Decoding the Multimodal Maze: A Systematic Review on the Adoption of Explainability in Multimodal Attention-based Models

Attention-based explanation methods dominate but have significant limitations: While attention mechanisms are the most commonly used technique for explaining multimodal AI models, the review finds they "often fall short in capturing the full spectrum of interactions between modalities," meaning current popular approaches may not fully reveal how these systems actually combine different types of information (like text and images).

Vision-language models receive disproportionate research focus: The study reveals that explainability research is heavily concentrated on vision-language and language-only models, suggesting that other important multimodal combinations (such as audio-visual or sensor-based systems) lack adequate transparency research and may pose blind spots for responsible AI deployment.

Evaluation of AI explanations lacks standardization and rigor: The review identifies that "evaluation methods for XAI in multimodal settings are largely non-systematic, lacking consistency, robustness, and consideration for modality-specific cognitive and contextual factors," meaning there is currently no reliable, standardized way to verify whether explanations of multimodal AI systems are actually accurate or trustworthy.

Architectural diversity creates barriers to explainability: The "architectural heterogeneity across domains" compounds challenges in explaining multimodal models, suggesting that as AI systems become more complex and specialized, developing universal explanation methods becomes increasingly difficult, potentially limiting accountability across different applications.

Gap between performance and interpretability persists: The paper highlights a tension between multimodal models achieving "significant performance gains" while simultaneously becoming harder to explain, underscoring an ongoing challenge in responsible AI: advanced capabilities don't automatically come with the transparency needed for safe deployment.