LATIDIA · Investigación

La atención ve, el MLP se niega: rastreo de la ruta de seguridad de la atención al MLP en los LLM

arXiv: 2610.13078v1Tipo de anuncio: nuevo Resumen: Los modelos de lenguaje grande (LLM) alineados con la seguridad siguen siendo vulnerables a los jailbreaks y pueden rechazar en exceso las solicitudes benignas. Comprender estas fallas requiere distinguir harmfu

La noticia

arXiv:2610.13078v1 Announce Type: new Abstract: 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