Reproduce “An Evaluation of Tabular Foundation Models as Training-Free 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
Xian Yeow Lee, Ryoichi Takase, Ahmed Farahat
The authors test whether off-the-shelf TabPFN predicts dynamics from recent transitions without task-specific weight updates. It treats state and action as inputs and state change as the target, then adds predicted change to the current state. With at least 10 transitions, it beats freshly trained neural baselines on one-step prediction, but only long-context HalfCheetah rollouts beat persistence and a full-data LSTM.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.