📄 中文摘要
大型语言模型在流媒体场景中的广泛应用使得传统的后期安全防护措施无法有效地实时拦截不安全内容。基于标记级监督训练的流媒体安全防护虽然能够解决这一问题，但需要昂贵的标注，并且容易出现严重的过拟合现象。该研究挑战了流媒体安全必须依赖标记级监督训练的传统观念，提出了NExT-Guard框架，利用后期安全防护的内在能力，通过监控稀疏特征的可解释潜在特征，实现无需训练的流媒体安全防护。NExT-Guard能够有效地编码标记级风险信号，从而在流媒体环境中提供实时的安全保障。
NExT-Guard: 无需训练的流媒体安全防护机制
📄 English Summary
NExT-Guard: Training-Free Streaming Safeguard without Token-Level Labels
Large language models are increasingly utilized in streaming scenarios, which renders conventional post-hoc safeguards ineffective in real-time interception of unsafe content. While streaming safeguards based on token-level supervised training could potentially address this issue, they require costly annotations and are prone to severe overfitting. This research challenges the notion that streaming safety must rely on token-level supervised training. Instead, it introduces NExT-Guard, a training-free framework that leverages the inherent capabilities of well-trained post-hoc safeguards. By monitoring interpretable latent features from Sparse representations, NExT-Guard effectively encodes token-level risk signals, enabling real-time safety measures in streaming environments.