2026/04/21 by Hao Yang, Jin Wang, Yi-Xiang Wang +2
Computer Science · #Constraint Satisfaction and Optimization #Distillation #Feature (linguistics) #Multimodal Machine Learning Applications #Pattern recognition (psychology) #Process (computing) #Visual Attention and Saliency Detection #Visualization
paper · pdf · doi:10.1109/icassp55912.2026.11463006
openalex publication_date 2026/04/21 · openalex created_date 2026/04/22 · openalex updated_date 2026/08/05
Multimodal chain-of-thought (CoT) reasoning integrates visual and textual cues through step-by-step inference. In small models with limited token budgets, modality-interaction fusion often suppresses tiny cross-modal differences. In particular, multimodal CoT often struggles when different images pair with identical text or different texts pair with an identical image, making such inputs nearly indistinguishable after fusion. This study proposes Visual Saliency Steering Distillation (VSSD). VSSD leverages the attention maps of multimodal large language models to generate perturbed images that capture task-sensitive feature directions, and then applies singular value decomposition to extract dominant steering vectors to guide inter-layer distillation. Experiments on ScienceQA and M3CoT demonstrate that VSSD improves rationale generation and answer inference. The code is available at https://github.com/BGWH123/VSSD.