Hallucination in Large Language Models: From Mechanistic Understanding to Novel Control Frameworks

Hallucination in large language models (LLMs) represents a fundamental challenge that extends beyond technical failure to reveal the mathematical inevitability of uncertainty in probabilistic language generation. This comprehensive survey presents the first unified mechanistic framework that systematically maps causal pathways from internal mechanisms to observable manifestations. We introduce a hierarchical taxonomy organizing hallucinations into four manifestation categories (P1: factuality errors, P2: faithfulness violations, P3: logical inconsistencies, P4: emergent behaviors) and trace their origins to three mechanistic sources (C1: data artifacts, C2: training biases, C3: inference failures). Our analysis reveals the dual nature of hallucination, simultaneously representing creative potential and factual unreliability, which challenges assumptions about complete elimination. We present seven novel frameworks for mechanism-aware hallucination management: HALO-M for layer-wise diagnostic analysis, KOQF for quantifying knowledge overshadowing effects, MIT for targeted mechanistic interventions, DPHCS for dual-process adaptive generation, HABS for comprehensive benchmarking, NSHPA for neuro-symbolic verification, and HACHMS for human-AI collaborative management. Through systematic task decomposition and compositional verification, these frameworks enable precise identification and mitigation of hallucination at its source. By establishing hallucination as an emergent property to be understood and managed rather than suppressed, we provide both theoretical foundations and practical tools for developing AI systems that balance creative capability with factual reliability, ultimately reframing the hallucination challenge from an obstacle to an opportunity for advancing artificial intelligence aligned with human values.