Unverified50% confidenceFactExact time
大语言模型存在「指令漂移」（instruction drift）现象，随着对话轮次增加和上下文窗口被大量代码填充，模型对早期指令的遵循概率会逐步下降
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026
Sources
Claude Code Hooks 完全指南：确定性自动化执行配置详解
bilibilibili_brook6/7/2026
Related Claims
Verified大语言模型在长时间对话中会因上下文窗口限制而丢失早期信息80% similarUnverified大语言模型在长文本生成中出现的话题漂移现象在学术上被称为语义漂移（Semantic Drift），根本原因在于自回归生成机制79% similarUnverifiedCode drift in pure prompt-based development is rooted in the statelessness of large language models, which re-infer system intent with each new prompt77% similarUnverified当前主流大语言模型上下文窗口已扩展至数十万乃至百万tokens，但超长上下文下模型对早期信息的注意力权重会显著衰减（中间丢失问题）75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/41406API
curl https://kongchang.com/api/v1/knowledge/claims/41406MCP
get_claim(id=41406)