Ghost Context: Cross-Context Hallucination in Long-Context Language Models

As context windows have grown, a class of failures has emerged in which the model answers from material the user never put in front of it, usually a sibling document or an earlier turn, without acknowledging the borrow.

We characterize ghost context across long-context models, identify structural conditions that make it likely, including genre overlap and formatting similarity, and propose retrieval-time and prompt-time mitigations.

Accepted at ACL TrustNLP 2026.