Continuous Sign Language Recognition aims to predict gloss sequences from unsegmented sign language videos. Existing methods typically rely on Connectionist Temporal Classification (CTC) loss for implicit alignment, which is the most dominant loss function for current sequence alignment tasks but causes low-entropy alignment and weak visual-semantic correspondence learning. To address this issue, we propose an explicit temporal-semantic supervision framework that provides dense signals beyond CTC. Specifically, we design the Gloss-Temporal Module (GTM) to construct sustained gloss-to-frame interval supervision by using an LLM to integrate three types of information from RGB videos. We further introduce a joint optimization method with frame-level and sequence-level objectives, coupled with a curriculum learning strategy, enabling a transition from local to global alignment and improving the model’s learning of visual-semantic features. The results show that dense temporal supervision substantially alleviates alignment collapse and enables even simple RGB-only baseline models to achieve competitive performance among state-of-the-art methods without extra cues, while the recognizer converges significantly faster.

Adversarial vulnerabilities in vision-language models pose a critical challenge to the reliability of large language systems, where typographic manipulations and adversarial perturbations can effectively bypass language model defenses. We introduce Attack as Defense (AsD), the first approach to proactively defend at the cross-modality level, embedding protective perturbations in vision to disrupt attacks before they propagate to the language model. By leveraging the semantic alignment between vision and language, AsD enhances adversarial robustness through model perturbations and system-level prompting. Unlike prior work that focuses on text-stage defenses, our method integrates visual defenses to reinforce prompt-based protections, mitigating jailbreaking attacks across benchmarks. Experiments on the LLaVA-1.5 show that AsD reduces attack success rates from 56.7% to 12.6% for typographic attacks and from 89.0% to 47.5% for adversarial perturbations. Further analysis reveals that the key bottleneck in vision-language security lies not in isolated model vulnerabilities, but in cross-modal interactions, where adversarial cues in the vision model fail to consistently activate the defense mechanisms of the language model.