Context Poisoning as Extreme-Value Attention Interference in Long-Context Language Models
Source
Meysam Ghaffari, Nina Fatehi, Bhaskar Sen, Nasim Sabetpour, Carlos Morato
Author
Meysam Ghaffari, Nina Fatehi, Bhaskar Sen, Nasim Sabetpour, Carlos Morato
Date
Key takeaways · AI-distilled
The paper models context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → poisoning as extreme-value interference: the decisive evidence's attentionThe mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.Full definition → score is bounded, while the maximum score among distractors keeps growing as more are added.
The N that matters is the effective count of confusable distractors, not raw context length, and the analysis links degradation to score aliasing, positional aliasing and softmax dilution.
In controlled tests, distractors in the same format as the evidence caused the largest accuracy drop among the tested distractor types at fixed context length.
Retrieval gating improved evidence use, but its net benefit depended on preserving evidence recall; the authors point to evidence bottlenecks, retrieve-then-reason designs and verifier-mediated memory as mitigations.
Terms in this piece · Glossary
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
Why it matters
Gives a theoretical and empirical account of why retrieval accuracy drops as context fills with hard negatives, a concrete failure mode to budget for when stuffing tool outputs or documents into an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition →'s context.
articleWhen Irrelevant Text Matters: Affine Margin Shifts in Multimodal Large Language ModelsYinfeng Wang, Zhiyuan Yao, Zheren Fu, Lei Zhang, Zhendong Mao
articleEvidence Integration in Large Language ModelsSebastien Kawada, Manolis Kellis
articleDo LLMs Make More Mistakes If They Do Not Believe the Input Data?Peter Kochelka, Ale\v{s} Manuel Pap\'a\v{c}ek, Vojt\v{e}ch Dvo\v{r}\'ak, Ond\v{r}ej Du\v{s}ek