CARD: Cluster-level Adaptation with Reward-guided Decoding for Personalized Text Generation
CARD personalizes LLMs with cluster LoRA adapters and decoding-time preference vectors, beating LaMP and LongLaMP baselines.
Researchers present CARD, a hierarchical framework that clusters users by writing style and trains group-specific LoRA adapters. An implicit preference step contrasts user-authored text with cluster-level generations so the model infers style without manual labels. At inference, personalization uses lightweight preference vectors and low-rank logit corrections while the base model stays frozen. On LaMP and LongLaMP, CARD reports higher generation quality than baselines plus better efficiency, scalability, and low-resource performance.
- Clusters users by style and learns group-specific LoRA adapters.
- Infers preferences by contrasting user text with cluster generations.
- Injects personalization only at decoding; the base model stays frozen.
- Reports better quality and efficiency than baselines on LaMP and LongLaMP.
From the original article
1 min read at huggingface.co
Adapting large language models to individual users remains challenging due to the tension between fine-grained personalization and scalable deployment. We present CARD, a hierarchical framework that achieves effective personalization through progressive refinement.
By Yutong Song, Jiang Wu, Weijia Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.