Visual Saliency Steering Distillation for Multimodal Chain-of-Thought Reasoning

Researchers proposed Visual Saliency Steering Distillation (VSSD), a new distillation method for multimodal chain-of-thought reasoning.

VSSD uses attention maps to generate task-sensitive image perturbations and singular value decomposition to extract steering vectors for inter-layer guidance.

The approach is designed to help compact vision-language models preserve subtle image-text differences that can be lost during fusion.

It improves reasoning performance on benchmarks such as ScienceQA and M^3^CoT.