Pulling The REINS：它和通用视觉自监督的关系在于：不是 SSL 训练法，但说明安全相关结构可在线性表示方向中分离，适合作为生成模型 hidden representation 读物
中相关；详见方法、贡献和实验边界。
Pulling The REINS: Training-Free Safety Alignment of Video Diffusion Models via Representation Steering arXiv 新增 + video diffusion representation steering 原文链接
编号2606.17257优先级P3类别arXiv 新增 + video diffusion representation steering会议arXiv 新增 + video diffusion representation steering方法不是 SSL 训练法，但说明安全相关结构可在线性表示方向中分离，适合作为生成模型 hidden representation 读物来源arXiv / OpenReview
Figureure 2 · : Qualitative results: For each prompt, we show baseline generation anFigure 2: Qualitative results: For each prompt, we show baseline generation and REINS-steered generation using identical random seeds for paired comparison. REINS redirects harmful content toward safe content. Some explicit content has been censored.这张图/表用于判断 Pulling The REINS 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致，而不是只看单个最高分。Figureure 3 · : Effective steering range across transformer layersFigure 3: Effective steering range across transformer layers. We apply REINS at a fixed steering strength across candidate layers and measure safety rate and VQ/MQ win rates. The shaded effective steering range highlights the band of intermediate layers (∼50% depth) where safety improves substantially while VQ and MQ remain near baseline. The operating region predicted by the information-steerability tradeoff. Outside this band, two diagnostic failure modes emerge: early layers lack composed safety features, leaving steering ineffective; late layers produce a degenerate safety rate of 1.0 via generation collapse (in (a), MQ falls past layer 25). The same band is recovered under both classifiers, confirming the profile is a property of the model’s representation rather than the calibration classifier.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 Pulling The REINS 的方法或实验，请结合正文精读段落一起看。
