Pinned
🚀As promised, excited to share our #CVPR2026 work: ALOE - Align Once to Explain 🌱
We turn frozen foundation ViTs (DINOv3, SigLIP2, sup.) into inherently interpretable B-cos models with a single label-free alignment step.
~1000× more data-efficient than training from scratch.
Interpretability shouldn’t be an afterthought. 🤓
We’ve been working on something exciting along these lines. 🧑🍳🍳
More details soon. 👀
Coming to arXiv and a model hub near you.😉😎


