In cognitive modeling, a central question is whether large language models behave like humans on linguistic and non‑linguistic tasks. Large reasoning models (LRM) differ from standard LLMs by being trained with reinforcement learning from verifiable rewards, which pushes them toward correct solutions rather than merely preference‑aligned replies. A recent study (de Varda et al., 2025) compared human reaction times with model reasoning traces to assess thinking cost in humans and LRMs. This work isolates the comparison by using abductive reasoning: unlike deductive problems, its difficulty cannot be inferred from formal structure and offers no shortcuts for a model to fake search, providing a firmer basis for claiming shared effort. Results show a clear alignment of reasoning effort between LRMs and humans, and both groups tend to make similar mistakes. Moreover, decoding strategies that let models explore multiple reasoning paths increase the alignment of thinking cost across the three tested models.
Review