Reproduce “Adaptive Latent Capacity for 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
Idan Achituve, Lior Dikstein, Idit Diamant, Arnon Netzer, Hai Victor Habi
ALeWM enables adjustable planning capacity in a learned world model. It trains a wide JEPA-style representation so earlier latent coordinates help predict the full next embedding, then selects a prefix for planning. MixSIGReg accommodates masked coordinates. On OGBench-Cube, it achieves 79.00 ± 0.76% success with a mean prefix of 16 dimensions, versus 72.33 ± 1.17% and 96 dimensions for tuned LeWM; these are simulated-task results, not runtime measurements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.