Inference Infrastructure & ServingGoogle (Leviathan et al.)Google

Speculative Decoding

Speculative Decoding

Why it matters

Sidesteps the memory-bound nature of decoding without changing the output distribution

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2211.17192

Added in Week 35, 2026 · See that week

Same week

OpenAI releases ChatGPT as a research previewOfficial

Nearby on the same axis

Accelerating LLM Decoding with Speculative SamplingPaperGeorgi Gerganov creates the llama.cpp repositoryOfficialOrcaOfficialAWQPaper