Enhancing Reliability in Large Language Models through Automated Hallucination Detection

Abstract

Large language models (LLMs) are increasingly used in settings where unsupported or fabricated statements can create material risk. This paper presents a hybrid hallucination detector that combines evidence retrieval, textual entailment, generation uncertainty, and lightweight linguistic anomaly features to identify unreliable model outputs before they are shown to end users. We evaluate the detector on an annotated benchmark of 18{,}400 generations spanning news summarization, open-domain question answering, and biomedical assistance.
The proposed system achieves a precision of 0.92, recall of 0.88, and F1 of 0.90 on the held-out test set, outperforming confidence-only, retrieval-only, and NLI-only baselines by 9--22 absolute F1 points. The gains are consistent across all three task families, with the largest improvement observed in biomedical assistance, where unsupported entity and dosage claims are particularly common.
We further study an operational deployment setting in which the detector acts as a gating module. Under an abstain-or-regenerate policy, the rate of unsupported responses falls from 22.8\% to 11.2\% while preserving 92.1\% response coverage. These results show that automated hallucination detection can substantially improve the reliability of LLM systems without incurring prohibitive latency, providing a practical path toward safer deployment in high-stakes domains.