Interpreting and Enhancing Emotional Circuits in Large Vision-Language Models via Cross-Modal Information Flow

Large Vision-Language Models (LVLMs) represent a significant leap towards empathetic agents, demonstrating remarkable capabilities in emotion understanding. However, the internal mechanisms governing how LVLMs translate abstract visual stimuli into coherent emotional narratives remain largely unexplored, primarily due to the scarcity of visual counterfactuals and the diffuse nature of emotional expression. In this paper, we bridge this gap by introducing a steering-vector-based causal attribution framework tailored for descriptive emotional reasoning. To this end, we construct a specialized dataset to demystify the emotional circuits underlying the three-stage '' Adapt-Aggregate-Execute'' mechanism. Crucially, we discover a functional decoupling: visual emotional cues are aggregated in middle layers via $\textit{sentiment-specific}$ attention heads, but are subsequently translated into narrative generation in deep layers through $\textit{emotion-general}$ pathways. Guided by these insights, we regulate the emotional information routing to strengthen attention flow and amplify the semantic activation to consolidate expression. Extensive experiments on the comprehensive MER-UniBench demonstrate that our methods significantly improve performance via inference-time intervention, effectively mitigating emotional hallucinations and corroborating the causal fidelity of the discovered circuits.

Vision-language AI systems are increasingly expected to understand not only objects in images, but also the emotions and social cues behind them. However, their emotional understanding is still unreliable: for example, a model may describe a sad or distressed scene as cheerful. To improve this ability, it is not enough to only train models on more examples; we also need to understand how they internally recognize emotional cues and turn them into written descriptions. Our work studies this hidden process inside large vision-language models. We build carefully paired image examples that separate emotional information from other visual content, and use them to trace which parts of the model carry and use emotion-related signals. We find that emotional understanding follows three stages: early layers adapt visual information, middle layers gather emotion-specific cues, and deeper layers turn these cues into general emotional language. Based on this understanding, we introduce VEENA, a training-free method that adjusts the model during inference so that emotional information is routed and expressed more reliably. Across a broad benchmark of multimodal emotion understanding tasks, VEENA improves performance and reduces emotional hallucinations without retraining the model. This work helps make emotion-aware AI systems more transparent, reliable, and socially responsible.