Mathematical Reasoning on GSM8K (test) (Accuracy, SpeedUp)
76.4AccuracyPBS (Single Token)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PBS (Single Token)Max Length=512, Backbone=DREAM-7B-Base2026.02 | 76.4 | 0.56 | |
| SOARMax Length=512, Backbone=DREAM-7B-Base2026.02 | 75.7 | 1.07 | |
| PBS (Single Token)Max Length=256, Backbone=DREAM-7B-Base2026.02 | 75.2 | 0.57 | |
| PBS (Parallel)Max Length=512, Backbone=DREAM-7B-Base2026.02 | 75.1 | 1.01 | |
| SOARMax Length=256, Backbone=DREAM-7B-Base2026.02 | 74.7 | 0.95 | |
| Greedy (Adaptive Parallel)Max Length=512, Backbone=DREAM-7B-Base2026.02 | 74.6 | 1.83 | |
| GreedyMax Length=512, Backbone=DREAM-7B-Base2026.02 | 74.5 | — | |
| GreedyMax Length=256, Backbone=DREAM-7B-Base2026.02 | 73.7 | — | |
| PBS (Parallel)Max Length=256, Backbone=DREAM-7B-Base2026.02 | 73.5 | 1.03 | |
| Greedy (Adaptive Parallel)Max Length=256, Backbone=DREAM-7B-Base2026.02 | 73.2 | 2.19 | |
| PBS (Single Token)Max Length=512, Backbone=LLADA-8B-Base2026.02 | 72.1 | 0.48 | |
| PBS (Single Token)Max Length=256, Backbone=LLADA-8B-Base2026.02 | 71.6 | 0.47 | |
| SOARMax Length=512, Backbone=LLADA-8B-Base2026.02 | 71.5 | 1.43 | |
| SOARMax Length=256, Backbone=LLADA-8B-Base2026.02 | 71.3 | 1.18 | |
| Greedy (Adaptive Parallel)Max Length=512, Backbone=LLADA-8B-Base2026.02 | 71 | 1.92 | |
| GreedyMax Length=512, Backbone=LLADA-8B-Base2026.02 | 70.9 | — | |
| GreedyMax Length=256, Backbone=LLADA-8B-Base2026.02 | 70.4 | — | |
| Greedy (Adaptive Parallel)Max Length=256, Backbone=LLADA-8B-Base2026.02 | 70.4 | 1.69 | |
| PBS (Parallel)Max Length=256, Backbone=LLADA-8B-Base2026.02 | 69.5 | 0.98 | |
| PBS (Parallel)Max Length=512, Backbone=LLADA-8B-Base2026.02 | 68.1 | 0.99 |