VAC Framework Uses Natural Language Feedback for Personalized QA

Researchers have introduced VAC, a novel framework for personalized question answering that replaces scalar reward signals with natural language feedback (NLF). Current personalization methods for large language models (LLMs) rely on retrieval-augmented generation (RAG) and reinforcement learning with scalar rewards, which the authors argue provide weak, non-instructive feedback that limits learning efficiency. VAC generates NLF conditioned on user profiles and question narratives, offering richer and more actionable supervision. This allows the policy model to iteratively refine outputs and internalize effective personalization strategies. The work is described in a paper on arXiv (2508.10695) and aims to improve both effectiveness and user satisfaction in information-seeking tasks.