Thinking Aloud Helps Decode How We Take Risks express
Every time you mull over a risky move, whether to invest in stocks or hold off, an inner dialogue runs through your head. Usually, scientists try to reconstruct that dialogue from the final choice alone, as if a detective were reconstructing a conversation between two people from a photo of the meeting spot. Such a picture is often insufficient: the same actions can stem from very different deliberations.
The authors of the study proposed recording the dialogue itself. Participants voiced their thoughts while playing a game where they had to choose between a sure win and a risky lottery. These recordings, along with the choices, were then fed to a program that seeks the best explanation for behavior. The result was striking: models trained on such a transcript became noticeably more accurate at predicting people's decisions on new examples. Moreover, for 69% of participants, the model architecture itself changed. Without the prompts, the program built a model of "explicit comparison"—as if the person puts the options on the scales and sees which side tips. With the verbal data, the model shifted to "integrated utility"—as if all the pros and cons fused into a single sense of value, and the decision emerged from that fusion.
So, the words spoken in moments of doubt are not just noise. They carry the imprint of how exactly we arrive at a decision, and that imprint can be decoded. This approach promises not only more accurate models but also a way to peek into the mechanisms of thought that are invisible from actions alone.
Original abstract
as written by the authors arXiv:2605.05091
Computational cognitive models discovered using large language models have so far relied solely on behavioral data. However, it is well-known that models produced from the behavioral trajectory alone are typically under-determined. In this work, we explore the use of Think Aloud traces as an additional form of data constraint during automated model discovery. When applied to the domain of risky decision-making, we find that the models discovered with think-aloud achieve significantly improved predictive performance on held-out data. Additionally, we find that the discovered models belong to different structural classes than those discovered from behavior alone for the majority of participants (69.4\%), specifically, it shifts from Explicit comparator towards Integrated utility. These results suggest that process-level language data not only improve model fit, but also systematically reshape the structure of the discovered cognitive models, enabling the identification of mechanisms that are not recoverable from behavior alone.
Think-aloud was originally used to study how experts solve problems — for example, how chess players choose moves. It turns out that this method helps peer into the "black box" of the mind even for simple choices between money options.