Context Poisoning as Extreme-Value Attention Interference in Long-Context Language Models
Under a softmax retrieval abstraction, a finite-sample upper bound is derived showing that maintaining a fixed accuracy target above base rate requires the evidence margin to scale as $\Omega(\sqrt{\log N})$, where N denotes the effective distractor count rather than necessarily the raw context length.
Meysam Ghaffari, Nina Fatehi, Bhaskar Sen et al.
· 0 citations