activity resembles genuine video understanding. The five-option multiple-choice format is far narrower than describing or reasoning about video in open form.
Data contamination risk: Medium — Ego4D source footage is a public research corpus and the questions are published. No run-specific overlap audit is registered.
Benchmark gaming risk: Medium — A five-option format over a fixed question set is susceptible to prior-based guessing and to tuning on the released questions.
What you still need to test yourself
- Test your own video type and length; egocentric footage is a narrow slice.
- Fix the frame-sampling protocol before comparing any published figures.
This benchmark supports decisions about:
- Screen whether a video model reasons over minutes rather than single frames.
Limitations
- Five-option multiple choice over egocentric footage only.
- Figures are not comparable across different frame-sampling protocols.
Official project source reviewed 2026-10-12. The source states no release date; the freshness date recorded here is the publication date of the primary paper attached 2026-10-25.
What this evidence does not cover
- Only one model is currently scored on this page, so no comparison is supported.
Sources
- EgoSchema — official project page Mangalam, Akshulakov, Malik (UC Berkeley) · accessed 2026-10-12Supports the over-5,000 question count, the three-minute clip length, the 250-plus hours of footage, the five-option format and the Ego4D derivation; it reports no human or model accuracy and does not support any model score.
- EgoSchema: A Diagnostic Benchmark for Very Long-form Video Language Understanding Mangalam et al. · accessed 2026-10-25The primary methodology paper, establishing the benchmark's question construction over long-form video and its release date. It does not establish any model's score.
How this benchmark is scored
| Category | multimodal |
|---|---|
| Maximum score | 100 % accuracy |
| Direction | Higher is better |
Primary source: https://egoschema.github.io/
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to EgoSchema Leaderboard.
Frequently asked questions
What does the EgoSchema benchmark measure?
Long-form first-person video question answering.
What is the methodology?
5-way MC over 3-minute clips.
Which AI model is the current EgoSchema leader?
Qwen2.5-VL 72B (Alibaba) currently leads with 76.2 % accuracy.