Conflicting Preference Reconciliation: A Game-Theoretic Approach to Multi-User LLM Alignment

While "HyPerAlign" (Garbacea & Tan, 2025) and "Personalized Soups" (Jang et al., 2023) tackle individual preference alignment, real-world LLM deployments often serve groups—families, teams, classrooms—where preferences may conflict. This research proposes to model multi-user alignment as a repeated game, where each user’s utility function (preferences) is explicitly modeled, and the LLM acts as a mediator, optimizing for socially fair or Pareto-optimal responses. Unlike prior work, which assumes either a single user or aggregates preferences, this approach would integrate concepts from cooperative game theory and negotiation (e.g., Nash bargaining), dynamically adjusting outputs based on evolving signals of consensus or dissent. The novelty lies in moving beyond simple preference aggregation toward principled, explainable reconciliation of conflicts, with empirical studies in domains like collaborative writing or group decision support. This could set a new standard for fairness and transparency in multi-user LLM interactions.

References:

1. HyPerAlign: Interpretable Personalized LLM Alignment via Hypothesis Generation. Cristina Garbacea, Chenhao Tan (2025). arXiv.org.
2. Personalized Soups: Personalized Large Language Model Alignment via Post-hoc Parameter Merging. Joel Jang, Seungone Kim, Bill Yuchen Lin, Yizhong Wang, Jack Hessel, Luke S. Zettlemoyer, Hannaneh Hajishirzi, Yejin Choi, Prithviraj Ammanabrolu (2023). arXiv.org.
3. Personalized Soups: Personalized Large Language Model Alignment via Post-hoc Parameter Merging. Joel Jang, Seungone Kim, Bill Yuchen Lin, Yizhong Wang, Jack Hessel, Luke S. Zettlemoyer, Hannaneh Hajishirzi, Yejin Choi, Prithviraj Ammanabrolu (2023). arXiv.org.