Mathematical Reasoning on GSM8K (Accuracy & Throughput)
81.3AccuracydKV-Cache
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| dKV-CacheGen Length=512, shots=5-shot2026.01 | 81.3 | 4.2 | 1.7 | |
| Streaming-dLLMGen Length=512, shots=5-shot2026.01 | 81.2 | 69.8 | 28 | |
| LLaDA-1.5Gen Length=512, shots=5-shot2026.01 | 81 | 2.5 | 1 | |
| Prefix-CacheGen Length=512, shots=5-shot2026.01 | 81 | 8.2 | 3.3 | |
| Streaming-dLLMGen Length=256, shots=5-shot2026.01 | 80.8 | 66.2 | 10.5 | |
| dKV-CacheGen Length=256, shots=5-shot2026.01 | 80.7 | 10.8 | 1.7 | |
| Fast-dLLMGen Length=256, shots=5-shot2026.01 | 80.7 | 50 | 7.9 | |
| Prefix-CacheGen Length=256, shots=5-shot2026.01 | 80.6 | 24.4 | 3.9 | |
| LLaDA-1.5Gen Length=256, shots=5-shot2026.01 | 80.5 | 6.3 | 1 | |
| Fast-dLLMGen Length=512, shots=5-shot2026.01 | 80.4 | 25.8 | 10.3 |