Skip to results
MLSift

About MLSift

Several hundred machine-learning papers land on arXiv every day. MLSift reads each abstract once, files the paper under one of 19 domains, names the key architecture, tags the topics and rates how far it moves the field, then shows the result as a feed you can filter. It is a triage tool: the papers stay arXiv's, and the PDF you read comes straight from arXiv.

How it works

Every morning at 06:00 UTC the pipeline asks the arXiv API for new submissions in cs.LG, cs.CV, cs.CL, cs.AI and stat.ML, looking back a week because arXiv announces papers a few days after submission. From 10:15 UTC a language model (DeepSeek V4 Flash, through a typed schema) classifies each new abstract; a bad answer marks that one paper as failed and never blocks the rest. The feed shows what has been classified, newest submission day first, and the footer on the front page says when the classifier last ran.

Classifications are model output. They are good enough to decide what to read first; they are not a peer review, and a wrong domain or an over- or under-stated innovation level is a known failure mode. The taxonomy is versioned, so papers can be re-filed when it changes.

What an account stores

Reading is anonymous. Signing in exists for one reason: saving papers to a list you can open on another device. It is done by email, with a link and a six-digit code, no password. An account holds:

Nothing is sold, shared or used for advertising. Visits are counted with Vercel Web Analytics, which aggregates page views and speed without cookies or a per-visitor identifier. The reader remembers your zoom and dim preference in your own browser. To delete an account with its saved papers and watchlist, write to hello@mlsift.com.

Contact

MLSift is a small independent project and is not affiliated with arXiv. Bugs, misclassifications and ideas: hello@mlsift.com. Thank you to arXiv for the open API that makes this possible; please use arXiv's own site to cite and download papers at scale.

Back to the feed