Mathematical Reasoning on MATH 4-shot (ACC, Latency, NFE)
53.8NFE (4-shot)dParallel
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| dParallelBackbone=LLaDA-Instruct-8B, Decoding Strategy=dParallel2026.04 | 53.8 | 30.22 | 1.7 | — | — | |
| dParallelBackbone=Dream-v0-Instruct-7B, Decoding Strategy=dParallel2026.04 | 57.3 | 35.94 | 1.6 | — | — | |
| R²-dLLM (T)Backbone=Dream-v0-Instruct-7B, Decoding Strategy=SFT to reduce redundancy2026.04 | 66.1 | 36.88 | 1.8 | — | — | |
| R²-dLLM (T)Backbone=LLaDA-Instruct-8B, Decoding Strategy=SFT to reduce redundancy2026.04 | 68.5 | 32.6 | 2.1 | — | — | |
| R²-dLLM (F)Backbone=Dream-v0-Instruct-7B, Decoding Strategy=training-free redundancy reduction2026.04 | 76.3 | 37.56 | 1.7 | — | — | |
| R²-dLLM (F)Backbone=LLaDA-Instruct-8B, Decoding Strategy=training-free redundancy reduction2026.04 | 79.7 | 32.26 | 2.9 | — | — | |
| LocalLeapBackbone=Dream-v0-Instruct-7B, Decoding Strategy=LocalLeap2026.04 | 81.5 | 37.18 | 1.9 | — | — | |
| LocalLeapBackbone=LLaDA-Instruct-8B, Decoding Strategy=LocalLeap2026.04 | 85.3 | 32.18 | 2.8 | — | — | |
| Fast-dLLMBackbone=Dream-v0-Instruct-7B, Decoding Strategy=Fast-dLLM2026.04 | 90.3 | 37.58 | 2.4 | — | — | |
| Fast-dLLMBackbone=LLaDA-Instruct-8B, Decoding Strategy=Fast-dLLM2026.04 | 107.8 | 32.6 | 3.3 | — | — | |
| VanillaBackbone=LLaDA-Instruct-8B, Decoding Strategy=Vanilla2026.04 | 256 | 33.2 | 9.7 | — | — | |
| VanillaBackbone=Dream-v0-Instruct-7B, Decoding Strategy=Vanilla2026.04 | 256 | 38.08 | 8.3 | — | — | |
| Fréchet profile decodingGen. Len.=256, Backbone=LLaDA-8B-Instruct, Margin=δ = 0.252026.06 | 358,178 | 32.5 | — | 102.5 | 3.48 | |
| FactorGen. Len.=256, Backbone=LLaDA-8B-Instruct2026.06 | 379,330 | 32.7 | — | 97.3 | 3.29 | |
| FrechetModel=Dream-v0-Base-7B, PrefixCache=Enabled, Block size=32, Hardware=H100, Len.=2562026.06 | 433,971 | 33.1 | — | 97 | 2.95 | |
| FactorModel=Dream-v0-Base-7B, PrefixCache=Enabled, Block size=32, Hardware=H100, Len.=2562026.06 | 442,791 | 33.2 | — | 92.4 | 2.8 | |
| ThresholdGen. Len.=256, Backbone=LLaDA-8B-Instruct, Threshold=τ = 0.92026.06 | 503,377 | 33.1 | — | 74.4 | 2.48 | |
| Fréchet profile decodingGen. Len.=512, Backbone=LLaDA-8B-Instruct, Margin=δ = 0.252026.06 | 545,993 | 35.5 | — | 77.7 | 3.96 | |
| ThresholdModel=Dream-v0-Base-7B, PrefixCache=Enabled, Block size=32, Hardware=H100, Len.=2562026.06 | 568,643 | 34 | — | 72.6 | 2.25 | |
| FactorGen. Len.=512, Backbone=LLaDA-8B-Instruct2026.06 | 585,176 | 35.3 | — | 69.1 | 3.7 | |
| FrechetModel=Dream-v0-Base-7B, PrefixCache=Enabled, Block size=32, Hardware=H100, Len.=5122026.06 | 595,000 | 33.5 | — | 127.5 | 4.3 | |
| FactorModel=Dream-v0-Base-7B, PrefixCache=Enabled, Block size=32, Hardware=H100, Len.=5122026.06 | 623,643 | 34.1 | — | 118.6 | 4.1 | |
| ThresholdGen. Len.=512, Backbone=LLaDA-8B-Instruct, Threshold=τ = 0.92026.06 | 764,793 | 36.1 | — | 56.4 | 2.83 | |
| ThresholdModel=Dream-v0-Base-7B, PrefixCache=Enabled, Block size=32, Hardware=H100, Len.=5122026.06 | 776,016 | 35.3 | — | 99.1 | 3.3 |