EXPLORE-Bench - Short Sequences: leaderboard
Metric: Unified score (0-100): 0.25 x object coverage + 0.35 x attribute score + 0.40 x relation score, the 0-5 scores rescaled to 0-100; attributes and relations are rated against human-corrected scene annotations by a Qwen3-8B scorer, over the 599 instances with 11 to 99 atomic actions; the model sees an initial egocentric scene image and a sequence of atomic actions (11 to 694) from real Ego4D, Ego-Exo4D and self-recorded videos and describes the final scene; greedy decoding; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B (Thinking) | 52.08 | |
| 2 | Qwen 3 VL 8B Instruct | 51.23 | #401 |
| 3 | Gemini 3 Flash | 50.27 | #93 |
| 4 | Gemini 3 Pro | 50.11 | #77 |
| 5 | Step3 VL 10B | 49.87 | #465 |
| 6 | GPT-5.2 Instant | 48.71 | #205 |
| 7 | Qwen 2.5 VL 7B Instruct | 43.58 | #643 |
| 8 | Qwen 2 VL 7B Instruct | 39.14 | #816 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=explore-bench-short-sequences · How It Works · Data refreshed daily, snapshot 2026-10-11.