Reproduce “Knowing When to Stop Thinking: A Dissection of Expected Free Energy as a Test-Time Compute Controller for Latent World-Model Planning” 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
Tong Zhang, Zhiyuan Shi, Yun Peng, Tao Xie
This paper tests whether image-based robot planners can use model-cost progress to decide when to stop refining candidate actions, and separately evaluates disagreement scores and online predictor updates. On DINO-WM’s Wall task, success peaks at 57.6% at depth 11, then falls to 35.6% at depth 30 as model cost decreases; cost-only stopping does not prevent the decline in this task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.