FiLoRA: Focus-and-Ignore LoRA for Controllable Feature Reliance
The episode discusses 'FiLoRA,' a method by NVIDIA authors that allows users to steer an AI model’s internal attention using natural language instructions. This structural control defines 'core' semantic features versus spurious distractions, providing a reliable way to manage feature reliance without needing full retraining or simply relying on fragile prompt-tuning methods.