A study combines fine-tuned large language models with mathematical models of choice to decode thousands of free-text explanations of human decisions in gambling tasks.

The research team notes open-access original work in PNAS and collaboration with TUD, SynoSys, and Neuroscience News.

The framework provides a scalable tool for studying behavior in complex real-world environments and could inform public policy, health, and technology adoption.

Participants described their thought processes after every round, yielding a rich corpus of verbal reports for analysis.

Findings show that decision-making strategies are dynamic and adapt to problem structure rather than being fixed traits.

A taxonomy of decision reasons (e.g., maximax, minimax, loss aversion) was developed and LLMs were used to tag which reasons appeared in each participant’s explanations.

LLMs’ classifications were validated by cross-referencing with objective mathematical models of actual choices, demonstrating high alignment and addressing concerns about hallucinations.