How this atlas is built

The Dataset Condensation Atlas is curated bySahar Rahimi Malakshan, whose PhD research at West Virginia University is on dataset condensation, including distribution matching and generative multimodal distillation. It aims to be the reference she wanted when starting in the field: organized by ideas, honest about evidence, and current.

Rules the atlas follows

1. No claim without a fetched source. Every paper's title and authors are checked against a fetched primary source (arXiv, CVF, OpenReview, ACL Anthology, publisher). Each entry records how deeply the atlas read the paper: metadata only, abstract, or full text. Summaries never claim more than that depth supports, and the structured "Explained" sections are written only from full text.
2. Abstracts only from arXiv. arXiv metadata is released under CC0, so arXiv abstracts are shown verbatim and labelled as such. Text from other venues is never copied.
3. No mixed-protocol leaderboards. Numbers produced under different label, backbone or budget protocols are never ranked against each other. See Evaluation.
4. No work under review. Papers under double-blind review are not listed, including the curator's own.
5. Discovery is mechanical, classification is judged. Candidates are harvested from the Awesome Dataset Distillation list and from targeted searches, then placed by hand into the four facets.

Tiers and depth

Landmark papers originated a family or changed its direction. Core papers are needed to understand a family today. Notable papers are meaningful steps inside a family.Catalogued papers are verified and classified but not narrated. Of 384 papers,384 are verified against their source and 108 have been read in full.

Keeping it current

The repository contains the scripts that harvest candidates, fetch metadata, validate every entry against a schema and build this site. A coverage check lists papers present in the community index but missing here. Data was last built on 2026-09-14.

Errors and contributions

If a paper is misclassified, a summary overstates a result, or an important paper is missing, pleaseopen an issue. All data is plain YAML indata/; the whole atlas can be downloaded as one JSON file.

Citing

@misc{rahimimalakshan2026dcatlas,
  author       = {Rahimi Malakshan, Sahar},
  title        = {Dataset Condensation Atlas: A Living Map of Dataset Distillation Research},
  year         = {2026},
  howpublished = {\url{https://saharr1372.github.io/dc-atlas/}}
}

Please also cite the community list that seeds discovery: Li, Zhao and Wang, Awesome Dataset Distillation (2022–).