Mathematical Reasoning on GSM8K CoT 8-shot (test)
51.7TPSCDLM-Dream
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| CDLM-DreamEvaluation Protocol (Shots)=8-shot, Method=Consistency Diffusion2025.11 | 51.7 | 2.1 | 44.1 | 107.4 | 78.8 | — | — | — | |
| Fast-dLLMEvaluation Protocol (Shots)=8-shot, Parallel Decoding=true, Dual-cache KV=true, KV Caching=false2025.11 | 36.6 | 2.5 | 60.8 | 90.3 | 77.3 | — | — | — | |
| Fast-dLLMEvaluation Protocol (Shots)=8-shot, Parallel Decoding=true, Dual-cache KV=false, KV Caching=false2025.11 | 19.2 | 5 | 53.7 | 96 | 79.9 | — | — | — | |
| dLLM-CacheEvaluation Protocol (Shots)=8-shot, Parallel Decoding=false, Dual-cache KV=false, KV Caching=true2025.11 | 6.9 | 12.6 | 256 | 87.5 | 75.2 | — | — | — | |
| Dream-7B-InstructEvaluation Protocol (Shots)=8-shot, Parallel Decoding=false, Dual-cache KV=false, KV Caching=false2025.11 | 4.1 | 23.5 | 256 | 95.1 | 79.1 | — | — | — | |
| Original LLaDAStrategy=std unmask2026.05 | — | — | — | — | — | 55.3 | 58.4 | 59.5 | |
| Original LLaDAStrategy=D3IM2026.05 | — | — | — | — | — | 48 | 51.6 | 55.3 | |
| SCOPEStrategy=std unmask2026.05 | — | — | — | — | — | 52.8 | 55.3 | 59.7 | |
| SCOPEStrategy=D3IM2026.05 | — | — | — | — | — | 68.3 | 71.9 | 73.8 |