Code Generation on MBPP (test) (Efficiency Metrics)
145.99Throughput (tokens/s)ES-dLLM
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ES-dLLMBackbone=LLaDA-8B-Instruct, Number of shots=32026.03 | 145.99 | — | 16.3 | 39.4 | |
| DualCacheBackbone=LLaDA-8B-Instruct, Number of shots=32026.03 | 117.85 | — | 13.1 | 38.4 | |
| Streaming-dLLMGen Length=512, Evaluation Protocol=3-shot2026.01 | 92.4 | 55.8 | 10.6 | — | |
| Streaming-dLLMGen Length=256, Evaluation Protocol=3-shot2026.01 | 80.2 | 56.4 | 7.3 | — | |
| Fast-dLLMGen Length=256, Evaluation Protocol=3-shot2026.01 | 67.2 | 56.4 | 6.1 | — | |
| Fast-dLLMGen Length=512, Evaluation Protocol=3-shot2026.01 | 63.1 | 55.2 | 7.3 | — | |
| Prefix-CacheGen Length=256, Evaluation Protocol=3-shot2026.01 | 32.3 | 53.2 | 2.9 | — | |
| Prefix-CacheGen Length=512, Evaluation Protocol=3-shot2026.01 | 24.5 | 53.8 | 2.8 | — | |
| dKV-CacheGen Length=256, Evaluation Protocol=3-shot2026.01 | 14.7 | 54 | 1.3 | — | |
| dKV-CacheGen Length=512, Evaluation Protocol=3-shot2026.01 | 11.6 | 53 | 1.3 | — | |
| DreamGen Length=256, Evaluation Protocol=3-shot2026.01 | 11 | 56.6 | 1 | — | |
| LLaDABackbone=LLaDA-8B-Instruct, Number of shots=32026.03 | 8.98 | — | 1 | 41 | |
| DreamGen Length=512, Evaluation Protocol=3-shot2026.01 | 8.7 | 55.6 | 1 | — |