Transformers Struggle with Tower of Hanoi Variant Despite Emergent World Models

A recent study published on arXiv (2608.07077) explores the challenges faced by large reasoning models (LRMs) in solving the Tower of Hanoi puzzle, specifically the flat-to-flat version where both the initial and target states can exist on any peg. Conducted by a team whose identity remains undisclosed, the research involves training small Transformers from the ground up using precomputed solution traces. Interpretability methods reveal that these models create an emergent world model, which is a geometrically accurate, linearly decodable depiction of the puzzle's state space, represented by the Sierpinski triangle. This model plays a crucial role in puzzle-solving. The study also examines two advanced reasoning models, Qwen3.6-27B and another unnamed model, which still encounter difficulties with the flat-to-flat variant, despite their advanced capabilities. The paper questions the assumption that current models genuinely 'think' or plan, indicating that their success in standard puzzles may be misleading. These findings underscore the limitations of AI reasoning and the need for enhanced planning skills. The abstract of the new arXiv paper emphasizes the disconnect between emergent world models and effective problem-solving in intricate situations.