Meta-LoRA Uses PAC-Bayes Regularization for Cross-Domain Personalization

Meta-LoRA implements a meta-learned LoRA initialization that adjusts update strength based on support-set size and predictive uncertainty. This method prevents overfitting during few-shot personalization in unseen conversational domains by calibrating adaptation to the quality of available evidence.