{"ok":true,"nota":{"id":"ce7d3262865b36fb","slug":"una-nueva-tecnica-de-entrenamiento-busca-ensenar-habilidades-a-modelos-ce7d32","titulo":"Una nueva técnica de entrenamiento busca enseñar habilidades a modelos de IA sin arruinarlos","resumen":"Presentan una idea de post-entrenamiento basada en aprendizaje por refuerzo, en la que el modelo se corrige a partir de sus propias respuestas. Buscan incorporar nuevas capacidades sin degradar las que ya tiene.","fuente":"tds","fuenteNombre":"Towards Data Science","color":"#2d6a3e","autor":"Jonathan Yahav","url":"https://towardsdatascience.com/could-on-policy-self-distillation-revolutionize-post-training/","imagen":"/hoy/portada/ce7d3262865b36fb","tema":"tecnologia","subtemas":["inteligencia artificial","entrenamiento de modelos","aprendizaje por refuerzo","modelos fundacionales","post-entrenamiento"],"entidades":["Towards Data Science"],"formato":"dato","entretenimiento":false,"pubTs":"2026-10-22T11:00:00.000Z","indexTs":"2026-10-22T16:38:44.280Z","hash":"b0002963f7f1e1d731dc0448aadaf61e8424f09ae94b9368edff698a23987a60","tsa":{"time":"Oct 22 16:39:58 2026 GMT","provider":"freetsa.org","sobre":"sello"},"stats":{"views":0,"likes":0,"clicks":0,"dwell":0},"tipo":"normal","registro":"hallazgo","porQueImporta":"Hoy agregar una habilidad a un modelo de IA suele hacerle perder otras, y una forma de evitarlo abarataría y haría más confiables las mejoras. Para el público general, esto explica por qué a veces un asistente mejora en una tarea y empeora en otra. Todavía es una propuesta en etapa temprana, sin resultados concluyentes en el texto. Cuando se habla de revolución en la IA, conviene esperar a ver pruebas repetidas por otros equipos.","quienFirma":"Towards Data Science","confianza":"divulgacion","consulta":null},"sello":{"hash":"b0002963f7f1e1d731dc0448aadaf61e8424f09ae94b9368edff698a23987a60","sello":"bcd43d703637f9fb0528f660c87a253d2c48326e4f05859e7efab93b9d800840","beacon":{"drand":{"cadena":"8990e7a9aaed2ffed73dbd7092123d6f289930540d7651336225dc172e51b2ce","ronda":6575205,"aleatorio":"8d668dee5f5988ce3b254235634f030221669521ad13430e3fc9c5056f311f5c","tiempo":"2026-10-22T16:39:30.000Z"},"nist":{"cadena":2,"pulso":1994802,"salida":"12DD79E00124B87AE9DF53F82C2769EC701A6E2C4BF455A27DBAB985BBB34D6404A618CEE876D315F983790BF0CE5F8356121E9A555C928585F812B2FC6DFF0C","tiempo":"2026-10-22T16:39:00.000Z"},"noAntes":"2026-10-22T16:39:30.000Z"},"verificacion":{"drand":true,"nist":true},"tsa":{"time":"Oct 22 16:39:58 2026 GMT","provider":"freetsa.org","sobre":"sello"},"original":{"titulo":"Could On-Policy Self-Distillation Revolutionize Post-Training?","descripcion":"Teaching foundation models new skills without breaking them is hard. A new reinforcement learning-based approach might have the power to change that. The post Could On-Policy Self-Distillation Revolutionize Post-Training? appeared first on Towards Data Science .","pubDate":"Thu, 22 Oct 2026 11:00:00 GMT","url":"https://towardsdatascience.com/could-on-policy-self-distillation-revolutionize-post-training/"},"indexado":"2026-10-22T16:38:44.280Z"}}