Reproduce “Adapt for Planning, Not just Prediction: Optimistic Test-Time World Model Adaptation” 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
Akshay Mete, Dileep Kalathil, P. R. Kumar
OpTTA adapts JEPA world models by combining prediction loss on recent observed transitions with the planner’s goal-reaching cost on candidate imagined rollouts. Across 15 shifted settings, its final-success gain over the frozen model averaged 13.5 percentage points, versus 6.9 for predictor-only AdaJEPA; OpTTA exceeded AdaJEPA in 14 of 15 settings, not all.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.