LATIDIA · Ciberseguridad

Atribución a nivel de token de marcas de agua LLM multimodo: regla óptima y ponderación de creencias

arXiv: 2610.23601v1Announce Type: new Resumen: Un verificador de marca de agua puede contener muchos modos, es decir, claves secretas, y un documento puede mezclar tramos de diferentes modos con texto sin marca de agua. Informar directamente del modo con el l

La noticia

arXiv:2610.23601v1 Announce Type: new Abstract: A watermark verifier may hold many modes, i.e., secret keys, and a document may mix spans of different modes with unwatermarked text. Directly reporting the mode with the largest pivotal statistic (pivot) inflates the Type I error, and restoring the nominal level reduces power as candidate modes are added. We formulate token-level multi-mode attribution for the Gumbel-max watermark as maximizing correct-label power under a Type I error constraint given a mode prior. We show that the optimal rule reports the mode with the largest prior-weighted alternative density and rejects when this density exceeds one common threshold. Based on this rule, we propose the Belief-weighted Likelihood