Verified70% confidenceFactExact time
LLM存在'Instruction Drift'现象：随着对话轮次增加，早期系统提示在模型注意力权重中逐渐衰减
4
Sources
70%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026
Sources
Claude Code Hooks: A System-Level Safety Lock for AI Programming
bilibiliAli厂长5/23/2026
Related Claims
UnverifiedLLM存在上下文窗口限制，在超长对话中面临注意力稀释问题，早期决策记录会被赋予更低注意力权重80% similarUnverified当前大模型并没有真正的持久化记忆机制，随着对话轮次增加，早期需求描述在注意力机制中的权重会逐渐衰减80% similarUnverifiedOverly long system instructions in LLMs lead to 'attention decay' on critical rules due to non-uniform attention distribution across the context window.80% similarUnverifiedLLM对上下文开头和结尾的信息保持较高敏感度，而对中间位置的信息存在'注意力衰减'现象78% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/52118API
curl https://kongchang.com/api/v1/knowledge/claims/52118MCP
get_claim(id=52118)