This paper revisits the argument that the large-scale reasoning model (LRM) lacks reasoning ability, raised in Apple's paper "The Illusion of Thinking". Apple's paper argues that LRM is simply a probabilistic parrot, and presents Towers of Hanoi and River Crossing problems as examples. This paper reproduces and improves the experiments on these two problems, introducing step-by-step prompts and interactive dialogues to show that the conclusions of previous studies are exaggerated. We show that LRM's failure in Towers of Hanoi is due to cognitive limitations as well as output constraints, and its failure in River Crossing is due to an unsolvable problem setting. When limited to solvable problems, LRM easily solves large-scale problems with more than 100 agent pairs. Therefore, LRM is a probabilistic, reinforcement learning-tuned explorer in a discrete state space, and suggests that further detailed analysis is needed for the development of symbolic and long-term reasoning.
