Papers with Code Is Back: Where ML Research Meets Leaderboards and Repos

⬅️ Back to Articles

I spent an hour on paperswithcode.co last night and kept opening tabs I did not plan to open. Transformer papers I half remembered. Leaderboards I had never seen for agent benchmarks. GitHub repos attached to the actual PDF. Some of the most influential work of the last decade lives here in one graph: paper, code, numbers, lineage. It is the site I missed when the original Papers with Code went quiet.

  1. The old site mattered. Robert Stojnic and Ross Taylor built Papers with Code into the default map of state-of-the-art ML. Meta bought it for a reported $40 million in 2019. Updates slowed. Leaderboards aged. The team behind it went on to LLaMA. Niels Rogge at Hugging Face relaunched paperswithcode.co in 2026 as an independent revival, not a corporate rebrand.

  2. The core loop is still paper → code → benchmark, but the surface area grew. Domains span vision, speech, embeddings, agents, time-series, and more. Each area gets leaderboards (COCO, MMTEB, Open ASR, coding-agent suites like Terminal Bench). Trending papers rank by GitHub star velocity, not just citation count. You can see what is moving this week, not what moved in 2021.

  3. Methods are first-class again. RLVR, Mamba variants, Gated DeltaNet, and similar tags link papers that share a technique. Paper pages show lineage banners when a model has a clear predecessor or follow-up (DINOv2, GLM-4.5, Mamba-3). That context is half the value when you are trying to understand why a benchmark jumped.

  4. Submission is open and AI-assisted. paperswithcode.co/submit accepts arXiv, bioRxiv, GitHub repos, and blog posts. The site auto-tags tasks, pulls linked repos and Hugging Face artifacts, and attaches evals where it can. Rogge reported roughly 3,000 eval rows so far, growing from Transformers-supported models. Multi-metric leaderboards now cover WER and latency on ASR, mAP and FPS on detection, and similar tradeoffs elsewhere.

  5. You might ask whether another leaderboard site adds noise in 2026. Fair. Hugging Face already hosts millions of models. The point of PwC was never storage. It is curation with evidence: which paper, which repo, which number, on which benchmark. When you need “best open model for X under constraint Y,” this beats scrolling model cards.

  6. It is early and community-maintained. Coverage gaps exist. Some domains are thin until someone adds results via the edit flow or GitHub tutorials. Sign-in with Hugging Face handles accounts. Treat it as a living index, not a finished encyclopedia.

The takeaway: Bookmark one domain you actually work in (agents, OCR, ASR, whatever). Pick a leaderboard, read the top three papers, clone the repo that matches your constraint. Submit a missing result if you have one. The site gets better when practitioners feed it numbers, not when you treat it as read-only wallpaper.

Related TMFNK Content

Crepi il lupo! 🐺