Curvature-Aware Parameter-Efficient Fine-Tuning for Vision-Language Models

Extended LoRA with Fisher-information-based (K-FAC-style) curvature approximations to stabilize low-rank adaptation in large vision-language models. Implemented adaptive rank reprojection and rank-stabilized LoRA to dynamically control model capacity under limited supervision.

Keywords: Parameter-Efficient Fine-Tuning, LoRA, Vision-Language Models, Curvature Approximation, K-FAC, Rank-Stabilized LoRA