Unverified50% confidenceTradeoffExact time
投机解码在代码生成等低熵任务中接受率可达70-85%，在创意写作等高熵任务中可能降至40-60%
1
Sources
50%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
小米静默发布MiMo-V2.5-DFlash：300B模型推理速度或将翻倍
redditr/LocalLLaMA7/12/2026
Related Claims
Unverified推测解码中当草稿通过率低于50%时，实际加速效果会大打折扣甚至产生负优化73% similarVerifiedThe cross-entropy loss in classification tasks is essentially the negative of the log-likelihood function, and minimizing cross-entropy is equivalent to maximizing the log-likelihood of the model on the training data.71% similarUnverified当不动动作占据80%以上帧时，模型通过最小化交叉熵损失可在不学习有意义特征的情况下达到80%准确率，这称为多数类捷径71% similarUnverified任务的输出熵越低（答案越可预测），推测性解码的加速效益越显著；GSM8K接受率从位置0的0.92衰减至0.53，MTBench从0.78急跌至0.1369% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/611058API
curl https://kongchang.com/api/v1/knowledge/claims/611058MCP
get_claim(id=611058)