arXiv cs.AI· Liuji Chen, Zeyu Zhang, Xinyuan Zhang, Shuai Nie, Qiang Liu, Shu Wu, Liang Wang·· 14 小时前AI 评分31
GLASS：用全局稀疏先验与局部语义校准实现个性化文本生成
From Profiles to Steering Vectors: Global Sparse Priors and Local Semantic Calibration for Personalized Text Generation
AI 导读
GLASS 是一个无需训练的个性化文本生成框架，通过稀疏自编码器从历史回复中提取全局用户风格先验，并在聚类交互场景上构建局部对比风格向量，推理时联合注入模型不同层。在 LaMP 和 LongLaMP 上，GLASS 在 ROUGE 指标和 LLM-as-judge 评估中均优于基于检索、微调和激活引导的基线。分析显示 SAE 表示对主题和长度变化更稳健，风格信息与语义残留的解耦更好。
正文 · 原文
Abstract:Personalized text generation requires models to capture user-specific writing styles from historical data. Existing approaches based on retrieval, parameter-efficient fine-tuning, or activation steering either introduce inference and storage overhead or struggle to separate stylistic signals from semantic content. We propose GLASS, a training-free framework for personalized generation via Global-Local Activation Steering with Sparse priors. GLASS uses sparse autoencoders to extract a global user-style prior from historical responses and constructs local contrastive style vectors over clustered interaction scenarios. During inference, it jointly injects global and local vectors into different model layers, enabling context-aware personalization without retrieval or parameter updates. Experiments on LaMP and LongLaMP show that GLASS outperforms retrieval-, fine-tuning-, and steering-based baselines across ROUGE metrics and LLM-as-judge evaluations. Further analyses show that SAE-based representations are more robust to topic and length shifts, suggesting better disentanglement of stylistic information from semantic residue.
| Comments: | EMNLP2026 Oral, SAC Highlight Award |
| Subjects: | Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2607.21620 [cs.AI] |
| (or arXiv:2607.21620v2 [cs.AI] for this version) | |
| https://doi.org/10.48550/arXiv.2607.21620 arXiv-issued DOI via DataCite |
Submission history
From: Liuji Chen [view email]
[v1]
Mon, 1 Jun 2026 12:42:17 UTC (607 KB)
[v2]
Thu, 29 Oct 2026 14:05:24 UTC (647 KB)
来源：arXiv cs.AI · arxiv.org