Bryce Grant @ ICLR 2026

Presenting two mechanistic interpretability works today at @iclr_conf Giving a talk from 2-2:15 at the Multimodal Intelligence workshop (204C) on our work “Not All Features are Created Equal: A Mechanistic Study of Vision-Language-Action Models.” Will also be at the Unifying Concept Representation Learning workshop (209) presenting “Gluing Local Contexts into Global Meaning: A Sheaf-Theoretic Decomposition of Transformer Representations.” Feel free to stop by if you want to talk VLAs, Robotics, or Algebraic Topology!

Thank you to @nvidia for providing the resources and compute for this project! We are looking forward to what the community does with these tools.

To let people explore all of this we built Action Atlas: action-atlas.com 📄 arxiv.org/abs/2603.19233… 🌐 cwru-aism.github.io/vla-interp-page 💻 github.com/CWRU-AISM/acti… In the coming weeks we will be releasing 7TB+ of activation data, SAE checkpoints, as well as results from our physical experiments where we're already seeing the same trends!

Vision Language Action Models: do they actually understand language, or just replay visual-motor habits? We studied 6 architectures across 400k+ episodes in to find out in the first large-scale VLA mechanistic interpretability study. Presenting "Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models" at the #ICLR2026 MM Intelligence Workshop w/ @xijiazhao Also releasing Action Atlas: our take on Neuronpedia for robot policies cwru-aism.github.io/vla-interp-page

To learn more, see more videos, and read a full research paper about Recap and π*0.6, see our blog post here: pi.website/blog/pistar06

I'm very excited to finally announce one of the most ambitious projects we've worked on — which makes the front cover of Science Robotics today: ☀️ Learning a Thousand Tasks in a Day ⭐️ Everyday tasks — like those below — can now be learned from a single demonstration each...