Constructive Belief Shifts: Leveraging Context-Driven Adaptation for Personalized Educational Feedback

Research Question: Can controlled, context-accumulating belief shifts in language models be harnessed to provide more effective, personalized educational feedback?

Hypothesis: When LMs are designed to shift beliefs in response to accumulating student context, their feedback becomes more tailored and effective, leading to improved student engagement and learning gains.

Experiment Plan: Develop an LM-based tutoring system that accumulates student data (essays, Q&A) over time, with an explicit mechanism for belief adaptation (e.g., on writing proficiency, content understanding). Conduct a classroom or simulated study comparing adaptive vs. static LM feedback. Measure educational outcomes (learning gains, engagement), and analyze whether constructive belief shifts correspond to better personalization and results. Investigate risks of maladaptive shifts (e.g., reinforcing misconceptions), and test guardrails.

References: 1. Warr, M., Oster, N., & Isaac, R. (2024). Implicit Bias in Large Language Models: Experimental Proof and Implications for Education. Social Science Research Network. 2. Geng, J., Chen, H., Liu, R., Horta Ribeiro, M., Willer, R., Neubig, G., & Griffiths, T. L. (2025). Accumulating Context Changes the Beliefs of Language Models.