Recent advances in large language models (LLMs) have led to the emergence of multimodal LLMs capable of jointly interpreting and generating information from both visual and linguistic inputs. Representative models such as GPT-4o, Gemini, Claude, and L...
Recent advances in large language models (LLMs) have led to the emergence of multimodal LLMs capable of jointly interpreting and generating information from both visual and linguistic inputs. Representative models such as GPT-4o, Gemini, Claude, and LLaVA demonstrate high performance in complex visual question answering, explanation generation, and creative reasoning tasks, thereby expanding their applicability across domains such as education, healthcare, law, and design. Despite these advancements, multimodal LLMs frequently generate responses that appear plausible yet are factually incorrect, a phenomenon commonly referred to as hallucination. This issue continues to pose a fundamental challenge to the reliability of AI systems.
In multimodal settings, hallucination errors cannot be sufficiently captured through simple correctness judgments. Errors often involve complex interactions among factors such as text–image inconsistency, overgeneralization of visual attributes, misinterpretation of user intent, and selective information bias. Prior studies have focused predominantly on hallucinations in text-only LLMs and have lacked systematic analyses of how hallucinations emerge and differ under multimodal conditions. In addition, the absence of a structured error taxonomy and the limited integration of quantitative and qualitative evaluation frameworks have constrained progress in building reliable multimodal AI systems.
To address these gaps, this study conducts a structured analysis of hallucination errors in state-of-the-art multimodal LLMs under three input conditions: text-only, image-only, and text+image. The proposed framework combines automated evaluation metrics (e.g., GPTScore, CLIPScore) with expert-based qualitative assessment. Furthermore, this study introduces a multimodal hallucination error taxonomy consisting of four core error types—factual error, interpretive error, exaggeration/distortion, and logical inconsistency—supported by a codebook and evaluation protocol. Representative response analyses are used to interpret underlying error mechanisms, including processing pathways, fusion failures, and interaction constraints between input modalities and model structure.
The results demonstrate distinct error patterns depending on input modality: text-only conditions predominantly yielded reasoning and interpretive errors, image-only conditions revealed failures in fine-grained attribute perception and contextual overgeneralization, while text+image conditions exhibited pronounced cross-modal alignment failures and information integration omissions. Notably, hallucination severity varied even under identical prompts, depending on input configuration. Model-wise comparison indicated that GPT-4o and Gemini produced more stable responses overall, whereas models such as LLaVA displayed tendencies to over-rely on visual cues or misinterpret query focus.
This study contributes in four significant ways. First, it reframes hallucination in multimodal LLMs not as a surface-level correctness issue, but as a structural phenomenon rooted in information processing and fusion mechanisms. Second, it establishes a reusable error classification and evaluation framework for future multimodal AI research. Third, it proposes targeted improvement directions for multimodal model design, including enhanced cross-modal integration, grounding stabilization, and attribute reasoning refinement. Fourth, it demonstrates the value of a combined quantitative–qualitative analysis pipeline, enabling both large-scale evaluation and in-depth interpretive insight.
Overall, the findings of this study provide foundational guidance for improving the reliability, transparency, and safety of multimodal LLM systems. As AI becomes increasingly embedded in everyday decision-making and collaborative contexts, understanding the origins and mechanisms of hallucination will be essential to developing models that are not only capable but trustworthy.