LATIDIA · Robótica

Se niega en el chat, cumple con los brazos: la negativa de LLM no sobrevive al control del robot

arXiv: 2610.14043v1Tipo de anuncio: Cross Resumen: Los modelos de lenguaje de frontera se niegan de manera confiable a explicar cómo causar daño físico. Preguntamos si esta negativa sobrevive cuando los mismos modelos controlan un robot. Presentamos Ro

La noticia

arXiv:2610.14043v1 Announce Type: cross Abstract: Frontier language models reliably refuse to explain how to cause physical harm. We ask whether this refusal survives when the same models control a robot. We introduce RoboHarm, a physical AI safety benchmark of harmful robot instructions, and test seven frontier models on it. We hold each model fixed and vary only the interface, from text to Code as Policies to vision agents that command end-effector poses or joint angles. In a MuJoCo simulation twin of a bimanual YAM arm robot, refusal falls from 73% of text requests to 18-25% once the model controls the robot, even when it is told what the objects are.