Researchers combine instruction-based refusal with a structural gate to reduce hallucinations in large language models by detecting unsupported claims at the output boundary.

Study frames hallucinations as a misclassification error where internally generated text is presented as if grounded in evidence

Proposed composite intervention combines instruction-based refusal with an abstention gate that computes a support deficit score from self-consistency, paraphrase stability, and citation coverage

Testing across 50 items, five epistemic regimes, and three models shows neither mechanism alone is sufficient—instruction-only prompting still exhibits over-cautious abstention and residual hallucination in GPT-3.5-turbo

Structural gate preserves accuracy on answerable questions but fails to catch confident confabulation when items contain conflicting evidence