Does physics live in the activations? We probed three open video diffusion models (Wan 2.1, CogVideoX-1.5, Open-Sora 2.0) with simulated videos: speed, direction, momentum and gravity are linearly decodable from their activations, even when the input barely contains them. 🧵
Next week I will be in Vancouver 🇨🇦 for #ICML2025. Together with @SimoneM44644 and @SebastianCygert we will present "No Task Left Behind:
Isotropic Model Merging with Common and Task-Specific Subspaces". Visit the poster on Wednesday, 16th July at 11 am.
🚀 What happens when you modify the spectrum of singular values of the merged task vector? 🤔
Apparently, you achieve 🚨state-of-the-art🚨 model merging results! 🔥
✨ Introducing “No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces”