Expired50% confidenceFactExact time
Overly long system instructions in LLMs lead to 'attention decay' on critical rules due to non-uniform attention distribution across the context window.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026(expired)
Sources
Three Essential Claude Code Configuration Techniques: A Practical Guide to CLAUDE.md, Memory & MCP
bilibiliAI极客玩家6/19/2026
Related Claims
VerifiedLLMs exhibit an 'Instruction Drift' phenomenon: as the number of conversation turns increases, early system prompts gradually decay in the model's attention weights.80% similarUnverifiedLLM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78% similarUnverified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76% similarUnverifiedOverly verbose, contradictory, or ambiguous system prompts cause the model's attention to become scattered, resulting in over-thinking or instruction-following failures; streamlining agents.md can improve the determinism of instruction following.74% 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 positions73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/43296API
curl https://kongchang.com/api/v1/knowledge/claims/43296MCP
get_claim(id=43296)