Papers with Code

What it is

Papers with Code is a free and open resource for machine learning papers and code^[1]. Its premise is that a research result is easier to trust and reuse when the paper sits one click away from the code that produced it. TechCrunch described the site as linking research papers to their open-source code and maintaining benchmark leaderboards that track the state of the art across tasks^[1].

A leaderboard of this kind gathers the scores that papers report for a single task and ranks the methods against one another. A reader can see which approach currently reports the best number, then follow the link back to the paper and the repository behind it. The value is mostly in the joining-up: papers, code and numbers live in one place instead of being scattered across PDFs, personal websites and code hosting sites.

Origins

Papers with Code was the custodian product of Atlas ML, a startup founded in 2018 by Robert Stojnic and Ross Taylor to make it easier to discover and apply deep learning research^[1]. Atlas ML was an alumnus of Entrepreneur First and raised seed funding from Episode1 and Kindred Capital^[1].

Ownership change

A regulatory filing for Deeptide Ltd., the company behind Atlas ML, shows that Facebook became a majority owner on 13 December 2019^[1]. On the same day, co-founder Robert Stojnic published a Medium post titled “Papers with Code is joining Facebook AI”^[1][2]. The acquisition was not announced by Facebook at the time, and TechCrunch reported a ballpark price of around $40 million^[1].

Company filings are one of the ways a smaller acquisition becomes public: a change of control has to be recorded even when no announcement is made. For people who rely on a resource like this one, the practical questions after such a move are whether it stays open and whether it keeps being maintained.

Relaunch

Hugging Face's Niels Rogge relaunched the site as paperswithcode.co and described the new features in a blog post published on 24 May 2026^[3]. Hugging Face publishes the last publicly available snapshot of the PWC datasets as the pwc-archive organisation on the Hub, and that archive powers paperswithcode.co^[4].

Why it matters

Machine learning moves fast, and much of its output arrives as preprints and conference papers rather than as finished software. A catalogue that pairs each paper with code lowers the cost of checking a claim, of reproducing a result, and of starting from someone else's implementation instead of from nothing. Leaderboards add a second shortcut: rather than reading a dozen papers to learn what currently works best on a task, a reader can look at one table and follow the links from there.

The same design has limits. A leaderboard is only as good as the numbers submitted to it, and comparisons drawn from different papers can be uneven when the experimental settings differ. That is why the link back to the paper, and to the code that produced the number, matters as much as the ranking itself.