List your company
ResearchFormerly WizWand

Mathematical Reasoning on GSM8K 4-shot (test)

54.6ThroughputFast-dLLM (Par.+D.C.)
0.540814.575428.6142.6446Sep 27, 2025Oct 6, 2025Oct 16, 2025Oct 26, 2025Nov 4, 2025Nov 14, 2025Nov 24, 2025
Updated 8mo ago

Evaluation results

MethodLinks
Fast-dLLM (Par.+D.C.)Shots=4-shot, parallel decoding=true, dual-cache KV=true2025.1154.6—4.276.585230.4
CDLM-LLaDAShots=4-shot2025.1154.3—3.373.957.7177.3
d2CacheModel=Dream-Inst, Few-shot=4-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.0946.6917.86.1375.7——
d2CacheModel=LLaDA-Inst, Few-shot=4-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.0938.1613.87.2676.9——
dLLM-CacheModel=Dream-Inst, Few-shot=4-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.093613.78.2174.3——
Fast-dLLMModel=Dream-Inst, Few-shot=4-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.0932.7512.58.3574.1——
Fast-dLLM (Par.)Shots=4-shot, parallel decoding=true2025.1126.7—8.778.677.5231.3
dLLM-CacheModel=LLaDA-Inst, Few-shot=4-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.0925.629.210.2677——
Fast-dLLMModel=LLaDA-Inst, Few-shot=4-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.0925.159.110.6577.6——
dLLM-CacheShots=4-shot, cache_mode=dLLM-Cache2025.1118.7—12.378.5256229.6
ParallelModel=Dream-Inst, Few-shot=4-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.0913.935.321.6874.2——
ParallelModel=LLaDA-Inst, Few-shot=4-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.098.533.133.9577.6——
LLaDA-8B-InstructShots=4-shot2025.118.2—28.377.1256232.1
VanillaModel=LLaDA-Inst, Few-shot=4-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.092.771110.2677.6——
VanillaModel=Dream-Inst, Few-shot=4-shot, Gen. Len.=256, Decoding Strategy=Semi-AR parallel decoding2025.092.62185.9476.7——