List your company
ResearchFormerly WizWand

Mathematical Reasoning on GSM8K (Accuracy, Speed)

98Accuracy (GSM8K)No Defense
70.9677.988592.02Jun 18, 2026
Updated 4mo ago

Evaluation results

MethodLinks
No DefenseModel=Qwen3-32B (Draft-1.7B)2026.06981
No DefenseModel=DeepSeek R1-70B (Draft-8B)2026.06981
Spec-DecodingModel=Qwen3-32B (Draft-1.7B)2026.06971.38
SRModel=Qwen3-32B (Draft-1.7B)2026.06971.07
RPOModel=Qwen3-32B (Draft-1.7B)2026.06971.16
SafeDecodingModel=DeepSeek R1-70B (Draft-8B)2026.06970.78
SafeDecodingModel=Qwen3-32B (Draft-1.7B)2026.06960.82
SafeSpecModel=Qwen3-32B (Draft-1.7B)2026.06962.06
SRModel=DeepSeek R1-70B (Draft-8B)2026.06961.09
SecDecodingModel=Qwen3-32B (Draft-1.7B)2026.06950.61
RPOModel=DeepSeek R1-70B (Draft-8B)2026.06951.12
Spec-DecodingModel=DeepSeek R1-70B (Draft-8B)2026.06941.7
SecDecodingModel=DeepSeek R1-70B (Draft-8B)2026.06940.57
SpecreasonModel=Qwen3-32B (Draft-1.7B)2026.06932.14
Specreason+SRModel=Qwen3-32B (Draft-1.7B)2026.06932.18
Specreason+RPOModel=Qwen3-32B (Draft-1.7B)2026.06932.21
SpecreasonModel=DeepSeek R1-70B (Draft-8B)2026.06901.81
Specreason+SRModel=DeepSeek R1-70B (Draft-8B)2026.06901.94
Specreason+RPOModel=DeepSeek R1-70B (Draft-8B)2026.06901.98
SafeSpecModel=DeepSeek R1-70B (Draft-8B)2026.06901.76
Draft ModelModel=Qwen3-32B (Draft-1.7B)2026.06866.42
Draft ModelModel=DeepSeek R1-70B (Draft-8B)2026.06724.53