To understand how a system decides, we do not always need to ask it for a persuasive explanation. We can watch what it chooses in a controlled setting: follow a familiar pattern, or spend effort to learn something new?
Research on active inference and reasoning used a simple task involving the movement of several balls. People did not only pursue the most likely outcome; at times, they chose information that could clarify the structure of the problem. That kind of design brings reasoning closer to observable behavior.
This is not evidence that language models use the same mental process. The human experiment concerns a bounded task. What I find useful is the methodological question: instead of accepting an explanation offered after a decision, can we design small choices that distinguish between competing hypotheses?