Towards Personalized Parameter Generation via Data-Conditioned Mapping

Ziheng Qin, Dongwen Tang, Mengxuan Wu, Yuhao Zhou, Zhiyuan Liang, Wangbo Zhao, Zekai Li, Panpan Zhang, Peihao Wang, Xuanlei Zhao, Zhaoyang Zeng, Tianlong Chen, Zhangyang Wang, Yang You, Kai Wang

Keywords: Parameter Generation

Abstract: Fine-tuning is the dominant strategy for adapting pre-trained models. However, it requires bulky gradient computation and model updates, which prevent real-time personalization. Even efficient variants such as LoRA incur a non-negligible latency and computation overhead. We explore a radically different approach: instead of training, we generate model parameters conditioned on data. Inspired by parameter generation via diffusion, we introduce a data-conditioned parameter generator that instantly produces personalized weights. In few-shot settings, our method outperforms LoRA in both performance and efficiency, making adaptation in less than one second. This demonstrates a feasible path toward practical, real-world, real-time model personalization.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 24406