[2603.22294] Efficient Embedding-based Synthetic Data Generation for Complex Reasoning Tasks

Computer Science > Machine Learning arXiv:2603.22294 (cs) [Submitted on 15 Mar 2026] Title:Efficient Embedding-based Synthetic Data Generation for Complex Reasoning Tasks Authors:Srideepika Jayaraman, Achille Fokoue, Dhaval Patel, Jayant Kalagnanam View a PDF of the paper titled Efficient Embedding-based Synthetic Data Generation for Complex Reasoning Tasks, by Srideepika Jayaraman and 3 other authors View PDF HTML (experimental) Abstract:Synthetic Data Generation (SDG), leveraging Large Language Models (LLMs), has recently been recognized and broadly adopted as an effective approach to improve the performance of smaller but more resource and compute efficient LLMs through fine-tuning. A key challenge in SDG is ensuring the quality and diversity of the generated data. In this paper, we analyze the diversity and distribution of generated data in the embedding space, and demonstrate a strong correlation between the density of examples within a specific neighborhood and the accuracy of predictions on examples drawn from that region. Building on this insight, we present a targeted pipeline for embedding-based sampling that enhances data diversity and consistently improves performance across several benchmarks. Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI) Cite as: arXiv:2603.22294 [cs.LG]   (or arXiv:2603.22294v1 [cs.LG] for this version)   https://doi.org/10.48550/arXiv.2603.22294 Focus to learn more arXiv-issued DOI via DataCite Submission history From:...