Neurocognitive Parallels: Mapping LLM Belief Dynamics to Human Brain Activation

Research Question: Do belief updates in LLMs, as characterized by the Bayesian model, exhibit representational or temporal similarities to neural activation patterns observed in human cognitive belief updating?

Hypothesis: There will be measurable correspondence between LLM activation changes during belief updates and human neural signatures of belief revision (e.g., as identified in fMRI studies of cognitive control and ToM reasoning).

Experiment Plan: - Use datasets like those from Peraza et al. (2025) or TimeToM (Hou et al., 2024) to find tasks where human cognitive belief updates are well-characterized.

• Run LLMs on analogous tasks with both prompt and activation interventions.
• Apply representational similarity analysis (RSA) between LLM internal states and brain activation maps.
• Look for alignment in the dynamics and localization of belief update processes.

References: ['Bigelow, E. J., Wurgaft, D., Wang, Y., Goodman, N. D., Ullman, T. D., Tanaka, H., & Lubana, E. (2025). Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering.', 'Peraza, J. A., Kent, J. D., Nichols, T., Poline, J., de la Vega, A., & Laird, A. (2025). NiCLIP: Neuroimaging contrastive language-image pretraining model for predicting text from brain activation images. bioRxiv.', "Hou, G., Zhang, W., Shen, Y., Wu, L., & Lu, W. (2024). TimeToM: Temporal Space is the Key to Unlocking the Door of Large Language Models' Theory-of-Mind. Annual Meeting of the Association for Computational Linguistics."]