Hanbo Xie, Ph.D. Student

Understanding Human Decision-Making and Learning from Think-Aloud Data with Large Language Models

The think-aloud protocol asks participants to verbalize their thoughts while they perform psychological tasks. Traditional work has mostly relied on behavioral outputs (often button presses) to infer latent cognitive processes. In many cases, candidate cognitive models are proposed and tested by researchers, which can limit the hypothesis space and introduce bias. By directly analyzing participants' verbal reports, we gain a richer and more direct view of cognition during task performance. However, most prior think-aloud research depends on manual coding by experts, which is labor-intensive, subjective, and difficult to scale.

Recent advances in LLMs make it possible to revisit this classic protocol with stronger computational tools. LLMs can help quantify, interpret, and even predict subsequent behavior from think-aloud language. Our work evaluates when and how these models can be used reliably, with the goal of building a more systematic and scalable framework for studying human thought processes.

Representative publications:

• Xie, H., Xiong, H., & Wilson, R. C. (2023). Text2Decision: Decoding Latent Variables in Risky Decision Making from Think Aloud Text. NeurIPS 2023 AI for Science Workshop.
• Xie, H., Xiong, H., & Wilson, R. C. (2024). From Strategic Narratives to Code-Like Cognitive Models: An LLM-Based Approach in A Sorting Task. First Conference on Language Modeling (COLM).
• Xie, H.†, Xiong, H. D., & Wilson, R. C. (2025). Rethinking Think-Aloud in the Age of Language Models. PsyArXiv. https://osf.io/preprints/psyarxiv/6ta3z_v1 (submitted).
• Zhang, Z.*, Xie, H.*, Baker, T., Peters, M., & Wilson, R. C. (2025). Linking strategies to think aloud in a stochastic learning task. In Proceedings of the Annual Meeting of the Cognitive Science Society.

Reverse Engineering Human Thoughts

Human thought is central to intelligence, yet it is difficult to define, measure, and model. A core challenge in both cognitive science and AI is to characterize thought processes across tasks, identify shared principles, and generalize those principles to make useful predictions. The difficulty is that thoughts are often implicit, while language is diverse and context-dependent. As a result, verbal reports are informative but still incomplete reflections of internal cognition.

Instead of focusing only on the forward direction (how thoughts generate behavior), this project emphasizes inverse inference: given observed behavior and related measurements, can we reconstruct plausible underlying thoughts? The broader goal is to build a stronger, more general bridge between behavior and cognition. This direction also supports a human-centered understanding of machine reasoning. If the computations of complex systems (e.g., AlphaGo-like models) can be approximated by human-trained explanatory models, we may be able to describe model reasoning in natural language that is useful for teaching, interpretation, and collaboration.

This project is currently in development and will be my primary research focus at Princeton University.

Representative publications:

• Zhu, J.-Q.*, Xie, H.*, Arumugam, D., Wilson, R. C., & Griffiths, T. L. (2025). Using reinforcement learning to train large language models to explain human decisions. arXiv preprint arXiv:2505.11614.

Understanding and Improving Artificial Intelligence Through Human Insights

This project examines AI through concepts from psychology and neuroscience. By comparing strengths and weaknesses of AI and human intelligence, we can design models that are both more capable and more interpretable. Beyond technical performance, I am interested in societal value: systems that support human decision-making, education, and collaboration. I also explore how people can learn from advanced AI models when we build the right frameworks to analyze and communicate their internal computations.

Representative publications:

• Pan, L.*, Xie, H.*†, & Wilson, R. C. (2025). Large Language Models Think Too Fast To Explore Effectively. arXiv preprint arXiv:2501.18009. NeurIPS 2025 Poster.

Accelerating Scientific Discoveries in Cognitive Science

The human mind is deeply complex. Although thoughts, emotions, and actions are part of everyday experience, formally describing and predicting cognition remains a major scientific challenge. Many cognitive theories are grounded in human intuition and then tested through experiments and computational models. These approaches are powerful, but they can remain constrained by the original hypothesis space.

In the AI era, there is an opportunity to rethink discovery pipelines in cognitive science and psychology. LLMs bring broad knowledge and strong inductive biases, and modern reasoning models can perform at levels that sometimes rival expert intuition. A central question for my work is whether we can build AI-assisted workflows that help discover new behavioral phenomena, generate computational models, and propose testable theories while reducing avoidable human bias. I view this as complementary to, not a replacement for, careful empirical research.

Representative publications:

• Xie, H., Xiong, H., & Wilson, R. C. (2023). Text2Decision: Decoding Latent Variables in Risky Decision Making from Think Aloud Text. NeurIPS 2023 AI for Science Workshop.
• Xie, H., Xiong, H., & Wilson, R. C. (2024). From Strategic Narratives to Code-Like Cognitive Models: An LLM-Based Approach in A Sorting Task. First Conference on Language Modeling (COLM).
• Zhu, J.-Q.*, Xie, H.*, Arumugam, D., Wilson, R. C., & Griffiths, T. L. (2025). Using reinforcement learning to train large language models to explain human decisions. arXiv preprint arXiv:2505.11614. ICLR 2026.
• Xie, H.*, & Zhu, J*. (2025, July 12). Centaur May Have Learned a Shortcut that Explains Away Psychological Tasks. https://doi.org/10.31234/osf.io/u7z4t_v1 (submitted).

Mentor and Committee

• Dr. Robert C. Wilson (Supervisor), Georgia Tech
• Dr. Thomas Griffiths, Princeton University
• Dr. Anna Ivanova, Georgia Tech
• Dr. Varma Sashank, Georgia Tech
• Dr. Travis Baker, Rutgers University

Social Cognition

• Prof. Xiaolin Zhou, East China Normal University
• Dr. Sensen Song, Central China Normal University
• Dr. Xiaoxue Gao, East China Normal University
• Dr. Yang Hu, East China Normal University
• Dr. Maarten Speekenbrink, UCL

Think Aloud

• Dr. Travis Baker, Rutgers University
• Dr. Megan Peters, UC Irvine
• Dr. Evan Russek, Princeton University
• Ionatan Kuperwajs, Princeton University

Large Language Models and Neural Networks

• Hua-Dong Xiong, Georgia Tech
• Dr. Jian-Qiao Zhu, Princeton University
• Dilip Arumugam, Princeton University

Mentees

• Zhenlong Zhang, Johns Hopkins University
• Lan Pan
• Yangtong Feng, Wash U St. Louis