CARD: Cluster-level Adaptation with Reward-guided Decoding for Personalized Text Generation
CARD personalizes generation by grouping users first, then applying individual preference signals only during decoding.
The paper proposes cluster-specific LoRA adapters for shared writing patterns, plus lightweight user preference vectors for differences inside each group. Its preference learning compares user-written text against cluster-level generations without manual labels. At inference, the base model stays frozen while low-rank logit corrections steer the output. The authors report stronger results on LaMP and LongLaMP, with better efficiency and scalability. HF Daily Papers' note
The paper proposes cluster-specific LoRA adapters for shared writing patterns, plus lightweight user preference vectors for differences inside each group. Its preference learning compares user-written text against cluster-level generations without manual labels. At inference, the base model stays frozen while low-rank logit corrections steer the output. The authors report stronger results on LaMP and LongLaMP, with better efficiency and scalability. HF Daily Papers' note
score 4