Researchers led by Evelina Fedorenko studied how humans and the latest reasoning-capable large language models (LLMs) approach complex tasks and found a notable parallel, tells MIT News. In their experiment, human volunteers and AI models were given the same set of problem types, ranging from straightforward arithmetic to more abstract tasks such as the ARC challenge (which involves transforming colored grids).
Humans were timed on how long it took to solve each problem; for the models, the metric was the number of internal tokens generated during the chain of reasoning (rather than raw processing time, which depends on hardware). The main finding: those classes of problems that took humans longer also required the models to generate more tokens. In other words, harder problems cost more internal work in both brains and machines.
The authors emphasize that this convergence is not because AI was deliberately built to mimic human thinking; as Fedorenko states, “People who build these models don’t care if they do it like humans. They just want a system that will robustly perform …” While the results don’t suggest AI is thinking exactly like humans, or generating solutions with similar wiring or neural mechanisms, they do hint that the “cost of thinking” can be a useful bridge between cognitive neuroscience and AI-research metrics.
The study raises further questions: Do models and humans rely on similar internal representations? Can AI handle tasks that involve real-world knowledge not embedded in its training texts? The researchers plan to explore these in follow-up work.
For engineers, AI designers, and cognitive scientists alike, this work ties together two disparate domains, i.e., neuroscience and machine learning, via a shared metric of effort, offering a window into the architecture of complex reasoning across biological and artificial systems.