安全、治理与可靠性 颠覆级 暂无讲解视频
发表时间
2026-10-12
arXiv
2610.13078

收录解读

第一段:安全对齐的大语言模型存在越狱和过度拒绝问题,现有工作未清晰区分“识别危害”与“写入拒绝”的计算过程。该论文将问题框架化为可测量的通路问题,在残差流更新中分离信息暴露与效应。

第二段:核心方法是将注意力更新和MLP更新分别视为感知与决策的载体,使用解码可读性分析“危害信息”与“拒绝证据”的变化,并引入因果干预验证注意力到MLP的通路依赖,最终提出Sentinel联合信号用于分类和路由。

全文在 18 对 Base–Instruct(Qwen/Llama/Gemma,1B–27B)与 6 个推理模型中区分 attention 有害性识别和 MLP 拒答写入,并用下游恢复检验功能通路。8 模型 Sentinel 联合路由平均错误 30.34%,单 P/D 为 38.09/34.96%;相较 Safety reminder,越狱 ASR 反而高 4.72 个百分点,但过度拒答低 29.52 个百分点。跨模型机制诊断和可复用路由满足 disruptive,不能宣传为所有安全指标都最好。

未达到 paradigm:通路读出和部分行为结果依赖探针/LLM 评判,白盒模型及受限测试分布,尚无独立复现或领域普遍采用。

原始摘要

Safety-aligned large language models (LLMs) remain vulnerable to jailbreaks and can over-refuse benign requests. Understanding these failures requires distinguishing harmfulness recognition from the computation that writes refusal. We reframe this as a pathway question and make it measurable on residual-stream updates in decoder-only Transformers, separating the harmfulness information a module update exposes from the change it produces in readable refusal evidence. Across 18 Base-Instruct pairs spanning Qwen, Llama, and Gemma, harmfulness is decoded more readily from attention updates, particularly in early layers, whereas MLP updates contribute more strongly to harm-selective refusal writing in aligned models. Post-training strengthens MLP refusal writing in all 18 pairs without systematically improving harmfulness decodability, while positive MLP writing becomes more concentrated in deeper layers after post-training. Across six reasoning-enabled models, harmfulness remains decodable during reasoning while strong refusal writing concentrates at answer onset, a pattern that persists with reasoning disabled. Causal interventions further support this pathway: resetting downstream MLP updates to their clean values substantially reduces the refusal shift caused by upstream attention interventions. Building on these findings, we introduce Sentinel, which combines an attention-derived perception score P with an MLP-derived decision score D. Their joint profiles characterize jailbreaks and over-refusals, and conditional routing with both signals achieves a lower aggregate error rate than single-signal routing and evaluated fixed reminders across eight models.

相关论文

链接