Hugging Face Daily PapersYutong Song, Jiang Wu, Weijia ZhangSep 20, 20261 min readpaperadvanced

CARD: Cluster-level Adaptation with Reward-guided Decoding for Personalized Text Generation

Summary

CARD is a hierarchical framework for personalized text generation that clusters users for group-specific LoRA adapters and learns individual preferences implicitly. It injects personalization at decoding via lightweight vectors and low-rank logit corrections, achieving superior quality and efficiency compared to baselines.

CARD clusters users by stylistic patterns to learn group-specific LoRA adapters for robust generalization.
Implicit preference learning infers user-specific styles by contrasting user-authored text with cluster-level generations.
Personalization is injected exclusively at decoding time using lightweight user preference vectors and logit corrections.
The base large language model remains frozen, significantly improving efficiency and scalability for personalized generation.

Engineers building personalized LLM applications should care about CARD's approach to balancing fine-grained personalization with scalable deployment challenges.

8/10

llm machine learning lora fine tuning personalization decoding

Related reading

Fathom: Per-Query Read Depth for Sparse Decoding over Offloaded KV Caches arxiv.org
Fathom introduces a per-query read depth mechanism for sparse decoding over offloaded KV caches, allowing each query to adaptively decide how many bits of each key channel to read. This method significantly speeds up decoding for large language models with long contexts by reducing host memory traffic, achieving 1.67x faster GPU decoding on Qwen3-8B at one million tokens.
Hugging Face Daily Papersarxiv.orgVivek Kalyanarangan19d1 minpaper
Decompose Radicals, Then Reward: Fine-Grained Inspection for Accurate Chinese Text Rendering arxiv.org
The paper presents IDSpect, a reward that uses Ideographic Description Sequences to give fine‑grained feedback on Chinese character radicals for text‑to‑image models. Experiments show it improves structural quality and semantic alignment on standard benchmarks without extra inference cost.
Hugging Face Daily Papersarxiv.orgYazhen Xie, Xingsong Ye, Zhineng Chen7d1 minpaper
EvolvingAvatar: Interactive 3D Head Generation That Adapts as Conversations Unfold arxiv.org
EvolvingAvatar is a novel 3D head generation model that adapts to ongoing conversations using test-time training and a self-supervised dyadic context prediction objective. It improves conversational motion statistics by learning from user audiovisual input during interaction, reducing expression mismatch by up to 11.1%.
Hugging Face Daily Papersarxiv.orgJunjie Chen, Fei Wang, Kun Li8d1 minpaper
AdaTutoRank: Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Research arxiv.org
AdaTutoRank is a setwise document reranker for retrieval‑augmented generation that trains via Adaptive Tutoring Optimization, providing quality‑matched hints from a frozen policy snapshot. Across ten benchmarks it achieves state‑of‑the‑art performance with fewer retrieval calls.
Hugging Face Daily Papersarxiv.orgKailin Jiang, Lei Liu, Jian Xi10d2 minpaper
Text-to-meowdio models kmjn.org
Lobsterskmjn.org6dHN3 lobste.rs2
AutoRef: Harness Optimization for Agentic Multi-Reference Image Generation arxiv.org
AutoRef uses a coding LLM to automatically rewrite the harness that orchestrates multi‑reference image generation, keeping the generator frozen. The resulting harness lifts FLUX.2‑4B scores from 5.72 to 7.37 on MultiBanana and generalizes across models and benchmarks.
Hugging Face Daily Papersarxiv.orgYuta Oshima, Ku Onoda, Yusuke Iwasawa8d2 minpaper

CARD: Cluster-level Adaptation with Reward-guided Decoding for Personalized Text Generation

Summary

Related reading

Fathom: Per-Query Read Depth for Sparse Decoding over Offloaded KV Cachesarxiv.org

Decompose Radicals, Then Reward: Fine-Grained Inspection for Accurate Chinese Text Renderingarxiv.org

EvolvingAvatar: Interactive 3D Head Generation That Adapts as Conversations Unfoldarxiv.org

AdaTutoRank: Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Researcharxiv.org

Text-to-meowdio modelskmjn.org

AutoRef: Harness Optimization for Agentic Multi-Reference Image Generationarxiv.org

Fathom: Per-Query Read Depth for Sparse Decoding over Offloaded KV Caches arxiv.org

Decompose Radicals, Then Reward: Fine-Grained Inspection for Accurate Chinese Text Rendering arxiv.org

EvolvingAvatar: Interactive 3D Head Generation That Adapts as Conversations Unfold arxiv.org

AdaTutoRank: Learning to Rerank Document Sets via Adaptive Tutoring Optimization for RAG and Deep Research arxiv.org

Text-to-meowdio models kmjn.org

AutoRef: Harness Optimization for Agentic Multi-Reference Image Generation arxiv.org