Abstraction and Reasoning on ARC-AGI
58.2ARC-1 ScoreSSL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SSLBackbone=Qwen2.5, Model Scale=7B, Reward Strategy=Sweet Spot Learning2026.01 | 58.2 | 40.3 | |
| RL-ContinuousBackbone=Qwen2.5, Model Scale=7B, Reward Strategy=Continuous2026.01 | 53.7 | 38.4 | |
| RL-BinaryBackbone=Qwen2.5, Model Scale=7B, Reward Strategy=Binary2026.01 | 52.8 | 39 | |
| SR²Params=3.4M2025.10 | 44.3 | 6.7 | |
| HRMParams=27.3M2025.10 | 40.3 | 5 | |
| SSLBackbone=Qwen2.5, Model Scale=3B, Reward Strategy=Sweet Spot Learning2026.01 | 33.3 | 20.3 | |
| RL-ContinuousBackbone=Qwen2.5, Model Scale=3B, Reward Strategy=Continuous2026.01 | 30.9 | 17.9 | |
| RL-BinaryBackbone=Qwen2.5, Model Scale=3B, Reward Strategy=Binary2026.01 | 28.3 | 14.9 | |
| TransformerParams=27.3M2025.10 | 21 | 0 |