SiT-Bench: leaderboard
SiT-Bench (Beijing Institute of Technology, 2026) tests spatial reasoning from text alone: 3,892 expert-reviewed items describe single- and multi-view scenes as coordinate-aware captions, across 17 subtasks from navigation and perspective shifts to robotic action prediction. Items are mostly multiple choice (cognitive mapping asks for a JSON map), scored as item-weighted accuracy in percent, higher is better; models with a switchable thinking mode appear in their standard run, and the paper's separate thinking reruns are not included.
Metric: Accuracy (%; item-weighted over the 17 subtasks). Source: arxiv.org. Saturation forecast: Around August 2028. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 59.46 | #141 |
| 2 | Qwen 3 VL 32B Instruct | 45.9 | #582 |
| 3 | GPT-4o | 45.7 | #741 |
| 4 | Qwen 2.5 VL 72B Instruct | 45.45 | #845 |
| 5 | InternVL3-14B | 42.58 | #1184 |
| 6 | Qwen 2.5 72B Instruct | 42.57 | #1093 |
| 7 | Qwen 3 VL 8B Instruct | 42.1 | #982 |
| 8 | Qwen 3 VL 4B Instruct | 38.67 | #1281 |
| 9 | InternVL3-8B | 38.42 | #1695 |
| 10 | Qwen 3 8B (Non-reasoning) | 37.91 | #1838 (Qwen 3 8B) |
| 11 | DeepSeek V3.2 (Non-reasoning) | 37.06 | #403 (DeepSeek V3.2) |
| 12 | Qwen 3 4B 2507 Instruct | 36.59 | #2174 |
| 13 | Qwen 3 30B A3B 2507 Instruct | 36.5 | #1078 |
| 14 | Qwen 2.5 7B Instruct | 36.43 | #2583 |
| 15 | Qwen 2.5 3B Instruct | 34.81 | #5206 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sit-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.