MetaOthello: A Controlled Study of Multiple World Models in Transformers
Published in arXiv preprint, 2026
Recommended citation: Chawla, A., Hall, G., & Lovato, J. (2026). MetaOthello: A Controlled Study of Multiple World Models in Transformers. arXiv preprint arXiv:2602.23164. https://arxiv.org/abs/2602.23164
We introduce MetaOthello, a suite of Othello game variants designed to investigate how transformers organize multiple world models within shared representations. We find that transformers trained on mixed-game data converge on a mostly shared board-state representation that transfers across variants.