About ML News

ML News is Hacker News filtered down to one thing: arXiv papers that would make a good thread on /r/MachineLearning. It has no accounts. Voting, commenting and replying happen on Hacker News; discussions are shown here read-only, and every reply link leads there.

How papers get in

Every Hacker News submission linking to an arXiv paper is a candidate. Papers whose arXiv categories are all outside the fields where machine-learning work is filed are left out. A language model reads the title and abstract of every other paper and decides whether machine learning is the substance of the paper, not just a tool it uses.

Summaries

The front page shows an explain-like-I'm-13 summary under each title: what is new, how it works, and what it found. Summaries are machine-written from the abstract, so check the paper before relying on a detail. Titles link to arXiv's HTML version of the paper when it exists, otherwise to the abstract page.

Ranking

The front page uses Hacker News' own formula, (points − 1)^0.8 / (age in hours + 2)^0.8, with a much gentler decay than HN's 1.8. ML papers are a small, slow stream on HN, so a well-discussed paper can stay on the front page for weeks. When a paper was submitted more than once, its best-ranked submission is shown.

Feeds

RSS follows the front page and a second feed carries every new ML paper. Both update hourly. Long abstracts are replaced by the short summary.