Elamur Research⌕
☰
home / papers / 2026 / The Data Frontier for Large Language Models: Selection, Synthesis, and Tools

The Data Frontier for Large Language Models: Selection, Synthesis, and Tools

Paper · 2026 · Lijun Wu, Wentao Zhang and Conghui He

Summary

This is a tutorial abstract, not an empirical study, so it reports no experiments or benchmark figures. It argues that LLM progress now depends more on training data (its quality, variety, and mixture) than on sheer volume, and that collecting and curating such data remains expensive and difficult. The presenters, from Shanghai AI Laboratory and Peking University, offer a structured overview of current data-centric research in two main blocks. The first covers selection: extracting the most useful examples from large, noisy collections. Examples include LLM-as-a-judge for automatic quality scoring and active learning for efficient use of annotation budgets. The second covers synthesis: generating training data artificially. This ranges from Chain-of-Thought reasoning traces to multi-agent pipelines that turn small seed sets into diverse instruction data. A final segment surveys open-source tools and platforms that support these workflows. The aim is to give practitioners a principled, actionable framework for building specialized and aligned models through better data.

Key points

  • Published in the Proceedings of ACL 2026 (64th Annual Meeting), Volume 5: Tutorial Abstracts, pages 3–4. It introduces no new corpus, model, or results.
  • The selection part covers LLM-as-a-judge for quality filtering and active learning for reducing annotation cost.
  • The synthesis part covers Chain-of-Thought trace generation and multi-agent workflows that build instruction data from raw seeds.
  • The closing part surveys open-source data tooling. The abstract names no specific tools, but the presenters' bios list involvement in DataFlow, MinerU (created by Conghui He), and Angel.
  • Conghui He's team curates training datasets for InternLM and InternVL. Wentao Zhang leads the Data-centric AI (DCAI) group at PKU.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksComputer Vision and Pattern Recognition · 3,486 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,894 works

Cite

Cite as Lijun Wu, Wentao Zhang and Conghui He. 2026. The Data Frontier for Large Language Models: Selection, Synthesis, and Tools. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 5: Tutorial Abstracts).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.acl-tutorials.2

@inproceedings{wu2026data,
  title = {{The Data Frontier for Large Language Models: Selection, Synthesis, and Tools}},
  author = {Lijun Wu, Wentao Zhang and Conghui He},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 5: Tutorial Abstracts)},
  year = {2026},
  doi = {10.18653/v1/2026.acl-tutorials.2},
  url = {https://doi.org/10.18653/v1/2026.acl-tutorials.2}
}

Updated

This page in Markdown → https://md.research.elamur.ai/free/en/papers/wu-2026-data-frontier-large-language-models/