Reproduce “Keeping JEPA World Models Plannable When Little of the Frame Moves” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Florian Strohm, Patrick Wagner, Jannik Schwab, Marco Huber
The authors add a training-only inverse-dynamics head to a JEPA world model, encouraging latent states to preserve action-relevant information. The head is discarded at test time. In SLIM, it averages 0.348 ± 0.016 success versus 0.003 ± 0.001 for plain LeWM across three seeds. On hard pushing, success is 0.157 versus 0.003. This supports a repair on this simulated benchmark, not a general guarantee.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.