Unverified50% confidenceOpinionExact time
The usable context is far smaller than the nominal context; the model loses middle content, over-weights the most recently appearing tokens, and fabricates bridges between facts that are far apart.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe limitation of specialized small models is that their generalization ability rapidly degrades when task boundaries are ambiguous, cross-domain reasoning is required, or unseen edge cases need to be handled.79% similarUnverifiedWhen the context approaches its limit, the model may forget critical early information or produce lower-quality output.75% similarVerified研究表明当上下文过长时，模型对中间部分信息的注意力会显著下降，即"Lost in the Middle"现象75% similarUnverifiedCovariate shift and model drift can trigger negative transfer, causing additional data to drag down model performance.75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1029999API
curl https://kongchang.com/api/v1/knowledge/claims/1029999MCP
get_claim(id=1029999)