Speculative decoding — a small model drafts, a big model checks
Speculative decoding speeds generation by letting a small draft model propose tokens that a large model verifies in parallel, cutting latency with no change to the output distribution.
Not yet scoped.
Why
The PoC contrasts plain and speculative decoding on the same prompt and measures the latency win and its dependence on draft-model agreement.
How it works
Not yet built.
Where it lands in Jayverse
Number: if the research pipeline or the daily alice-tech report ever chains LLM calls, apply the draft-then-verify pattern. A fast cheap pass drafts, a slower pass checks, cutting latency without changing output quality.
Auditor: an anomaly or resolution-check pipeline that runs a fast flag model and a careful verify model is the same shape as speculative decoding. Measure the fast model's agreement rate before deciding it's worth running.
Key expressions
Words and phrases from this page worth keeping, with the Korean meaning and the sentence they come from.
Expression
뜻 · 쓰이는 자리
draft (tokens)
후보(초안)를 먼저 제시하다 · "a small draft model propose tokens"
verify in parallel
병렬로 검증하다 · "a large model verifies in parallel"
cut latency
지연 시간을 줄이다 · "cutting latency with no change to the output distribution"
no change to
~에 변화가 없는 · "no change to the output distribution"
dependence on
~에 대한 의존도 · "its dependence on draft-model agreement"