Mathematical Reasoning on GSM8K (Throughput and FLOPs)
29.75Throughput (TPS)LLaDA Instruct + dLLM-Cache
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| LLaDA Instruct + dLLM-CacheModel Size=8B, Caching=dLLM-Cache2025.05 | 29.75 | 4.28 | 2.92 | 5.81 | 0.7854 | |
| LLaDA Base + dLLM-CacheModel Size=8B, Caching=dLLM-Cache2025.05 | 23.19 | 3.17 | 3.21 | 5.02 | 0.7066 | |
| LLaDA BaseModel Size=8B, Caching=None2025.05 | 7.32 | 1 | 16.12 | 1 | 0.6906 | |
| LLaDA InstructModel Size=8B, Caching=None2025.05 | 6.95 | 1 | 16.97 | 1 | 0.7748 |