Reproduce “Self-Supervised Latent Action-Conditioned World Modeling for Time Series” 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

Chia-Han Kuo, Chengshuai Zhao, Huan Liu, Lo Pang-Yun Ting

ACTS-WM learns task-conditioned latent actions from time-series transitions without explicit action labels, then uses them to predict target-series representations with a JEPA-based encoder. Its inverse and forward dynamics models learn transition codes, while a latent policy predicts codes from observations. On ETTh1, predicted representations alone lower MSE from 1.218 to 0.678, among 11 settings; results vary elsewhere.

Summary by alphaXiv.

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