A new AI agent skips the rulebook and writes its own, then beats the game anyway.
Researchers built Twin, a test-time world-model inference system in which a frontier coding agent writes an executable simulation of an unfamiliar game's rules and goal, using nothing but observation and interaction, not a hand-coded model built for that specific task. It checks its own work in a sandboxed twin of the game: the harness blocks any action until the program can reproduce every past observation, and each mismatch becomes a counterexample that patches the model. Tested on ARC-AGI-3, a benchmark of grid-based games with hidden rules, Twin cleared 179 of 183 levels (97.8%). On the 179 levels it cleared, it played more efficiently than human first-timers on 158 of them (88.3%), and it figured out the goal before ever seeing a reward signal on 156 of those levels (87.2%); the rest it solved by searching for the goal directly.
ARC-AGI-3 scores agents from 0 to 100 on a mix of completion and efficiency, benchmarked against humans playing each game for the first time. The base model alone scores 7.8% on that scale. Bolt on a generic off-the-shelf harness and it climbs to 61.1%. Give the same base model Twin's self-written world model and the score reaches 93.3%, with the system clearing 23 of 25 games in that run. The jump from 61% to 93% is the real story: it is not more raw model capability doing the work, it is the discipline of forcing an agent to write down and test a falsifiable theory of the game before it acts.
The paper's own admission is the honest part: building a usable world model turned out to be the easy half. Guessing what the game actually wants you to do is still the hard one.