Unverified50% confidenceFactExact time
Speculative decoding was proposed by Google DeepMind; its mathematical foundation is built on rejection sampling, where tokens from the draft model are accepted with a probability of min(1, p_large(x)/p_draft(x)), strictly guaranteeing that the output distribution is consistent with independent generation by the large model.
1
Sources
50%
Confidence
Long-term
Relevance
8/11/2026
First Seen
Sources
Related Claims
Unverified自回归大模型推理瓶颈并非计算量而是内存带宽，推测解码通过拒绝采样保证输出分布与原始大模型完全一致70% similarUnverifiedResearch from institutions such as Google DeepMind and MIT shows that multiple large language models can significantly reduce hallucinations and reasoning errors of individual models through debate70% similarUnverifiedSpeculative decoding generates drafts with a small model and verifies them in batch with a large model; mathematically it guarantees output fully consistent with the large model through rejection sampling, making it lossless acceleration69% similarUnverifiedSpeculative Decoding reduces computation by using a small model to predict the output tokens of a large model68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/729568API
curl https://kongchang.com/api/v1/knowledge/claims/729568MCP
get_claim(id=729568)