Large reasoning models form an accurate Sierpinski-triangle world model of the Tower of Hanoi at the prompt, then lose it during extended reasoning, and restoring it via activation steering improves planning accuracy.
Proceedings of the AAAI Conference on Artificial Intelligence , volume=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Transformers Struggle to Use Their Emergent World Models: Revisiting the Tower of Hanoi, and the Illusion of Thinking
Large reasoning models form an accurate Sierpinski-triangle world model of the Tower of Hanoi at the prompt, then lose it during extended reasoning, and restoring it via activation steering improves planning accuracy.