SFI-Bench: leaderboard

SFI-Bench (2026) tests the spatial-functional intelligence of multimodal models with human-verified four-option questions about egocentric videos of scanned indoor rooms, from counting, path and layout reasoning to object function, operation planning and troubleshooting. The score is the macro-average accuracy over the six tasks for models answering without a web search tool; percent, higher is better.

Metric: Macro-average accuracy (%) over the six tasks (global and conditional counting, cross-view path reasoning, layout inference, functional association, operation planning, troubleshooting), four-option multiple choice on egocentric indoor videos (ARKitScenes and ScanNet++ scans), zero-shot with the same prompt templates, models at their default configurations; the first four tasks are answered directly and, for the last two, models without a search tool also answer offline; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1LLaVA-Video-72B64.9#1193
2LLaVA-OneVision-72B61.3#1771
3Qwen 3 VL 235B A22B Instruct60.7#532
4Qwen 3 VL 32B Instruct59#606
5Qwen 3 VL Plus58.1#512
6Qwen 3 VL 235B A22B (Thinking)57.9#459
7Qwen 3 VL 32B (Thinking)55.9#573
8InternVL3.5-30B-A3B55.9#1267
9InternVL3.5-14B55.3#1504
10InternVL3.5-8B53.9#1604
11Qwen 3 VL 8B Instruct53.3#1029
12Qwen 3 VL 30B A3B Instruct52.7#895
13Qwen 3 VL 30B A3B (Thinking)52.1#719
14Qwen 3 VL 8B (Thinking)51.4#910
15LLaVA-Video-7B50.9#2276

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=sfi-bench · How It Works · Data refreshed daily, snapshot 2026-11-21.