Expired70% confidenceFactExact time
LLMs exhibit an 'Instruction Drift' phenomenon: as the number of conversation turns increases, early system prompts gradually decay in the model's attention weights.
4
Sources
70%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 9/30/2026(expired)
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% similarUnverifiedCurrent large models do not have a true persistent memory mechanism, and as the number of conversation turns increases, the weight of early requirement descriptions in the attention mechanism gradually decays.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% similarUnverifiedLLMs maintain higher sensitivity to information at the beginning and end of the context, while exhibiting an 'attention decay' phenomenon for information in the middle positions78% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/52118API
curl https://kongchang.com/api/v1/knowledge/claims/52118MCP
get_claim(id=52118)