Mathematical Reasoning on GSM8K (test) (Accuracy & Latency/Speedup)
81.3AccuracydKV-Cache
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| dKV-CacheGen Length=512, Evaluation Protocol=5-shot2026.01 | 81.3 | 48.1 | 1.7 | — | |
| Streaming-dLLMGen Length=512, Evaluation Protocol=5-shot2026.01 | 81.2 | 2.2 | 37.2 | — | |
| LLaDA-1.5Gen Length=512, Evaluation Protocol=5-shot2026.01 | 81 | 81.8 | 1 | — | |
| Prefix-CacheGen Length=512, Evaluation Protocol=5-shot2026.01 | 81 | 24.3 | 3.4 | — | |
| Streaming-dLLMGen Length=256, Evaluation Protocol=5-shot2026.01 | 80.8 | 2 | 17.3 | — | |
| dKV-CacheGen Length=256, Evaluation Protocol=5-shot2026.01 | 80.7 | 20.1 | 1.7 | — | |
| Fast-dLLMGen Length=256, Evaluation Protocol=5-shot2026.01 | 80.7 | 4.4 | 7.8 | — | |
| Prefix-CacheGen Length=256, Evaluation Protocol=5-shot2026.01 | 80.6 | 8.9 | 3.9 | — | |
| LLaDA-1.5Gen Length=256, Evaluation Protocol=5-shot2026.01 | 80.5 | 34.5 | 1 | — | |
| Fast-dLLMGen Length=512, Evaluation Protocol=5-shot2026.01 | 80.4 | 7.8 | 10.5 | — | |
| DreamGen Length=256, Prompting Strategy=5-shot, Chain-of-Thought=true2026.01 | 74.8 | 18.6 | 1 | — | |
| DreamGen Length=256, Evaluation Protocol=5-shot2026.01 | 74.8 | — | 1 | 9 | |
| Streaming-dLLMGen Length=512, Prompting Strategy=5-shot, Chain-of-Thought=true2026.01 | 74.7 | 2.4 | 21 | — | |
| Streaming-dLLMGen Length=512, Evaluation Protocol=5-shot2026.01 | 74.7 | — | 13.3 | 94.1 | |
| DreamGen Length=512, Prompting Strategy=5-shot, Chain-of-Thought=true2026.01 | 74.2 | 50.3 | 1 | — | |
| Prefix-CacheGen Length=512, Prompting Strategy=5-shot, Chain-of-Thought=true2026.01 | 74.2 | 21.6 | 2.3 | — | |
| DreamGen Length=512, Evaluation Protocol=5-shot2026.01 | 74.2 | — | 1 | 7.1 | |
| Prefix-CacheGen Length=512, Evaluation Protocol=5-shot2026.01 | 74.2 | — | 3.3 | 23.6 | |
| Fast-dLLMGen Length=512, Prompting Strategy=5-shot, Chain-of-Thought=true2026.01 | 74.1 | 10.7 | 4.7 | — | |
| Fast-dLLMGen Length=512, Evaluation Protocol=5-shot2026.01 | 74.1 | — | 5.9 | 41.7 | |
| Prefix-CacheGen Length=256, Prompting Strategy=5-shot, Chain-of-Thought=true2026.01 | 74 | 8.1 | 2.3 | — | |
| Streaming-dLLMGen Length=256, Prompting Strategy=5-shot, Chain-of-Thought=true2026.01 | 74 | 2.1 | 8.9 | — | |
| Prefix-CacheGen Length=256, Evaluation Protocol=5-shot2026.01 | 74 | — | 3.5 | 31.5 | |
| Streaming-dLLMGen Length=256, Evaluation Protocol=5-shot2026.01 | 74 | — | 8.4 | 75.5 | |
| dKV-CacheGen Length=256, Prompting Strategy=5-shot, Chain-of-Thought=true2026.01 | 73.6 | 15 | 1.2 | — | |
| dKV-CacheGen Length=256, Evaluation Protocol=5-shot2026.01 | 73.6 | — | 1.9 | 17 | |
| Fast-dLLMGen Length=256, Prompting Strategy=5-shot, Chain-of-Thought=true2026.01 | 73.5 | 5.3 | 3.5 | — | |
| Fast-dLLMGen Length=256, Evaluation Protocol=5-shot2026.01 | 73.5 | — | 5.3 | 47.9 | |
| dKV-CacheGen Length=512, Prompting Strategy=5-shot, Chain-of-Thought=true2026.01 | 71.6 | 39.9 | 1.3 | — | |
| dKV-CacheGen Length=512, Evaluation Protocol=5-shot2026.01 | 71.6 | — | 1.8 | 12.8 |