SiT-Bench - Logic & Anomaly Detection: leaderboard
SiT-Bench category (Beijing Institute of Technology, 2026): consistency of the internal world model, with object permanence across disjoint viewpoints and cardinal direction judgement (450 items). Item-weighted accuracy in percent, higher is better; scenes are given only as coordinate-aware text descriptions, never as images.
Metric: Accuracy (%; item-weighted over the category's subtasks). Source: arxiv.org. Saturation forecast: Around November 2027. 31 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 59.11 | #141 |
| 2 | GPT-4o | 45.33 | #741 |
| 3 | Qwen 3 VL 8B Instruct | 41.33 | #982 |
| 4 | Qwen 3 VL 32B Instruct | 40.22 | #582 |
| 5 | InternVL3-14B | 36.89 | #1184 |
| 6 | Qwen 3 VL 4B Instruct | 35.56 | #1281 |
| 7 | Qwen 2.5 VL 72B Instruct | 34.89 | #845 |
| 8 | Qwen 2.5 72B Instruct | 33.33 | #1093 |
| 9 | Qwen 3 4B 2507 Instruct | 33.11 | #2174 |
| 10 | Qwen 2.5 7B Instruct | 31.33 | #2583 |
| 11 | InternVL3-8B | 30.22 | #1695 |
| 12 | Qwen 3 8B (Non-reasoning) | 30 | #1838 (Qwen 3 8B) |
| 13 | Qwen 3 30B A3B 2507 Instruct | 29.11 | #1078 |
| 14 | Qwen 2.5 3B Instruct | 27.11 | #5206 |
| 15 | Qwen 3 4B (Non-reasoning) | 26.67 | #2297 (Qwen 3 4B) |
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-logic-anomaly-detection · How It Works · Data refreshed daily, snapshot 2026-11-12.