CogSense-Bench: leaderboard
CogSense-Bench (Li et al., 2026): 1,000 visual cognition questions drawn from RAVEN, PGM, MaRs-VQA, Bongard-RWR+, Bongard-HOI, Bongard-LOGO, KiVA, STARE, ARC-AGI, ARC-AGI-2 and CVR, covering fluid and crystallized intelligence, visuospatial cognition, mental simulation and visual routines. Sample-weighted accuracy (%); higher is better.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around 2029. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 40.3 | #172 |
| 2 | Gemini 2.5 Flash | 36.3 | #452 |
| 3 | Gemma 3 27B | 32.7 | #1638 |
| 4 | Claude Sonnet 4 | 32.6 | #397 |
| 5 | O3 | 32.3 | #224 |
| 6 | Llama 4 Scout | 31.8 | #1787 |
| 7 | Grok 4 Fast | 31.7 | #438 |
| 8 | Intern-S1 | 17.4 | #616 |
| 9 | Ministral 3 14B | 16.5 | #1852 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cogsense-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.