LATIDIA · Ciberseguridad

Suma por resta: hacia la mitigación de la alineación de seguridad frágil en modelos de lenguaje grandes

arXiv: 2611.06553v1Announce Type: cross Resumen: Los modelos de lenguaje grandes (LLM) se someten a una alineación de seguridad para evitar generaciones dañinas, pero siguen siendo vulnerables a los ataques de jailbreak. Trabajos recientes revelan una estructura más profunda

La noticia

arXiv:2611.06553v1 Announce Type: cross Abstract: Large language models (LLMs) undergo safety alignment to prevent harmful generations, yet they remain vulnerable to jailbreak attacks. Recent work reveals a deeper structural weakness: safety alignment can be highly sparse, such that pruning a small number of safety-critical neurons or ablating a single refusal direction can substantially undermine aligned behavior. In this work, we take the first steps toward strengthening safety alignment such that it is less dependent on sparse model components. We propose Addition by Subtraction, which gradually removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model. Our results show that the