Reproduce “WARP-JEPA: World-Model-in-the-Loop Amortized Rollout Planning for Frozen Hierarchical Joint-Embedding Predictive Architectures” 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

Junjie Li, Zhaojiacheng Zhou, Kaixiang Wang, Zihan Wang, QingSong Zhao, Jiong Lou, Jie Li

WARP-JEPA replaces repeated test-time action optimization in hierarchical visual control with learned proposals. It generates candidate plans, scores them using a frozen world model, and selects the lowest-cost plan; its top-level Transformer feeds imagined states back into generation. On Visual AntMaze L2, it achieves 78.7% success versus 45.3% for Online GD and runs 13.9× faster, under the reported evaluation conditions.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.