Unverified50% confidenceFactExact time
指令衰减（Instruction Decay）是指当上下文窗口逐渐被填满时，模型对早期系统指令或用户要求的遵循程度明显下滑
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
Unverified当LLM上下文窗口中的指令过多时，模型对每条指令的关注度会下降，导致遵循率降低73% similarUnverifiedIn long contexts, models are more easily influenced by the most recent few turns of conversation rather than the system prompt set at the beginning, a phenomenon known as instruction-following decay.68% similarUnverifiedExperiments show that after only a few generations of self-training, the variance of model outputs drops sharply and the text tends toward homogeneity68% similarUnverifiedCurrent model instruction-following failures commonly manifest as ignoring system prompt constraints, forgetting early instructions in long conversations, randomly violating formatted output requirements, and deviating from predefined processes in multi-step tasks.67% similarUnverified自训练循环存在模型坍缩、错误放大和安全评估时间窗口压缩等潜在风险67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/530271API
curl https://kongchang.com/api/v1/knowledge/claims/530271MCP
get_claim(id=530271)