We tried out Google’s new family of multi-modal models with variants compact enough to work on local devices. They work well.
This repository contains the introduction and the link to the source code for our ICML 2025 paper SAE-V: Interpreting Multimodal Models for Enhanced Alignment. SAE-V is a mechanistic interpretability ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results