EgoExo-Next: leaderboard

EgoExo-Next (2026) tests whether vision-language models can reason about how an observed action trajectory will look next: 2,503 human-curated four-choice questions from six public egocentric and ego-exo video sources, where each option is a candidate visual state (an image) rather than a text label. Four subtasks cover egocentric next-state prediction, ego-to-exo and exo-to-ego state correspondence, exocentric next-state prediction and their composition (Ego-to-Exo Next-State), each split into indoor and outdoor scenes. The board is the paper's Avg., the unweighted mean accuracy over the eight subtask-by-scene settings, zero-shot, mean of three runs; chance is 25% and human accuracy is 98.55%.

Metric: Accuracy (%), unweighted mean over the eight subtask-by-scene settings (four-choice visual-option questions, chance 25%), zero-shot, mean of three runs. Source: arxiv.org. Saturation forecast: Around 2029. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3.1 Pro (Preview)43.81#102
2GPT-5.243.16#190
3Qwen 3 VL 235B A22B40.93#486
4Qwen 3 VL 32B37.03#624
5InternVL3.5-38B36.81#1026
6Qwen 3 VL 8B36.27#927
7Claude Opus 4.633.59#114
8InternVL3.5-8B32.93#1618
9LLaVA-OneVision-1.5-8B31.85#1876

No result here: #2 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=egoexo-next · How It Works · Data refreshed daily, snapshot 2026-11-22.