Contextual override, the tendency of large language models to shift away from otherwise correct predictions under misleading prompt evidence, poses a fundamental risk to reliable deployment in context-rich environments. Despite its prevalence, the internal mechanism behind this shift remains poorly understood, particularly the causal structure underlying the override process. To bridge this gap, we formulate contextual override as an internal mediation process that attributes evidence-induced answer-preference shifts to fine-grained model components. Building on this formulation, we introduce a theoretically grounded mediated-effect framework that disentangles evidence sensitivity from causal mediation, allowing override-mediating components to be precisely identified and their causal role verified through targeted ablation. We empirically validate our framework across six models and three benchmarks, demonstrating consistent support for our mediation account across diverse architectures and task settings. Our study reveals that *contextual override is routed through sparse, structurally localized pathways, rather than being diffusely entangled with the model’s internal knowledge*. These findings provide a mechanistic diagnostic for auditing contextual override and a targeted interventional test for validating the identified pathways, rather than a general-purpose defense mechanism.

Visual evidence selection is a critical component of multimodal retrieval-augmented generation (RAG), yet existing methods typically rely on semantic relevance or surface-level similarity, which are often misaligned with the actual utility of visual evidence for downstream reasoning. We reformulate multimodal evidence selection from an information-theoretic perspective by defining evidence utility as the information gain induced on a model’s output distribution. To overcome the intractability of answer-space optimization, we introduce a latent notion of evidence helpfulness and theoretically show that, under mild assumptions, ranking evidence by information gain on this latent variable is equivalent to answer-space utility. We further propose a training-free, surrogate-accelerated framework that efficiently estimates evidence utility using lightweight multimodal models. Experiments on MRAG-Bench and Visual-RAG across multiple model families demonstrate that our method consistently outperforms state-of-the-art RAG baselines while achieving substantial reductions in computational cost. We release our code at https://github.com/Hcnaeg/utility-mrag.