OptimSyn: Influence-Guided Rubrics Optimization for Synthetic Data Generation

Zhiting Fan, Ruizhe Chen, Tianxiang Hu, Ru Peng, Zenan Huang, Haokai Xu, Yixin Chen, Jian Wu, Junbo Zhao, Zuozhu Liu · 2026

Read OptimSyn: Influence-Guided Rubrics Optimization for Synthetic Data Generation by Zhiting Fan, Ruizhe Chen, Tianxiang Hu, Ru Peng, Zenan Huang, Haokai Xu, Yixin Chen, Jian Wu, Junbo Zhao, Zuozhu Liu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-00536.txt.

Description

Fan and colleagues propose OptimSyn, a way to optimise rubrics for synthetic training data using influence signals. Supervised fine-tuning data in medicine, law, and finance is scarce. Expert curation costs much, privacy binds tightly, and labels prove hard to keep consistent. Practice prompts a generator over domain documents and filters outputs with handcrafted rubrics. But rubric design stays expert-dependent, transfers poorly, and gets tuned through a brittle loop of guesswork. OptimSyn aims to supply the missing quantitative feedback, guiding rubric choice by influence on downstream performance.

Catalog record

CREATORZhiting Fan, Ruizhe Chen, Tianxiang Hu, Ru Peng, Zenan Huang, Haokai Xu, Yixin Chen, Jian Wu, Junbo Zhao, Zuozhu Liu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE77665
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.00536

Browse the full library · About the Tower of Babel