LATIDIA · Investigación

Mitigar la degradación de la capacidad en la auto-mejora de LLM a través de la regularización de la geometría de representación

arXiv:2610.12855v1 Tipo de anuncio: nuevo Resumen: La autodestilación en la política es un paradigma de auto-mejora que permite a los grandes modelos de lenguaje (LLM) mejorar sus capacidades aprovechando el mismo modelo que ambos teac

La noticia

arXiv:2610.12855v1 Announce Type: new Abstract: On-Policy Self-distillation is a self-improvement paradigm that enables large language models (LLMs) to improve their capabilities by leveraging the same model as both teacher and student under different conditioning contexts. However, we find that OPSD methods can progressively degrade previously acquired capabilities: queries that are correctly responded at earlier training stages may become incorrectly responded as optimization proceeds, even the overall performance continues to improve. Further analysis reveals that this degradation coincides with a deterioration of the geometry of hidden representations, characterized by reduced intra-cluster compactness and diminished inter-cluster separation. Motivated by this observation, we propose Representation Geometry Regularization (RGR), which preserves representation geometry during