About

This is an unofficial reader's companion to the 224 Oral papers at ICLR 2026. It exists to make skimming, searching, and getting a sense of where the field is heading easier than clicking through 223 OpenReview pages one at a time.

What's on this site

• All oral papers with title, authors, abstract, PDF link, OpenReview link, session time, and room (when available).
• A hand-curated taxonomy of 12 topics; each paper is assigned a primary topic via keyword heuristic with manual overrides.
• Client-side fuzzy search across titles, authors, abstracts, and AI-generated summaries. Press / to focus.
• Per-paper AI-generated enrichment: a one-sentence summary, contributions, methods and datasets, limitations, and future work.
• A cross-paper trends page identifying recurring themes in what authors say comes next.

Data sources

• Paper metadata: OpenReview API v2, filtered to venue = "ICLR 2026 Oral".
• Session times and rooms: iclr.cc/virtual/2026/events/oral.
• Joined by fuzzy title match (rapidfuzz WRatio, threshold ≥ 88). Unmatched papers show a "session time not found on iclr.cc" notice.

How the AI content was produced

Per-paper summaries and extractions: Claude Haiku 4.5. Input is the paper title, abstract, and extracted Conclusion / Limitations / Future Work / Discussion sections from the PDF (pdfplumber, pypdf fallback). The model is instructed to return [] when information is not explicitly stated in the source text, and to never paraphrase the abstract. Responses are validated against a pydantic schema; invalid responses are retried once with error feedback and then stubbed.

Trend themes: Claude Sonnet 4.6, given the concatenated future-work and limitations bullets from every paper, asked to identify 5–10 cross-paper themes. Each theme's paper IDs are validated to reference real papers in the dataset.

How to tell authors' content from AI content

• Plain prose (no border, normal text color) means it comes directly from OpenReview: title, authors, abstract, author-provided TL;DR, keywords, session info.
• Dashed left border + "Auto-generated" header means the text was produced by an AI model. The specific model is named inline.
• Blockquotes marked "from the paper" are verbatim excerpts that the model pulled from the paper's PDF (e.g., limitations bullets). The structure was AI-selected but the words are the authors'.

Copyright and attribution

Paper titles, abstracts, and author lists are shown verbatim from OpenReview and are copyrighted by their authors. This site is not affiliated with ICLR, OpenReview, or Anthropic. Reviewer content is not used.

Known limitations

• Fuzzy title matching may miss papers whose titles were edited between OpenReview and the virtual site.
• PDF extraction occasionally fails or returns partial text; those papers show only the abstract-derived summary.
• Topic assignment is a keyword-weighted heuristic; some papers land in "Uncategorized" until manually overridden.
• The AI may misread or omit details. When in doubt, read the paper.

Code

Built with Astro, Tailwind, and MiniSearch. Hosted on Vercel. Source: github.com/bmtlima/iclr-2026-oral. If a paper is miscategorized or an extraction is wrong, open an issue.