When Text Hijacks Vision: Benchmarking and Mitigating Text Overlay-Induced Hallucination in Vision Language Models

2026-04-19

AI Summary

This work addresses the vulnerability of vision-language models to textual overlays in visual inputs, which can induce a phenomenon termed “Text-Overlay-Induced Hallucination” (TOIH)—where model predictions are dominated by spurious text semantics rather than actual visual content. The study formally defines and quantifies TOIH, introducing VisualTextTrap, a large-scale benchmark comprising 6,057 samples annotated with 88 fine-grained attributes and five-level hallucination severity labels. To mitigate this issue, the authors propose VTHM-MoE, a vision-text disentanglement framework featuring a dual-encoder architecture, four expert modules (temporal, action, object, and spatial), and an adaptive token routing strategy that dynamically resolves cross-modal conflicts. Experiments demonstrate that VTHM-MoE effectively suppresses TOIH while preserving strong performance on non-adversarial videos, significantly outperforming existing methods across multiple video question-answering benchmarks.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Recent advances in Vision-Language Models (VLMs) have substantially enhanced their ability across multimodal video understanding benchmarks spanning temporal, action, object, and spatial understanding. However, we identify a critical yet overlooked issue: when embedded on-screen text contradicts the visual scene, existing VLMs systematically hallucinate, prioritizing overlay textual semantics over the actual visual content. We define this phenomenon as Text Overlay-Induced Hallucination (TOIH). In this work, we propose VisualTextTrap, the first comprehensive benchmark, including large-scale human-validated samples with specifically designed evaluation metrics. In particular, we construct VisualTextTrap from widely-used public datasets using a scalable hybrid pipeline of VLMs assisted text generation and rigorous manual verification. The benchmark features 6,057 samples annotated across 88 fine-grained attributes within four dimensions, with hallucination intensity quantified on a five-level scale (L1--L5) that reflects the semantic contradiction between overlay text and visual reality. Moreover, we propose Visual Text Hallucination Mitigation Mixture-of-Experts (VTHM-MoE), a novel Vision-Text Disentanglement framework that employs a dual-encoder architecture. Concretely, four dimension-specialized expert modules spanning Temporal, Action, Object, and Spatial reasoning are first pre-trained to identify and leverage cross-modal discrepancies between textual semantics and actual video content. We develop an Adaptive Token Routing Strategy to enable dynamic expert allocation, conferring robust resistance to TOIH while preserving performance on uncontaminated videos. Extensive experiments conducted on our VisualTextTrap benchmark verify the effectiveness of VTHM-MoE, outperforming state-of-the-art counterparts with diverse video question answering tasks.

Problem

• Text Overlay-Induced Hallucination
• Vision-Language Models
• Multimodal Video Understanding
• Visual Hallucination
• Cross-modal Discrepancy

Innovation

• Text Overlay-Induced Hallucination
• Vision-Language Models
• VisualTextTrap
• Mixture-of-Experts
• Cross-modal Disentanglement