Personality and Bias: Probing the Limits of Bayesian Interventions for Social Traits

Research Question: How effective is the Bayesian belief update framework in modulating LLM behavior for social traits (empathy, personality), and can it reduce demographic biases compared to other control methods?

Hypothesis: Bayesian interventions (both prompt and activation) offer granular control over social trait expression, but may not fully eliminate pre-existing demographic biases without additional counter-bias interventions.

Experiment Plan: - Use datasets and frameworks from Handa et al. (2025) and Gabriel et al. (2024) to design tasks requiring empathy or personality alignment.

• Apply Bayesian-model-driven prompt and activation interventions.
• Measure trait expression, reasoning capability, and demographic bias using established benchmarks (e.g., MMLU, BBQ).
• Compare against ICL, PEFT, and direct mechanistic steering.

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.', 'Handa, G., Wu, Z., Koshiyama, A., & Treleaven, P. C. (2025). Personality as a Probe for LLM Evaluation: Method Trade-offs and Downstream Effects. arXiv.org.', 'Gabriel, S., Puri, I., Xu, X., Malgaroli, M., & Ghassemi, M. (2024). Can AI Relate: Testing Large Language Model Response for Mental Health Support. Conference on Empirical Methods in Natural Language Processing.']