SFI-Bench - Global and Conditional Counting: leaderboard

Counting questions (global counts and counts under a condition) from SFI-Bench (2026): four-option questions about egocentric videos of scanned indoor rooms, answered zero-shot from the video. Accuracy in percent; higher is better.

Metric: Accuracy (%) on the global and conditional counting questions, four-option multiple choice on egocentric indoor videos (ARKitScenes and ScanNet++ scans), zero-shot with the same prompt templates, models at their default configurations; answered directly with no tools; higher is better. Source: arxiv.org. Saturation forecast: Around June 2027. 24 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3.1 Pro (Preview)59.1#102
2GPT-558.4#185
3GPT-5.4High58.4#138
4LLaVA-Video-72B57.9#1211
5LLaVA-Video-7B55.4#2303
6Gemini 3.1 Flash Lite55#377
7Gemini 2.5 Pro54.4#280
8Qwen 3 VL 235B A22B (Thinking)53.8#462
9LLaVA-OneVision-72B52.8#1782
10Qwen 3 VL 235B A22B Instruct52.3#541
11Qwen 3 VL Plus51.3#514
12O4 Mini51#372
13Qwen 3 VL 32B Instruct50#607
14Qwen 3 VL 32B (Thinking)49.5#577
15InternVL3.5-30B-A3B48.5#1269

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

Interactive version: theaggregate.ai/benchmark?slug=sfi-bench-global-and-conditional-counting · How It Works · Data refreshed daily, snapshot 2026-11-22.