Beyond Isolated Fixes: A Comprehensive Survey on Hallucination Mitigation with a Three-Dimensional Taxonomy and Integrative Framework

Hallucinations in large language and vision–language models constitute a persistent and systemic failure mode that fundamentally undermines their reliability in high-stakes applications. Rather than framing hallucination as a localized generation error, this paper reconceptualizes it as a lifecycle phenomenon—emerging from upstream deficiencies in data collection, reinforced by training-time biases, and exacerbated by inference-time decoding pathologies. To address this challenge, we propose a unified framework for hallucination mitigation grounded in a structured three-dimensional taxonomy that spans mechanism-based strategies, phase-specific interventions, and cross-phase integrative approaches. This taxonomy elucidates how technical solutions, knowledge-enhanced methods, and framework-level designs can be systematically orchestrated across pre-, in-, and post-generation stages to support more robust, interpretable, and auditable model behavior. Within this paradigm, detection is positioned not as a terminal task, but as a diagnostic backbone that enables adaptive and multi-stage mitigation workflows. Building on this foundation, we introduce an integrative hybrid framework that consolidates previously fragmented techniques into coherent pipelines, demonstrating how cross-phase synergy and iterative refinement can enhance reliability in real-world deployment. This framework further reveals latent connections across diverse strands of research, facilitating the identification of underexplored intersections and the development of synergistic method combinations. By reframing hallucination mitigation as a system-level challenge rather than a set of isolated fixes, this work offers both a theoretical lens and a practical roadmap for advancing the trustworthiness of generative AI systems.