Nonlinear (small-MLP) probes reliably decode the full Othello board state from an 8-layer, 512-dim GPT-style transformer's internal activations at every move, and activation-patching interventions that flip the decoded board-state representation causally change the model's subsequent move predictions

Li, Hopkins, Bau, Viegas, Pfister & Wattenberg (2023, the original Othello-GPT paper) train an 8-layer, 8-head, 512-dim GPT-style transformer on sequences of legal Othello moves (synthetic self-play and championship human games), and show nonlinear (MLP) probes reliably decode the full board state from internal activations at each move -- a nonlinear emergent world-model representation, clearing scope on criterion (a). Causally, activation-patching interventions that flip the probe-decoded board-state representation change the model's subsequent move predictions accordingly (validated via move-legality and latent-saliency-map analyses; move-legality error rates ~0.01% synthetic, ~5.17% championship), clearing scope on criterion (b). This is the foundational world-model paper behind later linear reframings of the same phenomenon (Nanda et al. 2023, already in this map) and cross-architecture alignment follow-ups (Yuan & Sogaard 2025, already in this map). See [[nonlinear-world-model-decodability]].