Year: 2026
Type: conference-paper
Abstract: Long-context language models assemble prompts from heterogeneous sources, and every deployed system trusts the model to use the right span.We show that this trust is routinely violated.Irrelevant spans silently shape outputs and produce errors that are neither fabrication nor omission but misattributed grounding, a claim supported by the wrong part of the input.Unlike intrinsic hallucination, which contradicts the source, or extrinsic hallucination, which introduces unsupported claims, misattributed grounding uses real evidence from the wrong source, and standard faithfulness metrics are structurally unable to detect it.We call this phenomenon Ghost Context and present an empirical study of its prevalence, structure, and remediation.We evaluate three widely used models across a 272-case corpus using a mask-and-rerun causal attribution protocol with dual-judge calibration.We measure two complementary signals.Strict Ghost Context Rate (GCR) captures verifiable factual misattribution.Open-ended influence captures any detectable shaping, including nonfactual shifts.Under realistic contextual conflict, strict GCR spikes.Temporal contradictions trigger misattributed grounding in 38.3% of cases.Open-ended influence reaches 20.4% across all cases.Crucially, Ghost Context is not only detectable but correctable.Masking the single highest-attributed distractor span resolves 95.5% of detected errors (Fix@1), with collateral damage of only 2.4% and zero false positives on negative controls.We release the full corpus, pipeline, and results.
Author: Rohith Namboothiri
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166851221