Think-Aloud Data Improves AI-Discovered Cognitive Models
A new study on arXiv (2605.05091) demonstrates that incorporating think-aloud verbal traces into automated cognitive model discovery significantly enhances predictive accuracy and alters model structure. Using large language models, researchers found that models built with think-aloud data outperformed those relying solely on behavioral data in risky decision-making tasks. For 69.4% of participants, the discovered models shifted from Explicit comparator to Integrated utility structural classes, indicating that process-level language data systematically reshape model discovery beyond behavioral trajectories alone.
Key facts
- Study uses think-aloud traces as additional data constraint in automated cognitive model discovery.
- Applied to risky decision-making domain.
- Models with think-aloud data achieve significantly improved predictive performance on held-out data.
- For 69.4% of participants, discovered models belong to different structural classes than from behavior alone.
- Shift from Explicit comparator to Integrated utility structural class.
- Process-level language data improve model fit and reshape model structure.
- Published on arXiv with ID 2605.05091.
- Research highlights under-determination of models from behavioral data alone.
Entities
Institutions
- arXiv