OVO-S-Bench: leaderboard
OVO-S-Bench (InternLM, 2026) measures streaming spatial intelligence with human-annotated multiple-choice questions over egocentric and driving videos, each answerable only from the video before its query time, across four levels from egocentric perception to allocentric mapping. This board is overall accuracy in percent, the mean of the four levels, with 128 frames sampled from the video prefix; higher is better.
Metric: Overall multiple-choice accuracy (%) on the 1,680-question evaluation set, prefix-access mode (128 frames sampled uniformly from the video before the query timestamp, question given with the frames), macro-average of the four levels. Source: arxiv.org. Saturation forecast: Around 2029. 31 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 59.19 | #94 |
| 2 | Qwen 3 VL 235B A22B | 53.64 | #479 |
| 3 | Qwen 3.5 397B A17B | 52.1 | #246 |
| 4 | Qwen 3.5 27B (Non-reasoning) | 51.69 | #359 (Qwen 3.5 27B) |
| 5 | GPT-5.4 | 50.89 | #123 |
| 6 | Gemini 3.1 Flash Lite | 50.81 | #364 |
| 7 | Qwen 3 VL 32B Instruct | 48.8 | #585 |
| 8 | Qwen 3 VL 4B Instruct | 46.83 | #1279 |
| 9 | Qwen 3.5 9B (Non-reasoning) | 45.91 | #803 (Qwen 3.5 9B) |
| 10 | Qwen 3.5 4B (Non-reasoning) | 45.4 | #1175 (Qwen 3.5 4B) |
| 11 | Qwen 2.5 VL 7B Instruct | 44.19 | #1833 |
| 12 | Grok 4.1 Fast (Non-reasoning) | 43.73 | #430 (Grok 4.1 Fast) |
| 13 | Gemma 4 26B A4B | 42.56 | #615 |
| 14 | Gemma 4 E4B | 39.69 | #2214 |
| 15 | Gemma 4 E2B | 36.07 | #3892 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=ovo-s-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.