The World Is a Hypothesis: Epistemic Control for Executable World-Model Agents in Novel Interactive Environments
This paper introduces THEA, a prompt-defined revision policy that treats executable world models as falsifiable theories to orchestrate parent agents and specialist subagents, achieving state-of-the-art performance on 22 of 25 ARC-AGI-3 games with significantly reduced costs while ensuring hypothesis revision remains explicit, auditable, and cost-effective.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to teach a robot how to play a brand-new video game, but you can't give it a rulebook, a map, or even a hint about what the goal is. The robot has to figure out everything just by watching the screen, pressing buttons, and seeing what happens. This is the world of "executable world models," a fancy way of saying: "Let's write a computer program that acts like a simulation of the game so the robot can practice inside its own head before it actually moves."
Usually, scientists think that if this simulation matches the real game perfectly—frame for frame, pixel for pixel—the robot must be smart and ready to win. But here's the catch: a simulation can be a perfect liar. It might match the history of what happened by accident, like a broken clock that happens to show the right time twice a day, while completely misunderstanding why things happen. If the robot relies on this "perfect" but wrong simulation, it might try to solve a puzzle that doesn't exist or miss a hidden trap. The big question in this field is: How do we build an agent that knows when its own internal map is wrong and knows how to fix it without getting stuck in a loop of confusion?
This paper introduces a new system called THEA (Typed Hypothesis-driven Epistemic Agents) to solve exactly that problem. Think of THEA not as a single super-smart robot, but as a strict, organized scientific team running a game night. The "Parent" is the main player who holds the controller and keeps the big picture in mind. But THEA also has a team of specialized "detectives" (called specialists) who only jump in when things go wrong.
Here is how it works: The Parent plays the game and constantly updates its internal "theory" (a computer program) of how the game works. As long as the theory predicts the game's outcome correctly, the Parent keeps playing. But if the Parent makes a mistake, or if the game gets too weird, a detective is called in. These detectives don't touch the controller; they just look at the evidence. One detective might check if the robot is confusing a decoration with a real game mechanic. Another might try to prove that two different theories about the game are actually different by suggesting a specific "test move" that would give different results for each theory.
The paper tested this system on 25 different abstract puzzle games (part of the ARC-AGI-3 benchmark). The results were impressive: THEA managed to fully solve 22 out of 25 games. On average, it was incredibly efficient, using only about 92.63 "Relative Human Action Efficiency" (a score where 100 is perfect human efficiency). Perhaps most surprisingly, it did this while costing 19.52% less in computer processing power than the previous best system, which only solved 20 games.
However, the paper is very careful not to claim THEA is a magic bullet. It explicitly shows that this "detective team" approach isn't always cheaper or faster. For the 3 games THEA couldn't fully solve, and for some of the easier games it did solve, the extra layer of checking sometimes cost more time and money without changing the final result. The authors found that the system shines when the game is tricky and the robot needs to change its mind, but it can be overkill for simple tasks.
The paper also shares some "failure stories" to prove it's not just bragging. In one case, the robot had a perfect simulation of a game but was still failing because it was looking at the wrong part of the screen. In another, the robot figured out the correct rules, but only after it had already used up all its moves, proving that being right too late is just as bad as being wrong.
Ultimately, THEA teaches us that for an AI to be truly smart, it needs to treat its own understanding of the world as a "hypothesis"—a guess that can be proven wrong. By making the process of checking and fixing those guesses visible, auditable, and controlled by a central "Parent," the system can avoid getting stuck in its own mistakes. It's a step toward building AI that doesn't just memorize patterns, but actually understands the rules of the game well enough to know when it needs to rewrite the rulebook.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.