Abstract Reasoning on ARC-AGI v2 (test)
100AccuracyBest Human
Evaluation results
| Method | Links | |
|---|---|---|
| Best Human2026.05 | 100 | |
| Gemini 3 Pro#Params=N/A2026.05 | 31.1 | |
| Grok-4-thinking#Params=1.7T2026.05 | 16 | |
| GRAM#Params=10M, Supervision steps=162026.05 | 11.1 | |
| GPT 5.2 (low)#Params=N/A2026.05 | 9.7 | |
| TRM#Params=7M, Supervision steps=162026.05 | 7.8 | |
| HRM#Params=27M, Supervision steps=162026.05 | 5 | |
| o3-mini-high#Params=N/A2026.05 | 3 | |
| Deepseek-R1#Params=671B2026.05 | 1.3 | |
| Claude 3.7 16k#Params=N/A2026.05 | 0.7 | |
| Direct Pred#Params=27M, Supervision steps=162026.05 | 0 |