Planning needs a transition model: given a state and an action, what comes next? When a large language model plays that role, every next state is generated token by token, which makes searching over many possible futures slow and expensive. EmbedPlan embeds the state and the action with a frozen language model, predicts the next state's embedding with a small learned network, and returns the closest real state.

This organization is the home of the project:

• EmbedPlan/EmbedPlan: the library and the code behind the paper. Fit it on your own text transitions in a few lines, in the scikit-learn style (EmbedPlan().fit(X, y).evaluate(X_test, y_test)), or reproduce the paper's tables.

Tried EmbedPlan on your own domain, a new encoder, or a search algorithm on top of it? Tell us in Discussions. Issues labeled good first issue are a good place to start, and CONTRIBUTING.md explains the rest.

Eliezer Shlomi, Ido Levy, Eilam Shapira, Michael Katz, Guy Uziel, Segev Shlomov, Nir Mashkif, Roi Reichart and Sarah Keren (Technion and IBM). Eliezer Shlomi and Ido Levy contributed equally and maintain the code.

@article{shlomi2026textual,
  title   = {Textual Planning with Explicit Latent Transitions},
  author  = {Shlomi, Eliezer and Levy, Ido and Shapira, Eilam and Katz, Michael and Uziel, Guy
             and Shlomov, Segev and Mashkif, Nir and Reichart, Roi and Keren, Sarah},
  journal = {arXiv preprint arXiv:2602.04557},
  year    = {2026}
}