Reproduce “Few-shot LLM Synthetic Data with Distribution Matching” 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
Jiyuan Ren, Zhaocheng Du, Zhihao Wen, Qinglin Jia, Sunhao Dai, Chuhan Wu, Zhenhua Dong
SynAlign guides LLM generation using selected real examples and summarized attributes, then selects generated examples with maximum mean discrepancy (MMD) in embedding space. On AGNEWS, BERT-base-uncased reaches 0.9475 accuracy with MMD selection, compared with 0.9429 under random selection. This comparison tests selection within the generation pipeline; it does not establish that every component is necessary or that synthetic data always helps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.