Mechanistic Interpretability-Guided Selective Fine-Tuning of Vision-Language Models for Centimeter-Level Flood Depth Estimation
The episode discusses a paper using mechanistic interpretability to selectively fine-tune a large vision-language model for estimating centimeter-level flood depths from images. The team found that by identifying specific layers responsible for depth encoding, they could drastically reduce the number of trainable parameters while maintaining high accuracy and improving generalization to real-world data.