Definition
The tendency of Large Language Models to disproportionately favor information located at the end of the context window or, in RAG architectures, the prioritization of chronologically recent documents at the expense of historical accuracy. This often requires architectural trade-offs between 'freshness' and 'comprehensive context' to prevent the neglect of foundational data.
Not a social psychology concept, but a positional token weighting or metadata filtering priority in AI models.
"A conveyor belt where only the items closest to the exit are picked up, while vital components further back are ignored."
- Lost in the Middle(Correlated phenomenon where middle-context data is deprioritized)
- Positional Encoding(Technical driver within Transformer architectures)
- Time-weighted Vector Store(Mitigation strategy to balance relevance and age)
Conceptual Overview
The tendency of Large Language Models to disproportionately favor information located at the end of the context window or, in RAG architectures, the prioritization of chronologically recent documents at the expense of historical accuracy. This often requires architectural trade-offs between 'freshness' and 'comprehensive context' to prevent the neglect of foundational data.
Disambiguation
Not a social psychology concept, but a positional token weighting or metadata filtering priority in AI models.
Visual Analog
A conveyor belt where only the items closest to the exit are picked up, while vital components further back are ignored.