Abstract Reasoning on ARC-AGI 1
97AccuracyQwen3 + DIARC
Evaluation results
| Method | Links | |
|---|---|---|
| Qwen3 + DIARC#Params=3.6B2026.06 | 97 | |
| Qwen3 + SFT only#Params=3.6B2026.06 | 96.25 | |
| Grok-4#Params=1.7T2026.06 | 79.6 | |
| Minitron + DIARC#Params=6.9B2026.06 | 72.38 | |
| Minitron + SFT only#Params=6.9B2026.06 | 71.25 | |
| GPT-5.4#Params=N/A2026.06 | 68.2 | |
| Kimi K2.5#Params=1T2026.06 | 65.3 | |
| Human panel (average)#Params=–2026.06 | 64.2 | |
| Llama-3.2 + DIARC#Params=2.6B2026.06 | 61.5 | |
| VARC (ensemble)#Params=73M2026.06 | 60.4 | |
| Llama-3.2 + SFT only#Params=2.6B2026.06 | 59.38 | |
| VARC#Params=18M2026.06 | 54.5 | |
| GRAM#Params=10.9M2026.06 | 52 | |
| HRM#Params=27M2026.06 | 40.3 | |
| Opus 4.5#Params=N/A2026.06 | 40 | |
| Minitron#Params=8B2026.06 | 1.12 | |
| Qwen3#Params=4B2026.06 | 0.63 | |
| Llama-3.2#Params=3B2026.06 | 0.5 |