List your company

Mathematical Reasoning on GSM8K (Accuracy, ASR)

93AccuracyAutoregressive
13.02433.78754.5575.313Jan 14, 2025Apr 10, 2025Jul 5, 2025Sep 30, 2025Dec 25, 2025Mar 21, 2026Jun 16, 2026

Evaluation results

MethodLinks
AutoregressiveTarget Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=0.1, Draft Length (N)=52026.0493——
SpecCascadeTarget Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=0.1, Draft Length (N)=52026.0493—83.5
AutoregressiveTarget/Draft model pair=Gemma-3-12B/Gemma-3-4B, Temperature=1, Draft length (N)=52026.0493——
AutoregressiveTarget Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=1, Draft Length (N)=52026.0493——
SDTarget Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=0.1, Draft Length (N)=52026.0492—81.85
SD (Lossy)Target Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=0.1, Draft Length (N)=52026.0492—82.75
DiversedTarget Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=0.1, Draft Length (N)=52026.0492—90.26
SDTarget/Draft model pair=Gemma-3-12B/Gemma-3-4B, Temperature=1, Draft length (N)=52026.0492—84.15
SpecCascadeTarget/Draft model pair=Gemma-3-12B/Gemma-3-4B, Temperature=1, Draft length (N)=52026.0492—84.43
DIVERSEDTarget/Draft model pair=Gemma-3-12B/Gemma-3-4B, Temperature=1, Draft length (N)=52026.0492—90.7
SDTarget Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=1, Draft Length (N)=52026.0492—84.15
SpecCascadeTarget Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=1, Draft Length (N)=52026.0492—84.43
DIVERSEDTarget Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=1, Draft Length (N)=52026.0492—90.7
Static EnsembleTarget Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=0.1, Draft Length (N)=52026.0491—89.67
SDTarget/Draft model pair=Qwen3-8B/Qwen3-0.6B, Temperature=1, Draft length (N)=52026.0491—59.58
Static EnsembleTarget/Draft model pair=Gemma-3-12B/Gemma-3-4B, Temperature=1, Draft length (N)=52026.0491—87.62
Static EnsembleTarget Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=1, Draft Length (N)=52026.0491—87.62
AutoregressiveTarget/Draft model pair=Qwen3-8B/Qwen3-0.6B, Temperature=1, Draft length (N)=52026.0490——
SD (Lossy)Target/Draft model pair=Gemma-3-12B/Gemma-3-4B, Temperature=1, Draft length (N)=52026.0490—85.02
SD (Lossy)Target Model=Gemma-3-12B, Draft Model=Gemma-3-4B, Temperature=1, Draft Length (N)=52026.0490—85.02
SCASModel=Qwen2.5-7B2026.0589.38——
DIVERSEDTarget/Draft model pair=Qwen3-8B/Qwen3-0.6B, Temperature=1, Draft length (N)=52026.0488—76.48
GRACEModel=Qwen2.5-7B2026.0586.85——
SD (lossy)Target/Draft model pair=Qwen3-8B/Qwen3-0.6B, Temperature=1, Draft length (N)=52026.0486—60.1
SpecCascadeTarget/Draft model pair=Qwen3-8B/Qwen3-0.6B, Temperature=1, Draft length (N)=52026.0486—58.65
Static EnsembleTarget/Draft model pair=Qwen3-8B/Qwen3-0.6B, Temperature=1, Draft length (N)=52026.0486—67.58
HY-1.8B-FP16Model Scale=1.8B, Precision=FP16, Instruction-tuned=true2026.0284.08——
SCASModel=Qwen2.5-3B2026.0582.82——
Llama3.1Condition=Clean2025.0182.25——
Llama3.1Punctuation Percentage=10%2025.0181.04——
Llama3.1Punctuation Percentage=30%2025.0178.84——
HY-1.8B-INT4Model Scale=1.8B, Precision=INT4, Instruction-tuned=true2026.0278.7——
VoidPaddingBlock size B=5122026.0678.62——
VoidPaddingBlock size B=1282026.0678.39——
VoidPaddingBlock size B=Mean2026.0678.27——
VoidPaddingBlock size B=642026.0677.79——
MathstralCondition=Clean2025.0177.63——
HY-1.8B-2BitModel Scale=1.8B, Precision=2Bit, Instruction-tuned=true2026.0277.33——
VIGORBackbone=Qwen2.5-3B-Base, Training Dataset=CodeContests, RL Strategy=VIGOR2026.0577.1——
Llama3.1Punctuation Percentage=50%2025.0177.02——
GRACEModel=Qwen2.5-3B2026.0576.8——
OriginalBlock size B=642026.0676.8——
MathstralPunctuation Percentage=10%2025.0175.51——
Llama3Condition=Clean2025.0175.43——
INTUITORBackbone=Qwen2.5-3B-Base, Training Dataset=CodeContests, RL Strategy=INTUITOR2026.0575.13——
RainbowPaddingBlock size B=642026.0674.53——
DeepSeekCondition=Clean2025.0173.76——
DeepSeekPunctuation Percentage=10%2025.0173.76——
RainbowPaddingBlock size B=1282026.0673.39——
Llama3Punctuation Percentage=10%2025.0173.31——
Llama3Punctuation Percentage=30%2025.0173.08——
Llama3Punctuation Percentage=50%2025.0172.17——
RainbowPaddingBlock size B=Mean2026.0671.5——
MathstralPunctuation Percentage=30%2025.0171.34——
OriginalBlock size B=1282026.0670.96——
MathstralPunctuation Percentage=50%2025.0170.65——
DeepSeekPunctuation Percentage=30%2025.0170.43——
Before RL (Base)Backbone=Qwen2.5-3B-Base, Training Dataset=CodeContests, RL Strategy=None2026.0567.93——
DeepSeekPunctuation Percentage=50%2025.0167.24——
AutoregressiveTarget/Draft model pair=Llama-3.1-8B/Llama-3.2-1B, Temperature=1, Draft length (N)=52026.0467——
SDTarget/Draft model pair=Llama-3.1-8B/Llama-3.2-1B, Temperature=1, Draft length (N)=52026.0467—44.6
SpecCascadeTarget/Draft model pair=Llama-3.1-8B/Llama-3.2-1B, Temperature=1, Draft length (N)=52026.0467—61.53
DIVERSEDTarget/Draft model pair=Llama-3.1-8B/Llama-3.2-1B, Temperature=1, Draft length (N)=52026.0467—72.61
RainbowPaddingBlock size B=5122026.0666.57——
SD (Lossy)Target/Draft model pair=Llama-3.1-8B/Llama-3.2-1B, Temperature=1, Draft length (N)=52026.0466—59.81
Static EnsembleTarget/Draft model pair=Llama-3.1-8B/Llama-3.2-1B, Temperature=1, Draft length (N)=52026.0466—69.49
RSRModel=Qwen2.5-7B2026.0562.55——
Qwen2.5Condition=Clean2025.0161.1——
RSRModel=Qwen2.5-3B2026.0559.92——
OriginalBlock size B=Mean2026.0658.73——
Qwen2.5Punctuation Percentage=10%2025.0156.02——
RSR-ItemModel=Qwen2.5-7B2026.0555.12——
HY-0.5B-FP16Model Scale=0.5B, Precision=FP16, Instruction-tuned=true2026.0255.04——
Qwen2.5Punctuation Percentage=30%2025.0152.69——
Gemma2Condition=Clean2025.0149.65——
Qwen2.5Punctuation Percentage=50%2025.0149.35——
Gemma2Punctuation Percentage=10%2025.0145.1——
SCASModel=Qwen2.5-0.5B2026.0545.03——
MistralCondition=Clean2025.0142.07——
MistralPunctuation Percentage=10%2025.0141.62——
MistralPunctuation Percentage=30%2025.0137.75——
GRACEModel=Qwen2.5-0.5B2026.0537.65——
Gemma2Punctuation Percentage=30%2025.0136.46——
MistralPunctuation Percentage=50%2025.0136.39——
Gemma2Punctuation Percentage=50%2025.0135.63——
RSR-ItemModel=Qwen2.5-3B2026.0534.95——
OriginalBlock size B=5122026.0628.43——
RSRModel=Qwen2.5-0.5B2026.0526.26——
InfluenceModel=Qwen2.5-3B2026.0525.22——
ZephyrCondition=Clean2025.0123.27——
RouterModel=Qwen2.5-3B2026.0523.2——
RouterModel=Qwen2.5-7B2026.0521.99——
PPLModel=Qwen2.5-7B2026.0520.6——
IFDModel=Qwen2.5-7B2026.0520.32——
InfluenceModel=Qwen2.5-7B2026.0519.61——
PPLModel=Qwen2.5-3B2026.0518.88——
IFDModel=Qwen2.5-3B2026.0518.32——
ZephyrPunctuation Percentage=10%2025.0118.04——
ZephyrPunctuation Percentage=30%2025.0118.04——
RSR-ItemModel=Qwen2.5-0.5B2026.0516.1——