Personalized value alignment has become crucial as large language models (LLMs) are expected to serve diverse user preferences. Existing approaches typically enforce a static value profile across all prompts, ignoring that the importance of value dimensions varies widely with context. Inspired by Lewin's Field Theory, we treat personal values as priors and context‑dependent preferences as posteriors.
We introduce BaCVA, an inference‑time Bayesian Context‑aware personalized Value Alignment method. It approximates posterior preferences by integrating static personal values with scenario‑specific value salience. The procedure consists of: (1) estimating contextual value salience from generally normative responses; (2) employing a dual‑view personalization module that infers posterior preferences from both personal‑value and scenario‑driven perspectives. The Bayesian relationship is expressed as $$P(\text{preference}\mid\text{context}) \propto P(\text{preference}) \times P(\text{context}\mid\text{preference})$$, where the prior $P(\text{preference})$ comes from the user's static values and the likelihood $P(\text{context}\mid\text{preference})$ is provided by contextual salience.
Operating at inference time, the framework adapts to varying contexts, improves alignment accuracy, and leverages prior values for data efficiency. Extensive benchmark experiments demonstrate BaCVA's superiority over strong baselines across multiple metrics.
Review