End-to-end pipeline that turns 5,356 ICLR 2026 accepted papers into a clean, PDF-derived institutional-affiliation dataset and a publication-ready treemap of who is shaping AI research right now.
This avoids the OpenReview-profile drift problem (where authors' current job appears on every paper they ever wrote — e.g. listing Wyoming as the affiliation for a paper actually written at UBC). Affiliations come from the paper's title block PDF, not from author profiles.
Follow me for more analysis like this, plus AI engineering & research insights:
- LinkedIn — linkedin.com/in/dmytrolopushanskyy
- GitHub — github.com/DmytroLopushanskyy
If this dataset or the pipeline is useful to your work, a follow / star is the easiest way to encourage me to keep publishing this kind of analysis.
The headline chart
Each rectangle is one institution sized by the number of accepted papers it appears on (counted once per paper, regardless of how many of the paper's authors are affiliated with it). Region cells are sized by the cumulative count of their top-50 institutions. Lighter shade = academia / research institute, darker shade = industry.
Square version (for social posts):
charts/iclr2026_top50_treemap_unique_grouped_square.png
What's in data/
| File | What it is |
|---|---|
iclr2026_public.csv / .xlsx |
The main dataset. 5,356 accepted papers with PDF-derived authors and institutions, normalized institution canonical names, country/region, abstract, OpenReview URL. UTF-8 with BOM for Excel compatibility. |
iclr2026_institutions_ranked_unique.csv |
Top-N institutions ranked by unique-affiliation count (each institution +1 per paper). |
iclr2026_institutions_ranked_first_author.csv |
Same, but only counting the first author's institution. |
iclr2026_institutions_ranked_fractional.csv |
Same, with fractional 1/N credit per institution per paper. |
iclr2026_method_sensitivity.csv |
Side-by-side rank under all three counting methods, so you can see which institutions are robust and which are method artefacts. |
Columns in iclr2026_public.csv
| Column | Meaning |
|---|---|
Decision |
Oral / Poster |
Title |
Paper title (LaTeX math markup converted to Unicode — $\alpha$ → α, $\nabla$ → ∇, $\textrm{...}$ → plain text, etc.) |
Authors |
Semicolon-separated, in author order |
Institutions |
Same row order as Authors. PDF-extracted text per author (with OpenReview fallback for the ~6% of papers where PDF parsing failed). |
Institutions_canonical |
Normalized via ~250 rules. MIT / Massachusetts Institute of Technology / MIT CSAIL all collapse to MIT. Deduped per paper. |
Countries |
Per-paper deduped list. |
Regions |
High-level region per paper (China, USA, Hong Kong, etc.). |
Affiliation_source |
pdf (94%) / parse_fail (6%) / no_pdf (4 papers). Audit trail. |
Primary_Area |
OpenReview track. |
Keywords |
Author-supplied. |
Abstract |
Full text. |
OpenReview_URL |
Direct link to the paper. |
Quick start
Just regenerate the chart
git clone https://github.com/DmytroLopushanskyy/iclr2026-affiliations.git cd iclr2026-affiliations python3 -m venv .venv && source .venv/bin/activate pip install -r requirements.txt python3 make_iclr_treemap.py --source pdf
This reads data/iclr2026_public.csv and writes the treemap PNGs/SVGs into charts/.
Add --shape square for a 1:1 version. Add --source openreview to compare against the OpenReview-profile-only version (requires running the scraper first).
Reproduce the full pipeline from scratch
You only need this if you want to re-derive the dataset (e.g., for a new conference). It takes ~1–2 hours of network time and ~5 GB of disk for the PDF cache.
# 1. Scrape OpenReview metadata (requires an account) export OPENREVIEW_USERNAME=... export OPENREVIEW_PASSWORD=... python3 scrape_openreview.py # → data/iclr2026_accepted.{csv,xlsx} # 2. Download all accepted-paper PDFs (~5 GB; rate-limited; retry script handles 429s) python3 download_missing_pdfs.py python3 retry_missing_pdfs.py # picks up anything that hit a 429 the first time # 3. Parse PDFs and merge with OpenReview data python3 build_pdf_spreadsheet.py # → data/iclr2026_accepted_pdf.{csv,xlsx} + data/pdf_parse_summary.txt # 4. Build the public-facing CSV (sanitization + LaTeX-to-Unicode + canonical names) python3 build_public_spreadsheet.py # → data/iclr2026_public.{csv,xlsx} # 5. Render the charts python3 make_iclr_treemap.py --source pdf # → charts/iclr2026_top50_treemap_*.{png,svg}
How the parser works
parse_pdf_affiliations.py handles four layout patterns common in ICLR template papers:
| Pattern | Layout | Example |
|---|---|---|
| A | Numbered footnote markers | Author1,2 Author1,3 ... \n 1Inst A 2Inst B 3Inst C |
| B | No markers, single shared affiliation | Author1, Author2 \n Single Institution |
| C | Per-author stanzas separated by emails | Author1 \n Inst A \n a@x.edu \n Author2 \n Inst B \n b@y.edu |
| D | Alternating name / affil pairs (no emails) | Common for industry-only papers (Apple, Anthropic, etc.) |
Plus a footnote-text filter that catches and discards "Equal contribution", "Corresponding author", "Project lead", "These authors contributed equally" — these used to leak into affiliation strings before being filtered out.
Result: 96% of papers parse successfully; the remaining 4% fall back to OpenReview profile data (transparently flagged in the Affiliation_source column).
Methodology choices, briefly
- Counting: each institution counted once per paper, regardless of how many of its authors are listed. Same rule used by the AI World NeurIPS leaderboard. The repo also generates first-author-only and fractional 1/N variants for sensitivity.
- Canonicalization: ~250 regex rules collapse spelling/abbreviation variants (HKUST = Hong Kong University of Science and Technology = The Hong Kong University of Science and Technology, etc.). Institutions in the chart's top-50 are stable across all three counting methods (see
data/iclr2026_method_sensitivity.csv). - Region grouping: countries → 17 broad regions for the treemap. Hong Kong is shown separately from mainland China because Hong Kong universities operate under a separate higher-education system (different governance, language of instruction, listed separately in QS/THE rankings).
License
MIT. The data is derived from publicly available OpenReview submissions and ICLR 2026 paper PDFs; please cite this repository if you use it in published work.
Stay in touch
If you build something on top of this, ping me — I'm always interested in seeing where this kind of pipeline gets used. And if you want more posts like this (research-engineering deep dives, applied AI analysis, papers I'm reading), the best place is:
- LinkedIn — linkedin.com/in/dmytrolopushanskyy
- GitHub — github.com/DmytroLopushanskyy
— Dmytro Lopushanskyy