From interaction to executable world models

Baba in Wonderland

Online Self-Supervised Dynamics Discovery
for Executable World Models

SeungWon Seo*1 DongHeun Han*2 SeongRae Noh1 HyeongYeop Kang1

1Korea University2Kyung Hee University

Read the paper Code Play the environment

Same world. Unfamiliar words. Rules discovered through experience.

THE IDEA An unfamiliar vocabulary removes lexical shortcuts. Alice discovers the dynamics by interacting, finding prediction errors, and revising executable code.
01 / OVERVIEW

When words stop
giving the rules away.

Baba Is You is a puzzle game whose rules are movable word blocks. BABA IS YOU means you control Baba; FLAG IS WIN makes touching a flag a win. Rearrange the words into ROCK IS YOU, and you control the rocks instead.

Explore the original Baba Is You on Steam ↗

Baba in Wonderland keeps the same mechanics but changes familiar property words: YOU becomes STRANGE. Without a dictionary, rule descriptions, or a learning reward, the agent has to discover what these words do by acting and observing the result.

Alice learns a world model written in Python: give it a state and an action, and it predicts the next state. The research measures how accurately this code predicts changes in the world. In the demo below, you or a saved solution choose the actions; the learned program predicts their outcomes.

DEFAULT WORLD

BABAISYOU

BABA IN WONDERLAND

BABAISSTRANGE

The label changes. The behavior stays the same.

Read the abstract

Executable world models can be read, edited, executed, and reused for planning, but only if the program captures the environment's transition law rather than semantic shortcuts in its surface vocabulary. We study online executable world-model learning under prior misalignment, where an agent must induce state-dependent dynamics from interaction evidence alone, without rule descriptions, reward signals, or trustworthy lexical priors.

We introduce Alice, a closed-loop system that treats failed candidate updates as structural signal: when a candidate explains a new transition but loses previously explained ones, the preservation conflict reveals dynamics that the current program had conflated. Alice refines these conflicts into hypothesis classes that both provide compact, class-stratified preservation counterexamples for update and guide frontier exploration toward transitions that are novel and underrepresented with respect to the current program.

We evaluate Alice on Baba in Wonderland, a prior-misaligned variant of Baba Is You that preserves simulator dynamics while replacing semantically meaningful rule-property labels with unrelated words. Experiments show that Alice substantially improves executable world-model learning under prior misalignment, and ablations show that both class refinement and class-aware exploration contribute.

02 / WATCH THE PROGRAM CHANGE

See the mistake.
See what changes.

Two moments on the same map, replayed through saved programs. First, a model learns to move Baba. Then, it learns that pushing one object can move a whole chain.

DEFAULT WORLD YOU controls Baba. PUSH makes an object pushable.

Preparing the comparison images…

Fixed images from the supplied Default World programs on the first map, cropped to the action. Each row uses the same input and RIGHT action; the resolved prediction exactly matches the simulator. Red dashed outlines mark errors. These are illustrative replays, not a claim that training encountered these exact scenes.

NOW, TRY IT YOURSELF

Your next move.

Choose a map below. Solve it yourself or use Auto to replay a solution.

FIRST, UNDERSTAND. THEN, QUESTION.

What if the words stopped helping?

YOU identifies what you control, PUSH what you can push, and WIN the goal. Start here, then switch to Wonderland to keep the scene and change the vocabulary.

DEFAULT WORLD
Make a move to compare.

Actual environment

Use the controls to move through the environment.

Preparing the environment…

Learned world model

Make a move to see a prediction.

Same state + same actionOnly the program version changes.
THE EXECUTABLE WORLD MODEL

default/v021.py

Original saved Python, executed for the prediction on the left. Changes compares this revision with the previous saved revision.

Each world opens with its latest supplied program and remembers your selection. Changing worlds or versions keeps your game in place. Saved programs predict each move from the actual pre-action state; they are not trained live here. Revision numbers are specific to each run, not LLM-call counts. These examples do not measure full-benchmark accuracy.

03 / MEET ALICE

A failed update
can teach you something.

A fix can correct one prediction while breaking another. Alice uses these conflicts to separate past experiences into finer groups, called hypothesis classes. The same groups guide which old behaviors to preserve in the next code update and which unfamiliar dynamics to explore next.

  1. 01

    Observe a failure

    Interact with the environment and find a transition the current program cannot explain.

  2. 02

    Refine the hypotheses

    A rejected update separates preserved and lost transitions into finer hypothesis classes.

  3. 03

    Learn, then explore

    Use the same classes to choose preservation examples and seek underrepresented dynamics.

THE PROGRAMMER

Keep what works.
Separate what doesn't.

Representative counterexamples guide the next code revision. Acceptance is checked against all previously explained transitions.

Rejected updates expose distinctions that the previous program had conflated.
THE EXPLORER

Let the model's gaps
guide the next move.

Frontier candidates are scored by embedding novelty and expected coverage of rare hypothesis classes.

One source of evidence shapes both program updates and data collection.
04 / RESULTS

Learning beyond
familiar vocabulary.

Exact next-state prediction, including the uncommon transitions that are easy to overlook.

Online prediction accuracy in Baba in Wonderland
MethodAll accuracy ↑Balanced accuracy ↑LLM calls

All accuracy is exact one-step state matching. Balanced accuracy uses a class-reduced set grouped by action and state-change signature. Results use GPT-5.4, a 100-call cap, and one run per configuration; they are point estimates. Evaluation details ↗

05 / CITE THIS WORK

Baba in your bibliography.

BibTeX
@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 ↗