Unverified50% confidenceFactTime unknown

LLM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

1

Sources

50%

Confidence

Medium-term (~90 days)

Relevance

7/2/2026

First Seen

Valid until: 9/30/2026

Sources

Single-Agent vs Multi-Agent: A Three-Step Decision Framework for Choosing the Right Architecture

bilibili慢学AI

Related Claims

Unverified研究表明当上下文窗口填满时，LLM对序列首尾信息的关注度显著高于中间部分，即'迷失在中间'效应81% similar 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% similar VerifiedLLM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78% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/56454

API

curl https://kongchang.com/api/v1/knowledge/claims/56454

MCP

get_claim(id=56454)