AI Reasoning on ARC Easy (held-out)
55.4Pass@2 AccuracySPRM (Ours)
Evaluation Results
| Method | Links | |
|---|---|---|
| SPRM (Ours)#Params=14M, Pretraining Data=None2026.04 | 55.4 | |
| DRM (Ours)#Params=14M, Pretraining Data=None2026.04 | 55 | |
| TRM (Baseline)#Params=14M, Pretraining Data=None2026.04 | 53.2 | |
| SPRM (Ours)#Params=7M, Pretraining Data=None2026.04 | 50.7 | |
| DRM (Ours)#Params=7M, Pretraining Data=None2026.04 | 50.5 | |
| TRM (Baseline)#Params=7M, Pretraining Data=None2026.04 | 45.7 | |
| LLM (NVARC)#Params=4B, Pretraining Data=None2026.04 | 43.2 | |
| TRM w/o inner loop#Params=34M, Pretraining Data=None2026.04 | 39.8 | |
| Transformer w/ deep sup#Params=70M, Pretraining Data=None2026.04 | 37.2 | |
| TRM w/o deep sup#Params=7M, Pretraining Data=None2026.04 | 34.7 | |
| Transformer#Params=70M, Pretraining Data=None2026.04 | 25.5 | |
| Transformer#Params=7M, Pretraining Data=None2026.04 | 10.9 | |
| Masked Diffusion (Standard)#Params=70M, Pretraining Data=None2026.04 | 0.7 | |
| Masked Diffusion (Standard)#Params=7M, Pretraining Data=None2026.04 | 0 |