Quantifying modality imbalance and visual jailbreak robustness in LLaVA via projected gradient descent | Synapse