Revisiting the Othello World Model Hypothesis
This addresses the problem of understanding if language models can learn structured world representations, which is incremental as it builds on previous studies with more models and comprehensive probing.
The paper investigates whether language models can learn world models by testing seven models on the Othello board game, finding they achieve up to 99% accuracy in unsupervised grounding and induce board layouts, providing stronger evidence for the hypothesis than prior work.
Li et al. (2023) used the Othello board game as a test case for the ability of GPT-2 to induce world models, and were followed up by Nanda et al. (2023b). We briefly discuss the original experiments, expanding them to include more language models with more comprehensive probing. Specifically, we analyze sequences of Othello board states and train the model to predict the next move based on previous moves. We evaluate seven language models (GPT-2, T5, Bart, Flan-T5, Mistral, LLaMA-2, and Qwen2.5) on the Othello task and conclude that these models not only learn to play Othello, but also induce the Othello board layout. We find that all models achieve up to 99% accuracy in unsupervised grounding and exhibit high similarity in the board features they learned. This provides considerably stronger evidence for the Othello World Model Hypothesis than previous works.