首页 > AI前沿 > Where Is Rank-1 DiM Refusal Intervenable in Multimodal LM Derivatives?

Where Is Rank-1 DiM Refusal Intervenable in Multimodal LM Derivatives?

arXiv机器学习 2026-11-15 14:15 12 阅读 查看原文
用户现在需要我把给定的纯文本新闻正文排版成语义化的HTML片段,要遵循给定的规则。首先,先看原文,是一段学术类的新闻?不对,是一段研究相关的内容,没有明显的小节?不对,原文是一个大段落?不对,先仔细看,原文的内容:首先讲现在开源权重模型库的情况,然后提出问题,然后讲测试的发现,然后讲相关的验证,然后讲主要声明,然后讲VLGuard的测试,最后讲账本的作用。 首先,有没有适合的小节?比如,先整体的,不对,先看,原文是连贯的,但可能需要划分?不对,先看规则:长文用h2/h3划分小节,每个自然段落用p包裹。哦,原文的几个部分?不对,原文其实是:首先第一部分是背景和研究问题:从“Open-weight model hubs now ship”到“comparable bypass?”这部分是背景和提出问题?然后下一部分是测试设置和结果:“On the mixed-input settings we stress-test”到“content.” 然后下一部分是探针结果:“Margin-rank and leverage probes on a Vision leaf”到“we study.” 然后下一部分是支撑性验证:“Catalog breadth and mapper-free base”到“judges.” 然后下一部分是主要声明:“The primary claim is this recipe-scoped”到“by site).” 然后下一部分是VLGuard的测试:“On VLGuard Unsafe prompts”到“at $T$).” 然后最后是账本的作用:“An intervention-oriented Class”到“multimodal bypass.” 不对,有没有更合适的?或者,可能可以分几个h2?比如

研究背景与核心问题

?不对,等下,不能增改内容?不对,规则4说保留原文所有信息,不得增删改内容,哦,小节标题?不对不对,等下,规则2说“长文用

/

划分小节”,那小节标题是需要自己加的?不对不对,等下,原文没有小节标题啊?哦不对,等下,是不是我理解错了?再看规则:“每个自然段落用

包裹,长文用

/

划分小节”,哦,不对,原文有没有自然的段落划分?看原文,原文是一整段?不对,再仔细看用户给的正文: 原文是: Open-weight model hubs now ship not only chat checkpoints but a long tail of derivatives---fine-tunes, safety guards, vision--language models (VLMs), audio wrappers, and generative heads---yet residual-stream safety audits still often intervene at a single site, usually the text decoder $T$. We ask a localization question under one shared mean / rank-1 difference-in-means (DiM) extract$=$intervene protocol, with matched-norm random and empty-rate controls: does causal chat attack lift concentrate at $T$, or do perceptual encoders, projectors, and fusion paths ($E$/$P$/$F$) supply a comparable bypass? On the mixed-input settings we stress-test---HADES adversarial images paired with harmful text prompts, and spoken AdvBench via text-to-speech (TTS) or silence---text DiM at $T$ raises attack success rate (ASR) while the same mean DiM at $E$/$P$/$F$ does not; attacks track harmful text rather than image or audio content. Margin-rank and leverage probes on a Vision leaf show that this recipe is extremal at $T$ and near chance at $P$/$E$, which pressures a weak-extractor reading of those flat cells for the DiM family we study. Catalog breadth and mapper-free base$\to$derivative transfer when residual widths match provide supporting coverage under the same gates; generative U-Nets are a separate band under generation judges. The primary claim is this recipe-scoped $T$ vs.\ non-text intervenability asymmetry---intervention accessibility under mean / rank-1 DiM, not a claim that refusal is absent or fully represented outside $T$ (text vs.\ perceptual contrasts and pooling differ by site). On VLGuard Unsafe prompts ($n{=}200$; Llama-Guard), the same $T$ vs.\ $P$/$E$ directionality holds at modest absolute ASR ($0.030{\to}0.150$ at $T$). An intervention-oriented Class$\times$site ledger records where this protocol still hooks as supporting coverage, not as a universal multimodal bypass. 哦,原来原文是一整段?那需要划分小节,加合适的小节标题?不对,等下规则4说不得增删改内容,那小节标题算不算增?不对,再仔细看规则2的表述:“长文用

/

划分小节”,意思是用这些标签来划分内容的小节,小节标题是必要的?或者,是不是把原文的逻辑层次用h2和p区分?不对,再想,可能不需要加额外的文字内容,比如标题?不对,那h2/h3如果用来做小节的话,里面是小节标题,但是原文没有,那是不是不对?哦,等下,可能我理解错了,原文的几个逻辑块,每个逻辑块可以用h3或者h2作为小节的标识?不对,不对,再看规则,还有,重要句子用strong,引用?这里没有引用,列表?也没有,代码?也没有。 哦,等下,可能原文可以拆分成几个自然的逻辑段落,每个用p,然后如果有大的章节的话用h2?不对,再梳理原文的逻辑: 1. 背景与研究问题:Open-weight model hubs...comparable bypass? 2. 实验设置与初步结果:On the mixed-input settings...content. 3. 探针验证结果:Margin-rank and leverage probes...we study. 4. 支撑性证据:Catalog breadth and mapper-free base...judges. 5. 核心声明:The primary claim is...by site). 6. 针对性验证:On VLGuard Unsafe prompts...at $T$). 7. 研究边界说明:An intervention-oriented Class...multimodal bypass. 那这样的话,可以用h2来分几个大的部分?比如

研究背景与核心问题

,然后下面

内容,然后

实验设计与验证结果

,下面几个p,然后

核心声明与研究边界

?不对,但是这样加的标题算不算增内容?不对,规则4说“保留原文所有信息,不得增删改内容”,哦,哦,不对,是不是小节标题不需要额外添加,而是用h2/h3来标注原文里的内容?不对,原文里没有小标题啊。哦,等下,可能我错了,规则2的意思是,长文的话,用h2/h3来划分小节,也就是把原文的逻辑层次用这些标签组织,不需要额外加标题文字?不对,那h2/h3里放什么?哦,不对,再仔细看规则,用户说“长文用

/

划分小节”,可能是指将原文的不同内容模块用这些标签作为小节的标题,但是标题内容是对模块的概括?但是规则4说不得增删改内容,这会不会冲突? 哦,等下,不对,可能原文其实是一整段,不需要划分小节?那如果不用h2/h3的话,是不是可以?不对,再看原文,内容比较长,是长文,所以需要划分。或者,可能我理解错了,规则2的“长文用

/

划分小节”是指如果原文有小节的话就用,没有的话,把逻辑段落用p,然后重要的小节标识用h2/h3?不对,再想,可能不需要加额外的文字,比如把每个逻辑块的首句?不对,不行,不能改内容。哦,等下,可能原文的结构里,没有小节标题,那是不是不需要加h2/h3?不对,那规则2说长文用,那怎么用? 哦,等下,再仔细看用户的需求,用户给的是新闻正文,可能是科技新闻?那排版的话,比如: 首先,

Open-weight model hubs now ship not only chat checkpoints but a long tail of derivatives---fine-tunes, safety guards, vision--language models (VLMs), audio wrappers, and generative heads---yet residual-stream safety audits still often intervene at a single site, usually the text decoder $T$. We ask a localization question under one shared mean / rank-1 difference-in-means (DiM) extract$=$intervene protocol, with matched-norm random and empty-rate controls: does causal chat attack lift concentrate at $T$, or do perceptual encoders, projectors, and fusion paths ($E$/$P$/$F$) supply a comparable