World Modeling from Observations: self-improving world modelling via latent actions
Hosted by NICE AI Talk
Registration
Past Event
About Event
NICE Talk 156 🌟 invites Dr. Yifu Qiu, PhD candidate at the University of Edinburgh, jointly supervised at Cambridge University.
🥳We will talk about models' self-improving world modelling via latent actions!
SWIRL🍥, a self-improving framework that treats actions as latent variables and alternates optimization between a forward world model and an inverse dynamics model to learn solely from state sequences.
The ability to internally model the world is essential: predicting next states from current states and actions.
SWIRL has cross-modal SOTA, achieving a 16% improvement on AURORA-BENCH, 28% on ByteMorph, 16% on WORLD-PREDICTION-BENCH, and 14% on STABLETOOL-BENCH.
paper: https://arxiv.org/pdf/2602.06130