NeurIPS 2026

Baba in Wonderland

Online Self-Supervised Dynamics Discovery
for Executable World Models

1Korea University2Kyung Hee UniversityIIIXR Lab, South Korea

*Equal contribution†Corresponding author

Alice repeats four steps: play, find a bug, fix it, and test the fix. Watch on YouTube

At a glance

Same game, absurd words

BABA IS YOU and WALL IS STOP in Baba Is You become BABA IS STRANGE and WALL IS EAT in Baba in Wonderland.

Baba in Wonderland swaps the rule words of Baba Is You for unrelated ones. What they do has to be found by playing.

Rules learned as code

Alice writes the rules as Python code that predicts the next state, and improves the code by playing.

Still accurate with absurd words

Next-state prediction from BABA IS YOU to BABA IS STRANGE: the baseline, WorldCoder, drops from 73% to 26%; Alice goes from 99% to 95%.

Next-state prediction before and after the words change. Both use GPT-5.4; mean of 5 runs.

How Alice works

  1. 1Play

    Alice plays and saves every move. Under WATER IS EAT, water stops Baba, so its code says: if water, stay.

  2. 2Find a bug

    Under WATER IS BEGIN, Baba and the water both vanish, but the code predicts Baba stops. A move the code gets wrong is a bug.

  3. 3Fix

    A language model sees the bug and the code, not the saved moves, and writes a fix that gets the bug right: if water, vanish.

  4. 4Test

    The fix is checked on every saved move. Under WATER IS EAT it now predicts a vanish where Baba stays, so it is rejected.

  5. 5Learn from the failure

    The saved moves the rejected fix got wrong are kept as a group. The language model sees the bug with one of them, and its next fix gets both right: accepted.

  6. 6Explore

    Moves Alice has not tried yet are placed on a map of its saved moves. It tries the one farthest from what it has tried and most likely to land in its least-seen group: where its code is most likely still wrong.

Results

How accurately the learned code predicts the next state (balanced accuracy for Baba Is You). Mean of 5 runs with GPT-5.4; lines show ± one standard deviation.

All numbers

Inside a real run

Saved versions of Alice's code from a run: one move, first predicted wrong, then right after a fix.

Preparing the comparison images…

Your turn · a code world model demo

Choose a map below. Solve it yourself or watch the shortest solve.

Real gamev021predictsmake a move

Use the controls to move through the environment.

Preparing the environment…

Alice's code

v021

0/ 300 steps

Moves

    Simple puzzles we made ourselves for trying out Alice's code.

    Cite

    @article{seo2026baba,
      title={Baba in Wonderland: Online Self-Supervised Dynamics Discovery for Executable World Models},
      author={SeungWon Seo and DongHeun Han and SeongRae Noh and HyeongYeop Kang},
      journal={arXiv preprint arXiv:2605.16725},
      year={2026},
      url={https://arxiv.org/abs/2605.16725}
    }

    Questions about the paper or code? Get in touch ↗