Probing Emergent World Representations in Go Life-and- Death Problems

Large language models (LLMs) have demonstrated remarkable capabilities in learning complex tasks purely from sequential data. To explore whether such models can internalize strategic world representations, We investigate whether generative transformer models can learn structured world representations from sequential data. Using the domain of Go life-and-death problems as a controlled micro-world, we train a GPT-style generative model to predict moves from serialized board states. Focusing on localized life-and-death (tsumego) scenarios, we train the model to predict valid next moves from serialized board states without providing any explicit Go rules or strategic supervision. Probing the model’s internal activations reveals structured representations aligned with liberties, eyes, and tactical group status. To interpret these representations, we introduce the Multi-Aspect World Probe (MAWP), a modular probing framework that disentangles tactical concepts into orthogonal dimensions. We further apply interventional techniques to manipulate internal representations and causally evaluate their impact on model predictions. Our results show that the proposed model achieves 94.7% accuracy in sequence correctness and 92.1% in outcome validity on life-and-death tasks. This work extends interpretability research into spatially structured domains and offers tools for understanding decision-making in sequence models.