BaCVA Brings Context-Aware Personalization to LLM Value Alignment

Summary

Personalized value alignment aims to make large language models accommodate different users' preferences, but many existing methods apply a fixed value profile across prompts. This paper introduces BaCVA, an inference-time Bayesian Context-aware personalized Value Alignment method that treats personal values as priors and context-dependent preferences as posteriors. The approach is motivated by Lewin's Field Theory, which describes behavior as shaped by both personal dispositions and situational constraints. BaCVA first estimates which value dimensions are salient in a given scenario from generally normative responses. It then uses a dual-view personalization module to infer posterior preferences from both the user's personal values and the scenario itself. The authors argue that this formulation produces more accurate and adaptive alignment while improving data efficiency by reusing prior value information. Experiments on multiple benchmarks reportedly show that BaCVA outperforms strong baseline methods.