When an LLM is shown a random walk on a grid, it begins to model it. (https://arxiv.org/abs/2501.00070). And if you train a model on a game it will model its board states (https://www.neelnanda.io/mechanistic-interpretability/othello). But if you tell the model the rules of a game, then have it play one game, is there such a representation? How many games does it have to see? How complicated can the rules get?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-incontext-probable-game-2026,
author = {Holtzman, Ari},
title = {In-Context Probable Game States?},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/F229fVlcwYX3X6Lyluxp}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!