Executable World Models: AI That Understands the World by Writing Code, Not Guessing Words

Topic: Interactive reasoning world model based on code generation Paper: *Executable World Models for ARC-AGI-3 in the Era of Coding Agents* Author: Sergey Rodionov (SingularityNET, 2026)

---

1. What Does It Mean to Truly "Understand"?

Feynman famously said: "What I cannot create, I do not understand."

If you ask a current LLM: "If I move this red block next to the blue circle, what happens?" It will guess the most statistically likely answer based on billions of lines of text it has seen. That is "statistical simulation," not understanding.

Real understanding means having a simulator in your head. You don't need to look up the answer—you just run the action mentally and see the result.

The core idea of this paper: instead of giving answers directly, the AI first writes Python code that simulates the world's rules.

2. ARC-AGI-3: AI's "Intelligence Ceiling"

To prove this idea, the author took on ARC-AGI-3, a benchmark created by François Chollet specifically designed to catch AIs that rely on memorization.

ARC puzzles look like this: you're given several examples of a graphical transformation and must infer the rule. For example, "every red block rotates 90 degrees around the blue block." Humans spot it instantly; AI struggles because it has never seen that specific pattern before.

3. "Writing Code" Is "Modeling"

The paper's Executable World Models framework turns the AI into a programmer:

1. Observation and Hypothesis: After seeing the examples, the AI doesn't guess directly—it writes Python code: "I think the rule is color_fill(x, y, red)." 2. Self-Verification: It runs the code locally. If the output matches the examples exactly, its "world model" is correct. 3. Relentless Refactoring: This is the most brilliant step—the AI tries to make its code increasingly shorter.

• Why? Occam's razor: the simpler the explanation, the closer it tends to be to the truth (minimum description length, MDL).
• The AI keeps simplifying its code logic until it distills the core underlying rule.
4. Simulation-Based Planning: Once the model is built, the AI first runs it in the code simulator, confirms the result, and only then submits its final answer.

4. Core Thesis: Code Is the Skeleton of Rationality

The paper makes a strong claim: code generation ability is the key to general-purpose reasoning in AI.

• Explosive interpretability: If the AI gets something wrong, you can read its Python code and instantly see where the logic went off track.
• Striking efficiency: On the 25 hardest ARC games, this code-modeling approach solved 7 outright. It may not sound like much, but given ARC's brutal difficulty, this is a milestone.
• Cross-domain generality: As long as an environment can be described in code (physics, chemistry, programming), the AI can learn how to operate in it by "writing a simulator."

5. Takeaway: From "Probability" Back to "Logic"

We used to think that with enough data, AI would scale into intelligence. This paper argues otherwise: intelligence isn't about how much you remember—it's about whether you can distill messy phenomena into concise, runnable logic.

This is more than an AI framework; it's the scientific method, digitized: observe → model → verify → simplify.

Editor's note: This paper represents a necessary path toward AGI. Once AI learns to explore the unknown by "writing world models," it stops being a parroting machine and becomes a genuine "digital scientist."