In July 2025, a groundbreaking study published in Nature unveiled an AI model named "Centaur." This innovative model, built on established large language frameworks and enhanced through psychological experiment data, was crafted to emulate human cognitive functions. Its performance across 160 diverse tasks, including decision-making and executive control, garnered significant attention, hinting at a future where AI could closely mimic human thought processes.
New Insights Challenge Previous Findings
However, a recent investigation published in National Science Open raises critical questions about Centaur's capabilities. Researchers from Zhejiang University contend that the model's success might stem from overfitting, suggesting it has not genuinely understood the tasks but rather identified patterns in its training data to produce expected responses.
To explore this hypothesis, the researchers devised new evaluation scenarios. For instance, they substituted the original multiple-choice prompts with a simple directive: "Please choose option A." If Centaur truly grasped the task, it would consistently select option A. Contrary to expectations, the model continued to provide "correct answers" based on its original dataset, indicating a reliance on learned statistical patterns rather than true comprehension.
This phenomenon resembles a student who excels by memorizing test formats without a solid understanding of the material, highlighting a significant gap in the model's ability to interpret the meaning behind questions.
Implications for AI Assessment
The findings from this study underscore the necessity for caution when evaluating the capabilities of large language models. While these systems can adeptly fit data, their "black-box" nature complicates the understanding of how they generate outputs. This opacity can result in issues such as hallucinations or misinterpretations, making thorough and varied testing crucial to ascertain whether a model possesses the skills it appears to exhibit.
The Core Challenge: Understanding Language
Despite being touted as a model that simulates cognitive functions, Centaur's primary limitation lies in its language comprehension. It struggles to grasp and respond to the underlying intent of questions. The study suggests that achieving genuine language understanding is one of the most significant hurdles in developing AI systems capable of fully modeling human cognition.
