Mathematical Reasoning on MATH 4-shot (test)
37.2AccuracyLLaDA
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| LLaDAGen Length=512, Evaluation Protocol=4-shot2026.01 | 37.2 | 68.8 | 1 | — | — | — | — | |
| dKV-CacheGen Length=512, Evaluation Protocol=4-shot2026.01 | 36.3 | 42.7 | 1.6 | — | — | — | — | |
| Prefix-CacheGen Length=512, Evaluation Protocol=4-shot2026.01 | 36.2 | 23.3 | 3 | — | — | — | — | |
| Streaming-dLLMGen Length=512, Evaluation Protocol=4-shot2026.01 | 36.1 | 6.8 | 10.1 | — | — | — | — | |
| Fast-dLLMGen Length=512, Evaluation Protocol=4-shot2026.01 | 36 | 10 | 6.9 | — | — | — | — | |
| LLaDAGen Length=256, Evaluation Protocol=4-shot2026.01 | 33.5 | 28.6 | 1 | — | — | — | — | |
| VanillaBackbone=LLaDA-8B-Instruct, KV Cache=true, Generation Length=2562026.05 | 33.5 | — | 1 | 9.2 | — | — | — | |
| Fast-dLLM (Par.)Shots=4-shot, parallel decoding=true2025.11 | 33.4 | 9.9 | — | 25.1 | 96.9 | 248.6 | — | |
| dKV-CacheGen Length=256, Evaluation Protocol=4-shot2026.01 | 33.3 | 17.4 | 1.6 | — | — | — | — | |
| Prefix-CacheGen Length=256, Evaluation Protocol=4-shot2026.01 | 33.3 | 8.6 | 3.3 | — | — | — | — | |
| Streaming-dLLMGen Length=256, Evaluation Protocol=4-shot2026.01 | 33.3 | 3.5 | 8.2 | — | — | — | — | |
| dLLM-CacheShots=4-shot, cache_mode=dLLM-Cache2025.11 | 33.3 | 10.9 | — | 22.7 | 256 | 248.5 | — | |
| Fast-dLLMGen Length=256, Evaluation Protocol=4-shot2026.01 | 33.2 | 5.1 | 5.6 | — | — | — | — | |
| Fast-dLLMBackbone=LLaDA-8B-Instruct, KV Cache=true, Generation Length=2562026.05 | 33.2 | — | 5.6 | 51.5 | — | — | — | |
| TSPDBackbone=LLaDA-8B-Instruct, KV Cache=true, Generation Length=2562026.05 | 33.2 | — | 8.3 | 76.4 | — | — | — | |
| Credit DecodingBackbone=LLaDA-8B-Instruct, KV Cache=true, Generation Length=2562026.05 | 33.1 | — | 5.5 | 50.6 | — | — | — | |
| TSPDBackbone=LLaDA-8B-Instruct, KV Cache=false, Generation Length=2562026.05 | 33 | — | 4 | 36.8 | — | — | — | |
| ProphetBackbone=LLaDA-8B-Instruct, KV Cache=false, Generation Length=2562026.05 | 32.7 | — | 1.7 | 15.7 | — | — | — | |
| Fast-dLLM (Par.+D.C.)Shots=4-shot, parallel decoding=true, dual-cache KV=true2025.11 | 32.5 | 5 | — | 49.7 | 107 | 248.2 | — | |
| Learn2PDBackbone=LLaDA-8B-Instruct, KV Cache=false, Generation Length=2562026.05 | 32.3 | — | 3.2 | 29.4 | — | — | — | |
| CDLM-LLaDAShots=4-shot2025.11 | 28.3 | 4.2 | — | 50.2 | 75.3 | 210.3 | — | |
| LLaDA-8B-InstructShots=4-shot2025.11 | 24.1 | 25.7 | — | 8.8 | 256 | 226.4 | — | |
| DualCacheBackbone=LLaDA-8B-Base, Speedup=3.8×2026.03 | — | — | — | 57.89 | — | — | 30.48 | |
| ES-dLLMBackbone=LLaDA-8B-Base, Speedup=6.2×2026.03 | — | — | — | 94.46 | — | — | 30.6 | |
| LLaDABackbone=LLaDA-8B-Base, Speedup=1.0×2026.03 | — | — | — | 15.12 | — | — | 32.38 |