Unverified50% confidenceFactExact time
The extrapolation target takes the form p_S + β(p_T - p_S) (β>1), may introduce larger variance, and needs to be used with clipping or KL constraints.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn speculative decoding, if the draft model's per-position accuracy is p, the expected number of tokens accepted when generating k candidate tokens is p*(1-p^k)/(1-p).60% similarUnverified典型的导纳模型可表达为 M·ẍ + B·ẋ + K·x = F_ext，其中M、B、K分别是期望的虚拟惯量、阻尼和刚度矩阵59% similarUnverifiedIn token-level sequence generation tasks, the KL constraint is computed token by token, so tokens at different positions are subject to different constraint strengths.59% similarUnverifiedDP-μP proposes a spectral scaling formula that combines the SNR condition with matrix spectral properties to give concrete scaling rules for different optimizers.58% similarUnverifiedTensorRT-LLM、vLLM、llama.cpp等不同推理后端在KV Cache管理、注意力计算kernel实现、采样策略上各有不同，可能导致相同prompt在不同环境产生不同输出58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1074090API
curl https://kongchang.com/api/v1/knowledge/claims/1074090MCP
get_claim(id=1074090)