Tug-of-War No More: Harmonizing Accuracy and Robustness in Vision-Language Models via Stability-Aware Task Vector Merging

Proposed a stability-aware task-vector merging method that improves adversarial robustness of vision–language models without retraining.