Mathematical Reasoning on GSM8K (5-shot)
82.2AccuracyDepCap
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| DepCapBackbone=LLaDA-1.5, Parallel Decode=CAP-Decoding, Block=DepGA-Block2026.04 | 82.2 | 20.1 | — | 73.9 | — | — | — | |
| AdaBlockBackbone=LLaDA-1.5, Parallel Decode=Confidence, Block=AdaBlock2026.04 | 82.1 | 17.8 | — | 80.1 | — | — | — | |
| CAP-DecodingBackbone=LLaDA-1.5, Parallel Decode=CAP-Decoding, Block=|B0| = 322026.04 | 82.1 | 20.6 | — | 69.1 | — | — | — | |
| CLADBackbone Model=LLaDA-1.5, Shot Setting=5-shot2026.05 | 81.65 | 23.07 | 4.71 | — | — | — | — | |
| KLASSBackbone Model=LLaDA-1.5, Shot Setting=5-shot2026.05 | 81.5 | 4.77 | 0.97 | — | — | — | — | |
| ConfidenceBackbone=LLaDA-1.5, Parallel Decode=Confidence, Block=|B0| = 162026.04 | 81.4 | 17.5 | — | 81.9 | — | — | — | |
| Fast-dLLMBackbone Model=LLaDA-1.5, Shot Setting=5-shot2026.05 | 81.35 | 16.61 | 3.39 | — | — | — | — | |
| ConfidenceBackbone=LLaDA-1.5, Parallel Decode=Confidence, Block=|B0| = 322026.04 | 81.2 | 18.9 | — | 75.5 | — | — | — | |
| DAPDBackbone Model=LLaDA-1.5, Shot Setting=5-shot2026.05 | 81.2 | 10.86 | 2.22 | — | — | — | — | |
| VanillaBackbone Model=LLaDA-1.5, Shot Setting=5-shot2026.05 | 81.12 | 4.9 | 1 | — | — | — | — | |
| DAWNBackbone Model=LLaDA-1.5, Shot Setting=5-shot2026.05 | 81.12 | 22.63 | 4.62 | — | — | — | — | |
| VanillaBackbone=LLaDA-1.5, Parallel Decode=Vanilla, Block=|B0| = 322026.04 | 81.1 | 5.6 | — | 256 | — | — | — | |
| R²-dLLM (T)Backbone=Dream-v0-Instruct-7B, Decoding Strategy=SFT to reduce redundancy2026.04 | 80.82 | — | — | 44.9 | 1.3 | — | — | |
| ConfidenceBackbone=LLaDA-1.5, Parallel Decode=Confidence, Block=|B0| = 642026.04 | 80.7 | 19.5 | — | 72.8 | — | — | — | |
| DepGA-BlockBackbone=LLaDA-1.5, Parallel Decode=Confidence, Block=DepGA-Block2026.04 | 80.6 | 18.1 | — | 80.1 | — | — | — | |
| DepGA-BlockBackbone=LLaDA-8B-Instruct, Parallel Decode=Confidence, Block=DepGA-Block2026.04 | 79.2 | 18.7 | — | 81.4 | — | — | — | |
| Fast-dLLM (threshold)Generation Length=2562025.12 | 79 | 90.4 | 1 | — | — | — | — | |
| LLaDABlock size=5122026.06 | 79 | 5.89 | — | — | — | — | — | |
| Bf16Backbone=Llama-3.1 8B-Instruct, Bits=16, Prompting=5-shot CoT2025.02 | 78.85 | — | — | — | — | — | — | |
| LLaDABlock size=10242026.06 | 78.85 | 1.86 | — | — | — | — | — | |
| d3LLM-TF-LLaDABlock size=10242026.06 | 78.85 | 46.21 | — | — | — | — | — | |
| DepCapBackbone=LLaDA-8B-Instruct, Parallel Decode=CAP-Decoding, Block=DepGA-Block2026.04 | 78.8 | 21.4 | — | 74.8 | — | — | — | |
| VanillaBackbone=Dream-v0-Instruct-7B, Decoding Strategy=Vanilla2026.04 | 78.77 | — | — | 256 | 10.4 | — | — | |
| PolarQuant44Backbone=Llama-3.1 8B-Instruct, Bits=4.25, Prompting=5-shot CoT2025.02 | 78.77 | — | — | — | — | — | — | |
| ZipCache-4Backbone=Llama-3.1 8B-Instruct, Bits=4.25, Prompting=5-shot CoT2025.02 | 78.7 | — | — | — | — | — | — | |
| dParallelBackbone=Dream-v0-Instruct-7B, Decoding Strategy=dParallel2026.04 | 78.47 | — | — | 38.7 | 1.3 | — | — | |
| KIVI-4Backbone=Llama-3.1 8B-Instruct, Bits=4.25, Prompting=5-shot CoT2025.02 | 78.32 | — | — | — | — | — | — | |
| ConfidenceBackbone=LLaDA-8B-Instruct, Parallel Decode=Confidence, Block=|B0| = 162026.04 | 78.3 | 18.3 | — | 84.1 | — | — | — | |
| d3LLM-TF-LLaDABlock size=2562026.06 | 78.24 | 78.85 | — | — | — | — | — | |
| d3LLM-TF-LLaDABlock size=5122026.06 | 78.24 | 68.33 | — | — | — | — | — | |
| LocalLeapBackbone=LLaDA-Instruct-8B, Decoding Strategy=LocalLeap2026.04 | 78.16 | — | — | 64.4 | 2.3 | — | — | |
| FactorGen. Len.=256, Backbone=LLaDA-8B-Instruct, Cache=PrefixCache2026.06 | 78.1 | 96 | — | 79,047 | — | 3.9 | — | |
| CadLLMGeneration Length=2562025.12 | 78.01 | 120.07 | 1.33 | — | — | — | — | |
| Fast-dLLMBackbone=LLaDA-Instruct-8B, Decoding Strategy=Fast-dLLM2026.04 | 78.01 | — | — | 86.4 | 2.7 | — | — | |
| Fast-dLLM-LLaDABlock size=5122026.06 | 78 | 64.55 | — | — | — | — | — | |
| AdaBlockBackbone=LLaDA-8B-Instruct, Parallel Decode=Confidence, Block=AdaBlock2026.04 | 77.9 | 18.4 | — | 83.9 | — | — | — | |
| R²-dLLM (T)Backbone=LLaDA-Instruct-8B, Decoding Strategy=SFT to reduce redundancy2026.04 | 77.86 | — | — | 51.9 | 1.7 | — | — | |
| LLaDABlock size=2562026.06 | 77.86 | 15.19 | — | — | — | — | — | |
| VanillaBackbone=LLaDA-8B-Instruct, Parallel Decode=Vanilla, Block=|B0| = 322026.04 | 77.6 | 6 | — | 256 | — | — | — | |
| ThresholdGen. Len.=256, Backbone=LLaDA-8B-Instruct, Cache=PrefixCache, Threshold=τ = 0.92026.06 | 77.6 | 73.8 | — | 107,135 | — | 2.88 | — | |
| R²-dLLM (F)Backbone=LLaDA-Instruct-8B, Decoding Strategy=training-free redundancy reduction2026.04 | 77.56 | — | — | 63.1 | 2.3 | — | — | |
| ConfidenceBackbone=LLaDA-8B-Instruct, Parallel Decode=Confidence, Block=|B0| = 642026.04 | 77.5 | 20.1 | — | 75.6 | — | — | — | |
| Fast-dLLM-LLaDABlock size=10242026.06 | 77.41 | 35.2 | — | — | — | — | — | |
| AsyncLane-LLaDABlock size=5122026.06 | 77.33 | 131.66 | — | — | — | — | — | |
| ConfidenceBackbone=LLaDA-8B-Instruct, Parallel Decode=Confidence, Block=|B0| = 322026.04 | 77.3 | 19.6 | — | 78.6 | — | — | — | |
| CAP-DecodingBackbone=LLaDA-8B-Instruct, Parallel Decode=CAP-Decoding, Block=|B0| = 322026.04 | 77.3 | 21.6 | — | 71.5 | — | — | — | |
| Fréchet profile decodingGen. Len.=256, Backbone=LLaDA-8B-Instruct, Cache=PrefixCache, Margin=δ = 0.252026.06 | 77.2 | 103.8 | — | 72,881 | — | 4.24 | — | |
| VanillaBackbone=LLaDA-Instruct-8B, Decoding Strategy=Vanilla2026.04 | 77.03 | — | — | 256 | 13.1 | — | — | |
| AsyncLane-LLaDABlock size=10242026.06 | 76.8 | 136.34 | — | — | — | — | — | |
| Fast-dLLMBackbone=Dream-v0-Instruct-7B, Decoding Strategy=Fast-dLLM2026.04 | 76.57 | — | — | 63.6 | 1.7 | — | — | |
| Fast-dLLM-LLaDABlock size=2562026.06 | 76.57 | 84.01 | — | — | — | — | — | |
| ThresholdGen. Len.=512, Backbone=LLaDA-8B-Instruct, Cache=PrefixCache, Threshold=τ = 0.92026.06 | 76.5 | 45.3 | — | 129,791 | — | 2.75 | — | |
| AsyncLane-LLaDABlock size=2562026.06 | 76.42 | 130.29 | — | — | — | — | — | |
| Int-4Backbone=Llama-3.1 8B-Instruct, Bits=4.25, Prompting=5-shot CoT2025.02 | 76.35 | — | — | — | — | — | — | |
| Fast-dLLM (factor)Generation Length=2562025.12 | 76.19 | 119.18 | 1.32 | — | — | — | — | |
| LocalLeapBackbone=Dream-v0-Instruct-7B, Decoding Strategy=LocalLeap2026.04 | 76.19 | — | — | 61 | 1.4 | — | — | |
| Fréchet profile decodingGen. Len.=512, Backbone=LLaDA-8B-Instruct, Cache=PrefixCache, Margin=δ = 0.252026.06 | 75.6 | 59.4 | — | 91,239 | — | 3.9 | — | |
| R²-dLLM (F)Backbone=Dream-v0-Instruct-7B, Decoding Strategy=training-free redundancy reduction2026.04 | 75.51 | — | — | 53.3 | 1.4 | — | — | |
| CadLLMGeneration Length=5122025.12 | 75.44 | 107.79 | 1.37 | — | — | — | — | |
| dParallelBackbone=LLaDA-Instruct-8B, Decoding Strategy=dParallel2026.04 | 75.36 | — | — | 39.7 | 1.4 | — | — | |
| Fast-dLLM (threshold)Generation Length=5122025.12 | 75.28 | 78.77 | 1 | — | — | — | — | |
| VanillaBackbone Model=Dream-v0-Base-7B, Shot Setting=5-shot2026.05 | 75.13 | 7.16 | 1 | — | — | — | — | |
| VanillaBackbone=Dream-v0-base-7B, Parallel Decode=Vanilla, Block=|B0| = 322026.04 | 75.1 | 7.8 | — | 256 | — | — | — | |
| AdaBlockBackbone=Dream-v0-base-7B, Parallel Decode=Confidence, Block=AdaBlock2026.04 | 75.1 | 13.1 | — | 154.9 | — | — | — | |
| DepGA-BlockBackbone=Dream-v0-base-7B, Parallel Decode=Confidence, Block=DepGA-Block2026.04 | 75 | 13.5 | — | 149 | — | — | — | |
| DepCapBackbone=Dream-v0-base-7B, Parallel Decode=CAP-Decoding, Block=DepGA-Block2026.04 | 74.8 | 13 | — | 146.6 | — | — | — | |
| ConfidenceBackbone=Dream-v0-base-7B, Parallel Decode=Confidence, Block=|B0| = 162026.04 | 74.7 | 13.7 | — | 148.1 | — | — | — | |
| ConfidenceBackbone=Dream-v0-base-7B, Parallel Decode=Confidence, Block=|B0| = 642026.04 | 74.6 | 12.8 | — | 157.9 | — | — | — | |
| FactorGen. Len.=512, Backbone=LLaDA-8B-Instruct, Cache=PrefixCache2026.06 | 74.6 | 53.3 | — | 99,580 | — | 3.61 | — | |
| ConfidenceBackbone=Dream-v0-base-7B, Parallel Decode=Confidence, Block=|B0| = 322026.04 | 74.5 | 13.2 | — | 152.4 | — | — | — | |
| Fast-dLLM (factor)Generation Length=5122025.12 | 74.45 | 100.66 | 1.27 | — | — | — | — | |
| Fast-dLLMBackbone Model=Dream-v0-Base-7B, Shot Setting=5-shot2026.05 | 74.15 | 12.04 | 1.68 | — | — | — | — | |
| CAP-DecodingBackbone=Dream-v0-base-7B, Parallel Decode=CAP-Decoding, Block=|B0| = 322026.04 | 74.1 | 12.7 | — | 149.9 | — | — | — | |
| CLADBackbone Model=Dream-v0-Base-7B, Shot Setting=5-shot2026.05 | 73.54 | 12.67 | 1.77 | — | — | — | — | |
| DAWNBackbone Model=Dream-v0-Base-7B, Shot Setting=5-shot2026.05 | 72.71 | 13.39 | 1.87 | — | — | — | — | |
| KLASSBackbone Model=Dream-v0-Base-7B, Shot Setting=5-shot2026.05 | 72.4 | 7.1 | 0.99 | — | — | — | — | |
| DAPDBackbone Model=Dream-v0-Base-7B, Shot Setting=5-shot2026.05 | 71.95 | 10.62 | 1.48 | — | — | — | — | |
| ZipCache-4Backbone=Llama-2 7B-Chat, Bits=4.25, Prompting=5-shot CoT2025.02 | 23.12 | — | — | — | — | — | — | |
| PolarQuant44Backbone=Llama-2 7B-Chat, Bits=4.25, Prompting=5-shot CoT2025.02 | 22.61 | — | — | — | — | — | — | |
| KIVI-4Backbone=Llama-2 7B-Chat, Bits=4.25, Prompting=5-shot CoT2025.02 | 21.61 | — | — | — | — | — | — | |
| Bf16Backbone=Llama-2 7B-Chat, Bits=16, Prompting=5-shot CoT2025.02 | 20.92 | — | — | — | — | — | — | |
| Int-4Backbone=Llama-2 7B-Chat, Bits=4.25, Prompting=5-shot CoT2025.02 | 19.79 | — | — | — | — | — | — | |
| Confidence (Top-1)Generation Length (L)=128, Block Size (B)=128, Number of decoding blocks (K)=12026.06 | — | — | — | — | — | — | 62 | |
| Confidence (Top-1)Generation Length (L)=128, Block Size (B)=32, Number of decoding blocks (K)=42026.06 | — | — | — | — | — | — | 72.8 | |
| Confidence (Top-1)Generation Length (L)=128, Block Size (B)=16, Number of decoding blocks (K)=82026.06 | — | — | — | — | — | — | 74.4 | |
| Confidence (Top-1)Generation Length (L)=256, Block Size (B)=256, Number of decoding blocks (K)=12026.06 | — | — | — | — | — | — | 62.8 | |
| Confidence (Top-1)Generation Length (L)=256, Block Size (B)=64, Number of decoding blocks (K)=42026.06 | — | — | — | — | — | — | 76 | |
| Confidence (Top-1)Generation Length (L)=256, Block Size (B)=32, Number of decoding blocks (K)=82026.06 | — | — | — | — | — | — | 75.8 | |
| Confidence (Top-1)Generation Length (L)=512, Block Size (B)=512, Number of decoding blocks (K)=12026.06 | — | — | — | — | — | — | 76.5 | |
| Confidence (Top-1)Generation Length (L)=512, Block Size (B)=128, Number of decoding blocks (K)=42026.06 | — | — | — | — | — | — | 81 | |
| Confidence (Top-1)Generation Length (L)=512, Block Size (B)=64, Number of decoding blocks (K)=82026.06 | — | — | — | — | — | — | 81 | |
| EntropyGeneration Length (L)=128, Block Size (B)=128, Number of decoding blocks (K)=12026.06 | — | — | — | — | — | — | 60 | |
| EntropyGeneration Length (L)=128, Block Size (B)=32, Number of decoding blocks (K)=42026.06 | — | — | — | — | — | — | 74 | |
| EntropyGeneration Length (L)=128, Block Size (B)=16, Number of decoding blocks (K)=82026.06 | — | — | — | — | — | — | 70 | |
| EntropyGeneration Length (L)=256, Block Size (B)=256, Number of decoding blocks (K)=12026.06 | — | — | — | — | — | — | 62 | |
| EntropyGeneration Length (L)=256, Block Size (B)=64, Number of decoding blocks (K)=42026.06 | — | — | — | — | — | — | 79 | |
| EntropyGeneration Length (L)=256, Block Size (B)=32, Number of decoding blocks (K)=82026.06 | — | — | — | — | — | — | 79 | |
| EntropyGeneration Length (L)=512, Block Size (B)=512, Number of decoding blocks (K)=12026.06 | — | — | — | — | — | — | 79.2 | |
| EntropyGeneration Length (L)=512, Block Size (B)=128, Number of decoding blocks (K)=42026.06 | — | — | — | — | — | — | 79.8 | |
| EntropyGeneration Length (L)=512, Block Size (B)=64, Number of decoding blocks (K)=82026.06 | — | — | — | — | — | — | 80.2 |