About Look et al.
A reference library of academic web pages, for researchers and their coding agents.
What is included
180 paper project pages, ten per venue and year: NeurIPS, ICML, ICLR, CVPR and ACL in 2025 and 2026, ICCV 2025, ECCV 2026, SIGGRAPH 2025 and 2026, CoRL 2025, CHI 2026 and EMNLP 2025, plus ten journal papers from TMLR, TPAMI, JMLR and Nature-family journals (2025 and 2026). Candidates come from arXiv comments that state the acceptance and link a project page; a page had to load, and workshop papers were left out. That gave a pool of 907 pages. Most venue-years had 40 or more candidates; CHI 2026 had 19 and EMNLP 2025 had 37.
50 researcher homepages from a pool of 236, drawn by three fixed rules. The first covers ACM Turing Award laureates (2018–2025) and first or last authors of Best Paper, Outstanding Paper and Test of Time awards at NeurIPS, ICML, ICLR, CVPR and ACL (2023–2025). The second covers first and last authors of ACL Outstanding Papers (2023–2025). The third covers cognitive science and computational neuroscience: Rumelhart Prize laureates (2010–2025), Glushko Dissertation Prize winners (2018–2025), authors of the CogSci Computational Modeling Prizes (2020–2025) and CCN keynote speakers (2023–2025). Only personal pages count; institutional directory pages, Google Scholar and LinkedIn profiles do not.
40 code READMEs, ten in each of four groups: vision (CVPR, ICCV, ECCV), learning (NeurIPS, ICML, ICLR and the journals), language (ACL, EMNLP), and graphics, robotics and HCI (SIGGRAPH, CoRL, CHI). Candidates are the official code repositories that the project pages above link to: 744 of the 907 pages link one, and 647 of those repositories have a README of at least 2.5 KB. Each README was rendered with the HTML that GitHub's own API returns, in GitHub's light theme at the 920px width of a repository page, and judged the same way as the pages below, on header, visuals, structure and polish. Their DESIGN.md files record structure and presentation (header block, badges, section outline, components) rather than CSS, because GitHub fixes the fonts and colors.
How pages were chosen
By visual design only, in two rounds. Every candidate was screenshotted at 1440×900. In the first round a model reviewer looked at every first screen in a pool and kept 20; in the second it looked at a 2700px-tall capture of each and kept 10. The first researcher pool was judged as one group, keeping 60 and then 30. The criteria were typography, a look of its own beyond template defaults, layout and rhythm, how the work is presented, and polish. Research content, venue and fame were ignored, and near-ties went to the page that added a style not yet picked. Pages that showed a loader, a blank screen or a bot wall could not be judged and were dropped. The reason for each pick is on its style page.
How a DESIGN.md is made
From the rendered page first. A headless browser reports the computed styles of the title, headings, body, buttons and containers, the fonts that actually loaded, the first-viewport anatomy in DOM order, and the color share of the first screen. The page's stylesheet and markup then supply selectors, hover and responsive rules. When the CSS and the render disagree, the render wins. Each file follows a fixed section order (Overview, First Viewport, color, type and spacing tokens, Layout Metrics, Components, Page Modules, guidance for other content, an agent prompt guide and a Quick Start) and passes an automatic check that every color used in the prompts appears in the token table. 270 of 270 entries were checked against the rendered page. DESIGN.md files are written in English only, so that every reader and every coding agent works from the same text.
The method was revised after a reconstruction test. Builders given only a DESIGN.md and different content rebuilt five pages, a blind judge tried to match each rebuild to its original, and an analyst scored similarity from 1 to 5 against the live page. With the first version, written from CSS alone, the judge matched all five, but overall similarity averaged 3.0 (2 to 4), mostly because of values the CSS declared and the page never showed. With the current version, the same test on the same five pages again matched all five and averaged 4.0 (all five scored 4). A separate audit of ten other files found 4 wrong values in about 129 checked claims; those were corrected.
Screenshots
Each entry shows an above-the-fold screenshot of the original page, reduced to 1200×750, credited to the authors and linked back to the live page. Screenshots are shown for reference and comparison, the same way design reference libraries do. DESIGN.md files describe visual style; they are not official guides of the authors.
License and citation
The DESIGN.md files, the selection and the site text are under CC BY-NC 4.0. You can reuse them for non-commercial work with credit to Look et al. (Dengzhe Hou). For commercial use, ask first. The code is MIT. Screenshots belong to the pages' authors and are not covered by either license. To cite the library, use Hou, D. (2026), Look et al.: A Design Library of Academic Web Pages for Humans and AI Agents, Zenodo, doi:10.5281/zenodo.23718470.
Removal
If you are an author and want your page or screenshot removed or corrected, open a removal request (a short GitHub form) and it will be taken down. Requests are handled by hand, usually within a few days.