LATIDIA · Investigación

ADTGuard: Protección adaptativa de profundidad-tiempo para la inferencia en tiempo real en modelos de lenguaje grandes

arXiv:2610.36574v1 Anuncio Tipo: nuevo Resumen: Los modelos de lenguaje grandes (LLM) pueden generar respuestas fluidas a diversas solicitudes de los usuarios, pero la transmisión puede exponer a los usuarios a contenido dañino antes de que se complete la generación. Existen

La noticia

arXiv:2610.36574v1 Announce Type: new Abstract: Large Language Models (LLMs) can generate fluent responses to diverse user requests, yet streaming may expose users to harmful content before generation is complete. Existing guardrails face two limitations in this setting. First, post-hoc moderation at response or chunk boundaries delays intervention, risking harmful exposure when content is released before verification. Second, even with token-level detection, coarse response-level supervision does not specify when harmful content begins, while fixed-layer readouts may miss relevant safety signals at other depths. To address these limitations, we introduce ADTGuard, which checks each token before release, progressively refines token-level training targets, and adapts layer selection to the current prefix. ADTGuard models