Multi-task Language Understanding on MMLU (Speedup, τ, Acc.)
1.49SpeedupSD + DM
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SD + DMModel=Qwen3 4B/32B, L=102026.02 | 1.49 | 8.6 | 82.06 | |
| SDModel=Qwen3 4B/32B, L=102026.02 | 1.12 | 6.48 | 83.26 | |
| SD + DMModel=Llama-3.1 8B/70B-Instruct, L=52026.02 | 1.09 | 4.43 | 85.5 | |
| SD + DMModel=Qwen3 4B/32B, L=52026.02 | 1.05 | 5.15 | 82.15 | |
| SD + DMModel=Llama-3.1 8B/70B-Instruct, L=102026.02 | 1.02 | 6.06 | 85.34 | |
| SDModel=Llama-3.1 8B/70B-Instruct, L=52026.02 | 1 | 4.02 | 85.94 | |
| SDModel=Qwen3 4B/32B, L=52026.02 | 1 | 4.62 | 83.21 | |
| SDModel=Llama-3.1 8B/70B-Instruct, L=102026.02 | 0.86 | 5.25 | 86.06 |