Unverified50% confidenceBenchmarkExact time
研究发现位于上下文首尾的信息被模型召回的概率远高于中间位置，衰减在上下文长度超过4K token后急剧加速
1
Sources
50%
Confidence
Long-term
Relevance
7/20/2026
First Seen
Sources
Related Claims
Unverified研究发现Lost in the Middle衰减在上下文长度超过4K token后急剧加速76% similarVerifiedA Stanford 2023 research experiment demonstrated that models have a significantly higher forgetting rate for information located in the middle of the context than for content at the beginning and end.74% similarUnverifiedThe quality of long-context understanding in small models degrades rapidly with length, typically deteriorating after exceeding a 4K–8K token window.72% similarUnverifiedLost in the Middle 效应指模型对位于极长输入序列中间位置的信息关注度会系统性下降，在超过 20K token 时开始显著，100K token 级别尤为突出71% similarUnverified斯坦福大学2023年研究显示当关键信息位于超长文档中间段落时模型检索准确率显著低于位于开头或结尾70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/566480API
curl https://kongchang.com/api/v1/knowledge/claims/566480MCP
get_claim(id=566480)