# xoult

> A free, personalized daily feed of new arXiv papers in machine learning. Each weekday it
> ranks the new announcements for every reader from what they save and dismiss, and gives
> each paper a one-line TLDR.

xoult covers cs.CL (Computation and Language), cs.IR (Information Retrieval), cs.LG (Machine Learning), cs.AI (Artificial Intelligence), cs.CV (Computer Vision), stat.ML (Machine Learning (Statistics)). That is about 1,900 new papers a weekday.

- No account is needed to start: readers pick categories and get a ranked list at once.
  Optional Google sign-in keeps saved papers across devices.
- Each recommendation says which saved paper it follows from ("because you saved X").
- TLDRs are written by an LLM from the abstract; the abstract and arXiv link are always shown.
- Readers can download or delete all their data from Settings.

## How the ranking works

- Papers are embedded with all-MiniLM-L6-v2. A reader's profile is the mean of their saved
  papers moved away from dismissed ones (Rocchio). The 200 papers from the past week in
  their categories that are closest to it are reranked by embedding similarity blended
  50/50 with TF-IDF similarity to the saved papers, plus recency and category match.
- Offline evaluation on CiteULike-a (5,551 users' libraries): NDCG@10 of 0.1118 for the
  production ranker, against 0.1005 for embeddings alone and 0.1016 for TF-IDF alone.

## Links

- [Home](https://xoult.onrender.com/): the feed
- [Daily list](https://xoult.onrender.com/papers/cs.CL): every paper from the newest announcement in one
  category, with TLDRs; /papers/<category> for the others
- [Ranking stats](https://xoult.onrender.com/stats): public save rate per ranker version, JSON, last 14 days