List your company

Multi-path speculative decoding on Llama (held-out test)

1.21Throughput Ratio ImprovementNaiveTree NDE
0.951.01751.0851.1525Feb 19, 2026

Evaluation results

MethodLinks
NaiveTree NDEVerification algorithm=NaiveTree, Neural Dynamic Expansion=true2026.021.21
Khisti NDEVerification algorithm=Khisti, Neural Dynamic Expansion=true2026.020.97
NSS NDEVerification algorithm=NSS, Neural Dynamic Expansion=true2026.020.96
SpecInfer NDEVerification algorithm=SpecInfer, Neural Dynamic Expansion=true2026.020.96
SpecTr NDEVerification algorithm=SpecTr, Neural Dynamic Expansion=true2026.020.96