Reproduce “How Many Candidates Does a One-Step World Model Need? A Random-Codebook Account of Nearest-of-m Training on Frozen DINO Tokens” 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

Yong Hwi Song, Keon Myung Lee

The authors introduce IMLE-WM, which trains a world model by drawing multiple possible next states and updating the candidate closest to the observed outcome; inference then needs one network evaluation. On A2-div, with four possible modes, m=16 achieves 0.93 coverage and 0.82 precision, versus 0.79 and 0.51 for 50-step flow matching. On continuous BAIR video, samples remain under-dispersed.

Summary by alphaXiv.

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