🤖 AI Summary
Hallucinations in large language and vision models critically undermine the reliability and safety of generative AI deployments, yet existing research lacks a unified, systematic understanding of their root causes. This project introduces the first cross-modal, multi-level hallucination analysis framework, jointly considering task- and modality-specific dimensions. It identifies hallucinations as arising from the synergistic interaction between data distribution shifts and inherited model biases, and characterizes their propagation mechanisms across the full model lifecycle—training, inference, and deployment. Leveraging hierarchical classification, cross-modal comparative analysis, and large-scale mechanistic attribution studies, we establish a unified theoretical model covering both textual and visual modalities. Our framework provides a generalizable foundation for hallucination attribution, enabling the design of robust, interpretable mitigation strategies and significantly enhancing the trustworthiness and generalization capability of generative AI systems.
📝 Abstract
The widespread adoption of large language and vision models in real-world applications has made urgent the need to address hallucinations -- instances where models produce incorrect or nonsensical outputs. These errors can propagate misinformation during deployment, leading to both financial and operational harm. Although much research has been devoted to mitigating hallucinations, our understanding of it is still incomplete and fragmented. Without a coherent understanding of hallucinations, proposed solutions risk mitigating surface symptoms rather than underlying causes, limiting their effectiveness and generalizability in deployment. To tackle this gap, we first present a unified, multi-level framework for characterizing both image and text hallucinations across diverse applications, aiming to reduce conceptual fragmentation. We then link these hallucinations to specific mechanisms within a model's lifecycle, using a task-modality interleaved approach to promote a more integrated understanding. Our investigations reveal that hallucinations often stem from predictable patterns in data distributions and inherited biases. By deepening our understanding, this survey provides a foundation for developing more robust and effective solutions to hallucinations in real-world generative AI systems.