How Many Exposures Does a World Model Need? Learning Unseen Game Mechanics at Test Time

How Many Exposures Does a World Model Need? Learning Unseen Game Mechanics at Test Time. By Ku Onoda and three others, from Third Intelligence and the University of Tokyo. Published on the ninth of November.

World models let agents plan and learn from imagined experience. In an unfamiliar environment, though, a world model first has to learn how that environment works, and it has only limited interaction to learn from. The work frames this with two terms. A mechanic is a rule for how an action or an interaction changes part of the environment. Each occurrence of a mechanic that the model observes counts as one exposure. The question is how many exposures a world model needs before it can predict the effects of a new environment's mechanics. For half the mechanics in this study, the answer is about eighty, within a single play of the game, provided the model keeps updating as it plays.

The setting is ARC-AGI-3, a benchmark of interactive games defined by their programs. From those programs the authors build MechEval, which labels sixty-nine recurring mechanics across twenty games. It scores each prediction on the cells the mechanic changes, at each exposure, so accuracy is measured where the mechanic actually acts.

The test bed is WARC, short for World model for ARC-AGI-3. It is a thirty-five million parameter vision transformer, pretrained on other games, that updates all of its parameters after each transition. On held-out games, WARC learns to predict four in five mechanics with at least eighty percent accuracy on the cells they change, and keeps that accuracy to the end of the recordings. Half of them reach that level within about eighty exposures. Without updates, only one in seven gets there at all.

In closed-loop play on the public games, WARC-Agent pairs WARC with a fixed graph explorer. It triples the explorer's official efficiency score, from zero point one five to zero point four six, mostly through test-time training. On the hidden games of the Kaggle leaderboard, the entry scores zero point two.

These numbers have edges. Mechanics are labelled from the game programs of ARC-AGI-3, so the measure depends on environments whose rules can be read off their source. The learner is one particular test bed, a pretrained transformer that updates every parameter after every transition, and the counts describe that learner rather than world models in general. Even within it, one mechanic in five never reaches the eighty percent mark. The agent results pair the model with a fixed explorer, and its score on the hidden games sits below its score on the public ones.

Updating is what separates the two outcomes. With test-time training, most mechanics become predictable inside one playthrough. Without it, very few do. MechEval turns that difference into something countable: exposures, one at a time, each scored on the cells that change.