Abstract Reasoning on ARC-AGI 2
60AccuracyHuman panel (average)
Evaluation results
| Method | Links | |
|---|---|---|
| Human panel (average)#Params=–2026.06 | 60 | |
| Qwen3 + DIARC#Params=3.6B2026.06 | 33.06 | |
| Qwen3 + SFT only#Params=3.6B2026.06 | 30.83 | |
| Grok-4#Params=1.7T2026.06 | 29.4 | |
| GPT-5.4#Params=N/A2026.06 | 29.2 | |
| Kimi K2.5#Params=1T2026.06 | 11.8 | |
| GRAM#Params=10.9M2026.06 | 11.1 | |
| VARC (ensemble)#Params=73M2026.06 | 11.1 | |
| Minitron + DIARC#Params=6.9B2026.06 | 10.28 | |
| Minitron + SFT only#Params=6.9B2026.06 | 9.86 | |
| VARC#Params=18M2026.06 | 8.3 | |
| Opus 4.5#Params=N/A2026.06 | 7.8 | |
| Llama-3.2 + DIARC#Params=2.6B2026.06 | 6.53 | |
| HRM#Params=27M2026.06 | 5 | |
| Llama-3.2 + SFT only#Params=2.6B2026.06 | 4.44 | |
| Llama-3.2#Params=3B2026.06 | 0 | |
| Minitron#Params=8B2026.06 | 0 | |
| Qwen3#Params=4B2026.06 | 0 |