Unverified50% confidenceBenchmarkExact time
接受率随位置逐级衰减：GSM8K从位置0的0.92下滑至约0.53，MTBench从0.78急跌至0.13，因误差在连续猜测中累积
1
Sources
50%
Confidence
Long-term
Relevance
7/9/2026
First Seen
Sources
DeepSeek DSpec Hands-On: How Speculative Decoding Accelerates LLM Inference by 6x
bilibili程序员-智能译站7/8/2026
Related Claims
Unverified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0.1383% similarUnverified研究表明，即使在 GSM8K 或 MATH 数据集上接近满分的模型，面对稍作变形或需要跨领域组合的问题时性能下降可达30-50个百分点70% similarUnverifiedThe optimal variant showed only a 1.83% decline on the GSM8K mathematical reasoning benchmark, while the MMLU general knowledge benchmark remained unchanged.70% similarUnverifiedThe released checkpoint was trained with a chunk size of 16; changing the sampling chunk size to 64 causes HumanEval to drop from 76% to approximately 20%, and GSM8K to drop from a high position to 2.2%.66% similarUnverifiedGitHub Copilot满意度已降到9%，远低于Claude Code的46%和Cursor的19%63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/362552API
curl https://kongchang.com/api/v1/knowledge/claims/362552MCP
get_claim(id=362552)