Reproduce “LATER: Learning Action-Conditioned Temporal Evidence Retrieval for Predictive World Models” 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

Zihan Zhou, Wenxi Xu, Qilin Wang, Hao Tang

LATER extends CF-JEPA with recurrent memory that stores completed interactions and lets candidate actions selectively read the fixed record without writing imagined outcomes into it. In simulated distracted Two-Stage PushT, it reaches the goal at step 50 in 25.0% of trials, versus 15.5% for reconstructed CF-JEPA. This result is limited to the tested simulation, not real-robot performance.

Summary by alphaXiv.

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