Temporal Dynamics of Belief: Real-Time Tracking of LLM Belief Shifts with Interactive Tools

Research Question: How do LLM internal beliefs and output probabilities evolve over a sequence of interventions, and can interactive visualization tools provide actionable insights for users and developers?

Hypothesis: Real-time tracking will reveal non-linear, phase-transition-like shifts in beliefs—as predicted by the Bayesian model—and empower users to anticipate and guide LLM behavior more effectively.

Experiment Plan: - Develop an interactive dashboard that logs and visualizes belief changes (logits, internal activations) as a function of sequential prompt and activation interventions.

• Test the tool using multi-step reasoning and control tasks (inspired by FastMem, Zhu et al., 2024).
• Analyze the presence of predicted sigmoidal curves, additivity in log-belief space, and dramatic behavioral transitions.

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.', 'Zhu, J., Liu, S., Yu, Y., Tang, B., Yan, Y., Li, Z., Xiong, F., Xu, T., & Blaschko, M. B. (2024). FastMem: Fast Memorization of Prompt Improves Context Awareness of Large Language Models. Conference on Empirical Methods in Natural Language Processing.']