Towards Trustworthy AI: A Unified Framework for LLM Hallucination Detection and Mitigation via System Reliability Theory

Hallucination has been a fundamental issue for large language models (LLMs) and multimodal large language models (MLLMs) in areas where reliability and explainability are particularly valued. Current studies usually perceive the hallucination issue from two aspects: hallucination detection and mitigation. However, these two aspects are not independent in nature, and a more profound similarity exists between them. This paper surveys more than 100 recent studies on hallucination in the LLM and MLLM fields (2023-2026) and unifies the framework structure for hallucination from two perspectives: signal source and intervention time. Grounded on System Reliability Theory, by comparing methods including probing-based uncertainty estimation, multi-agent self-consistency, retrieval-based grounding, and verification/graph-based decoding control, we reveal the similarities and differences in these methods in terms of their costs, latency, and model accessibility. Ultimately, we present a novel and unifying point of view on hallucination mitigation, establishing a solid foundation towards trustworthy AI.