This paper addresses the lack of a sufficient understanding of the internal decision-making mechanisms of multimodal large-scale language models (MLLMs), which have achieved remarkable success in various visual-language tasks. Existing interpretability research has primarily focused on cross-modal features, but has tended to overlook unimodal dependencies. To address this, this paper proposes a method that leverages unimodal interactions to enhance interpretability. For visual branching, we introduce *Multiscale Explanation Aggregation (MSEA),* which dynamically adjusts receptive fields by aggregating features across multi-scale inputs, generating more holistic and spatially consistent visual explanations. For text branching, we propose *Activation Rank Correlation (ARC),* to measure the relevance of contextual tokens, suppressing erroneous activations in irrelevant contexts and maintaining semantically consistent activations.
