Multi-level Constraint-Based Two-Stage Few-Shot Knowledge Distillation for Vision-Language Models. IJCAI. 2026;50(13). doi:10.31449/inf.v50i13.13666