Question Answering on ARC (pass@1)
88.68pass@1MCDIFFUSE
Evaluation Results
| Method | Links | |
|---|---|---|
| MCDIFFUSEModel Type=Diffusion, Parameters=7B, Slot Ordering Strategy=MCTS2026.02 | 88.68 | |
| Qwen3 8BModel Type=Autoregressive, Parameters=8B2026.02 | 88.4 | |
| ReFusion 7BModel Type=Diffusion, Parameters=7B, Slot Ordering Strategy=Vanilla2026.02 | 87.98 | |
| Qwen2.5 7BModel Type=Autoregressive, Parameters=7B2026.02 | 87.8 | |
| ReFusion 7BModel Type=Diffusion, Parameters=7B, Slot Ordering Strategy=Sequential2026.02 | 87.09 | |
| SCR (Ours)Backbone=Qwen2.5-7B-Instruct2026.01 | 86.43 | |
| GRPOBackbone=Qwen2.5-7B-Instruct2026.01 | 85.49 | |
| Dream 7BModel Type=Diffusion, Parameters=7B2026.02 | 84.95 | |
| SFT + GRPOBackbone=Llama3.1-8B-Instruct2026.01 | 84.73 | |
| SCR (Ours)Backbone=Llama3.1-8B-Instruct2026.01 | 84.56 | |
| BaseBackbone=Qwen2.5-7B-Instruct2026.01 | 84.39 | |
| LLaDA 8BModel Type=Diffusion, Parameters=8B2026.02 | 84.09 | |
| Llama-3.1 8BModel Type=Autoregressive, Parameters=8B2026.02 | 83.32 | |
| SCR-Stage IBackbone=Llama3.1-8B-Instruct2026.01 | 82.17 | |
| GRPOBackbone=Qwen2.5-3B-Instruct2026.01 | 81.14 | |
| ReFusion 7BModel Type=Diffusion, Parameters=7B, Slot Ordering Strategy=Random2026.02 | 80.98 | |
| SFT + GRPOBackbone=Qwen2.5-7B-Instruct2026.01 | 80.63 | |
| SCR-Stage IBackbone=Qwen2.5-7B-Instruct2026.01 | 80.63 | |
| SCR (Ours)Backbone=Qwen2.5-3B-Instruct2026.01 | 79.35 | |
| SFT + GRPOBackbone=Qwen2.5-3B-Instruct2026.01 | 79.01 | |
| SCR-SFTBackbone=Llama3.1-8B-Instruct2026.01 | 74.06 | |
| SCR-SFTBackbone=Qwen2.5-7B-Instruct2026.01 | 72.95 | |
| GRPOBackbone=Llama3.1-8B-Instruct2026.01 | 70.31 | |
| BaseBackbone=Qwen2.5-3B-Instruct2026.01 | 69.2 | |
| SCR-Stage IBackbone=Qwen2.5-3B-Instruct2026.01 | 67.32 | |
| Self-RefineBackbone=Qwen2.5-3B-Instruct2026.01 | 63.57 | |
| Self-RefineBackbone=Qwen2.5-7B-Instruct2026.01 | 57.51 | |
| BaseBackbone=Llama3.1-8B-Instruct2026.01 | 57.08 | |
| SCR-SFTBackbone=Qwen2.5-3B-Instruct2026.01 | 29.69 | |
| Self-RefineBackbone=Llama3.1-8B-Instruct2026.01 | 8.79 |