List your company

Mathematical Reasoning on GSM8K (test) (Accuracy, Throughput, Speedup)

5.6Relative SpeedupWindow-Diffusion
0.8162.0583.34.542Jan 28, 2026Feb 8, 2026Feb 20, 2026Mar 3, 2026Mar 15, 2026Mar 26, 2026Apr 7, 2026

Evaluation results

MethodLinks
Window-DiffusionModel=LLaDA-Base, #Shots=4, Max generation length=256, external window length=64, internal window length=16, cache refresh interval=32, early stopping=disabled2026.015.668.525.14
Fast-dLLM (Dual-Cache)Model=LLaDA-Base, #Shots=4, Max generation length=256, block size=32, parallel decoding=disabled2026.013.8567.917.28
Fast-dLLM (Prefix-Cache)Model=LLaDA-Base, #Shots=4, Max generation length=256, block size=32, parallel decoding=disabled2026.013.2868.814.71
Drafter5Drafter Specialization=5: Math2026.042.454——
MetaSpS-UCBBandit Selection=UCB2026.042.28——
MetaSpS-SHBandit Selection=SH2026.042.144——
DKV-CacheModel=LLaDA-Base, #Shots=4, Max generation length=256, cache refresh interval=8, block size=322026.011.868.78.07
MetaSpS-EXP3Bandit Selection=EXP32026.041.69——
PLDMethod Type=Other2026.041.653——
OFADrafter Type=One-Size-Fits-All2026.041.571——
LookaheadMethod Type=Other2026.041.533——
Drafter4Drafter Specialization=4: QA2026.041.486——
Drafter1Drafter Specialization=1: Code2026.041.483——
MetaSpS-RandBandit Selection=Rand2026.041.471——
Drafter3Drafter Specialization=3: Summarization2026.041.378——
Drafter2Drafter Specialization=2: Translation2026.041.228——
LLaDAModel=LLaDA-Base, #Shots=4, Max generation length=2562026.01170.74.49