Researchers have introduced CARD, a hierarchical framework for personalized text generation that seeks to reconcile two competing demands: fine-grained adaptation to individual users and deployment at scale. The work, published as an AI research paper, describes a progressive refinement approach rather than a single monolithic fine-tuning step.
CARD begins by clustering users according to shared stylistic patterns and training group-specific LoRA adapters. According to the authors, this cluster-level stage supports robust generalization and performs well in low-resource settings, where per-user data is scarce.
To capture differences among users inside the same cluster, the framework introduces an implicit preference learning mechanism. It contrasts text written by a user with generations produced at the cluster level, allowing the model to infer individual style preferences without requiring manual annotation.
At inference time, personalization is injected exclusively during decoding through lightweight user preference vectors and low-rank logit corrections, while the base model remains frozen. This design keeps the underlying model unchanged and limits the cost of serving many personalized variants.
Experiments on the LaMP and LongLaMP benchmarks reportedly show that CARD delivers better generation quality than the compared baselines, alongside notable gains in efficiency and scalability for practical personalized text generation.