Towards strong generalization in LLMs

Abstract: In order to truly solve open-ended problems, future language models must demonstrate new capabilities beyond their training data — a paradigm of strong generalization. I’ll present how simple-to-hard curricula; via teacher-trajectory distillation, skill-targeted math training, and cross-modal alignment in VLMs; consistently unlock strong generalization under limited data and model capacity.  I’ll conclude with open directions on architectures, optimization, and autonomy for systems that keep learning new capabilities.