NXAI publications & resources page

Two files, no build step, no dependencies beyond a Google Fonts CDN link:

• index.html — page shell, styling, rendering logic. You should not need to touch this for routine updates.
• links.json — every entry on the page. This is the only file you edit to add, remove, or correct a link.

Deploy on GitHub Pages

Option A: dedicated repo (recommended for a short URL like nx-ai.com/links via CNAME later)

1. Create a repo, e.g. NX-AI/linktree.
2. Commit both index.html and links.json to the main branch, in the same folder.
3. Repo Settings > Pages > Source: main branch, / (root).
4. Page is live at https://nx-ai.github.io/linktree/.

Option B: fold into an existing repo (e.g. the tirex repo) as a /docs folder if you want to reuse Pages already configured there. Keep both files together in that folder.

Adding or editing an entry

Open links.json and add an object to the array:

{ "model": "tirex", "type": "paper", "icon": "arxiv", "title": "...", "meta": "Author et al. · Venue Year", "tag": "2611.xxxxx", "url": "https://arxiv.org/abs/2611.xxxxx" }

Field reference:

• model: one of xlstm, tirex, vision-lstm, bio-xlstm, lram. To add a new model family, also add it to the MODELS / MODEL_KEY lists near the top of the script in index.html — that's the one structural change that does require touching the HTML, since it drives the filter tabs.
• type: paper or code. Controls which subheading it appears under.
• icon: arxiv, github, huggingface, docker, jupyter, book, article, or package. Brand icons render in their own brand colour; the rest render in teal.
• meta: freeform subtitle line, kept to the format Author(s) et al. · Venue Year for consistency.
• tag: short label shown as a pill on the right (arXiv ID, "GitHub", "Docs", etc).
• date and venue (papers only): power the Timeline view. date is "YYYY-MM" — use the actual conference month (not the arXiv submission month) once a paper is accepted; fall back to the arXiv month only for genuine preprints. venue is one of NeurIPS, ICLR, ICML, or preprint, and drives the colour-coded badge. Entries without both fields are simply left out of the timeline.

Same content on two platforms (like the "TiRex on the edge" post): instead of "url", use a "links" array so both destinations get their own pill instead of implying two different articles:

{ "model": "tirex", "type": "code", "icon": "article", "title": "Post title", "meta": "Same article, cross-posted",
  "links": [
    { "tag": "NXAI", "url": "https://www.nx-ai.com/..." },
    { "tag": "Hugging Face", "url": "https://huggingface.co/..." }
  ]
}

No LAST_UPDATED constant to remember anymore — the footer date is generated automatically from the page load.

Local testing

Opening index.html directly from disk (file://...) will fail silently or show a load error, because browsers block fetch() of local files for security reasons. Serve the folder over HTTP instead:

python3 -m http.server 8000

then open http://localhost:8000/. This restriction disappears once deployed to GitHub Pages, since that's real HTTP.