Mathematical Reasoning on GSM8K (Accuracy & TPS)
81.96AccuracyDSB Cache (greedy)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DSB Cache (greedy)Backbone=LLaDA-1.5, Decode=Confidence, Cache=DSB, Block=DSB (greedy)2026.02 | 81.96 | 95.54 | |
| DSB Cache (const.)Backbone=LLaDA-1.5, Decode=Confidence, Cache=DSB, Block=DSB (const.)2026.02 | 81.73 | 93.99 | |
| DSB (const.)Backbone=LLaDA-1.5, Decode=Confidence, Cache=None, Block=DSB (const.)2026.02 | 81.2 | 48.6 | |
| Dual NaiveBackbone=LLaDA-1.5, Decode=Confidence, Cache=Dual, Block=Naive2026.02 | 80.82 | 86.42 | |
| DSB (greedy)Backbone=LLaDA-1.5, Decode=Confidence, Cache=None, Block=DSB (greedy)2026.02 | 80.74 | 49.63 | |
| VanillaBackbone=LLaDA-1.5, Decode=Vanilla, Cache=None, Block=Naive2026.02 | 80.52 | 14.05 | |
| DSB Cache (greedy)Backbone=LLaDA-8B-Instruct, Decode=Confidence, Cache=DSB, Block=DSB (greedy)2026.02 | 80.29 | 99.61 | |
| DSB Cache (const.)Backbone=LLaDA-8B-Instruct, Decode=Confidence, Cache=DSB, Block=DSB (const.)2026.02 | 80.14 | 98.1 | |
| Confidence + NaiveBackbone=LLaDA-1.5, Decode=Confidence, Cache=None, Block=Naive2026.02 | 80.14 | 49.63 | |
| DSB (greedy)Backbone=LLaDA-8B-Instruct, Decode=Confidence, Cache=None, Block=DSB (greedy)2026.02 | 78.54 | 51.03 | |
| DSB (const.)Backbone=LLaDA-8B-Instruct, Decode=Confidence, Cache=None, Block=DSB (const.)2026.02 | 78.17 | 50.36 | |
| VanillaBackbone=LLaDA-8B-Instruct, Decode=Vanilla, Cache=None, Block=Naive2026.02 | 77.79 | 14.94 | |
| Dual NaiveBackbone=LLaDA-8B-Instruct, Decode=Confidence, Cache=Dual, Block=Naive2026.02 | 77.4 | 92.26 | |
| Confidence + NaiveBackbone=LLaDA-8B-Instruct, Decode=Confidence, Cache=None, Block=Naive2026.02 | 77.26 | 48.7 | |
| VanillaBackbone=Dream-v0-Instruct-7B, Decode=Vanilla, Cache=None, Block=Naive2026.02 | 77.03 | 11.38 | |
| Confidence + NaiveBackbone=Dream-v0-Instruct-7B, Decode=Confidence, Cache=None, Block=Naive2026.02 | 75.51 | 45.79 | |
| VanillaBackbone=Dream-v0-Base-7B, Decode=Vanilla, Cache=None, Block=Naive2026.02 | 75.06 | 21.09 | |
| DSB (greedy)Backbone=Dream-v0-Instruct-7B, Decode=Confidence, Cache=None, Block=DSB (greedy)2026.02 | 75.06 | 38.52 | |
| DSB (const.)Backbone=Dream-v0-Base-7B, Decode=Confidence, Cache=None, Block=DSB (const.)2026.02 | 74.83 | 34.16 | |
| DSB (greedy)Backbone=Dream-v0-Base-7B, Decode=Confidence, Cache=None, Block=DSB (greedy)2026.02 | 74.75 | 34.7 | |
| Confidence + NaiveBackbone=Dream-v0-Base-7B, Decode=Confidence, Cache=None, Block=Naive2026.02 | 74.3 | 37 | |
| DSB Cache (const.)Backbone=Dream-v0-Base-7B, Decode=Confidence, Cache=DSB, Block=DSB (const.)2026.02 | 74.22 | 68.71 | |
| DSB (const.)Backbone=Dream-v0-Instruct-7B, Decode=Confidence, Cache=None, Block=DSB (const.)2026.02 | 74.22 | 45.22 | |
| Dual NaiveBackbone=Dream-v0-Base-7B, Decode=Confidence, Cache=Dual, Block=Naive2026.02 | 73.31 | 68.04 | |
| DSB Cache (greedy)Backbone=Dream-v0-Instruct-7B, Decode=Confidence, Cache=DSB, Block=DSB (greedy)2026.02 | 73.08 | 75.27 | |
| DSB Cache (greedy)Backbone=Dream-v0-Base-7B, Decode=Confidence, Cache=DSB, Block=DSB (greedy)2026.02 | 72.48 | 68.09 | |
| DSB Cache (const.)Backbone=Dream-v0-Instruct-7B, Decode=Confidence, Cache=DSB, Block=DSB (const.)2026.02 | 71.27 | 73.03 | |
| Dual NaiveBackbone=Dream-v0-Instruct-7B, Decode=Confidence, Cache=Dual, Block=Naive2026.02 | 67.32 | 72.18 |