Reproduce “UtilGen: Utility-Centric Generative Data Augmentation with Dual-Level Task Adaptation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiyu Guo, Shuo Yang, Yiming Huang, Yancheng Long, Xiaobo Xia, Xiu Su, Bo Zhao, Zeke Xie and 1 more
Researchers from Harbin Institute of Technology, NUS, and collaborators developed UtilGen, a framework that optimizes synthetic data generation by directly adapting to a downstream task's utility requirements. This approach outperforms prior generative augmentation methods, yielding an average accuracy of 37.17% using only synthetic data and 62.04% when combined with real data across various classification benchmarks.
Key insights
- Optimizing data generation directly for task-specific utility, as opposed to solely visual characteristics, leads to demonstrably better downstream model performance.
- Utility-centric synthetic data can achieve higher accuracy than models trained exclusively on real data, especially with increased synthesis ratios.
- A dual-level approach, combining generative model fine-tuning with instance-specific prompt and noise optimization, is essential for maximizing synthetic data utility and diversity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.