Tag
This paper introduces 'context poisoning' as a phenomenon where attention interference degrades performance in long-context language models, deriving bounds showing evidence margin must scale with log of distractor count and proposing mitigations like evidence bottlenecks.