CARD: Cluster-level Adaptation with Reward-guided Decoding for Personalized Text Generation
Yutong Song, Jiang Wu, Weijia Zhang, Chengze Shen, Shaofan Yuan, Weitao Lu, Jian Wang, Yu Wang and 2 others
University of California, Irvine · Independent Researcher · University of Amsterdam · TikTok
Published September 20, 2026 · on the feed September 28, 2026 · ▲ 7 · arXiv 2601.06352
The abstract
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. CARD first clusters users according to shared stylistic patterns and learns group-specific LoRA adapters, enabling robust generalization and strong low-resource performance. To capture individual differences within each cluster, we propose an implicit preference learning mechanism that contrasts user-authored text with cluster-level generations, allowing the model to infer user-specific style preferences without manual annotation. At inference time, CARD injects personalization exclusively at decoding via lightweight user preference vectors and low-rank logit corrections, while keeping the base model frozen. Experiments on the LaMP and LongLaMP benchmarks show that CARD achieves superior generation quality compared to baselines, while significantly improving efficiency and scalability for practical personalized text generation.
The Brief
2:04A short podcast summary of the paper. Written and voiced by AI.
Transcript
CARD. Cluster-level Adaptation with Reward-guided Decoding for Personalized Text Generation. By Yutong Song and nine others, from the University of California, Irvine, Independent Researcher, and two other institutions. Published on the twentieth of September. Its code repository is a placeholder: a readme, and nothing to run yet.
Making a large language model write the way one particular user writes pulls in two directions at once. Fine-grained personalization wants something close to a model per person. Scalable deployment wants one model serving everybody. Adapting to individuals remains hard precisely because of that tension, and a system that leans fully toward either side gives up the other.
CARD answers with a hierarchy, reaching personalization through progressive refinement rather than in a single step. The coarse level is the group. Users are clustered according to the stylistic patterns they share, and a group-specific LoRA adapter is learned for each cluster. Sharing an adapter across similar users is what gives robust generalization, and it is also why the approach stays strong when an individual user has little data.
A cluster still hides individual differences, so the second level goes finer. The authors propose an implicit preference learning mechanism that contrasts text a user actually wrote with what the cluster-level model generates. The gap between the two lets the model infer that user's style preferences, with no manual annotation required. Then, at inference time, personalization enters at one place only: decoding. Lightweight user preference vectors and low-rank logit corrections adjust the output, while the base model itself stays frozen throughout.
On the LaMP and LongLaMP benchmarks, CARD achieves superior generation quality compared to the baselines. It does so while significantly improving efficiency and scalability, which is the other half of the tension the paper started from.
Read as a whole, the design divides the labour by granularity. Clusters carry what users have in common, a contrast against the cluster recovers what makes each one different, and all of the per-user part lives in small decoding-time adjustments to a model that never changes.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
One model for everyone sounds like no one. One model per person never ships. Start there.
You run a writing assistant used by thousands of people, and each of them wants replies that sound like their own writing: their slang, their emojis, their rhythm.
The road ahead
- 1 · The problem ~7 min A writing assistant must sound like each of its many users, but a model per person cannot scale and one shared model sounds like nobody.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Share one adapter per group of similar writers, learn each person's quirks by comparing their text with the group's, and apply them only while decoding.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Cluster users, train one LoRA adapter per cluster, build preference pairs against the cluster output, then train user vectors that steer decoding.
- 6 · The experiment ~6 min You’ll guess: CARD personalizes only at decoding, keeping the base model frozen. On LaMP and LongLaMP, did it give up generation quality to gain efficiency?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.