Probing-Informed Reliability Gates: Multilingual Hallucination Detection and Mitigation in Large Language Models

Large language models can produce fluent but unsupported answers. This paper studies whether probing features can predict hallucination risk and inform a reliability gate that routes uncertain answers through retrieval before responding. We evaluate multilingual question answering and dialogue tasks in high- and low-resource languages. Probe-informed classifiers outperform confidence baselines for detecting hallucinations, and a gate that uses these signals reduces unsupported answers while preserving accuracy. Results reveal cross-lingual differences in hallucination rates and show how probing can move beyond analysis to support trustworthy conversational AI. The paper includes a system diagram, precision–recall analyses, multilingual error rates, ablations, and a latency–accuracy trade-off, along with an appendix covering datasets, hyperparameters, error cases, and reproducibility details.