How Aftergleam ranks papers

Fetching

arXiv announces new papers once each weekday evening, at 20:00 US Eastern, from Sunday to Thursday. The app fetches once after each announcement and not in between, so refreshing twice does not ask arXiv twice. Requests to the arXiv API are kept at least three seconds apart, as arXiv asks. The other servers have no fixed schedule, so the app asks them for the last few days, through the bioRxiv API, the OSF API and the Crossref REST API.

By default this happens at 05:00, on Wi-Fi while the phone charges, and the digest is built straight after. Opening the app in the morning then only reads a database.

The model

Each paper is represented by the words and two-word phrases in its title and abstract, weighted by TF-IDF. A logistic regression learns from your reactions, with older papers from the phone as negative examples, and is retrained on the phone whenever your reactions change. There is no neural network and nothing to download.

That choice was measured rather than assumed. In a leave-one-out test on a real library of 38 papers, TF-IDF placed 21 of 24 held-out computer vision papers in the top ten of 301 candidates. A small sentence-embedding model, MiniLM, placed 19. At that sample size the difference is not significant, so the simpler model was kept: nothing to download, no native code, and reasons that quote real words.

A later comparison found that combining TF-IDF with a sentence embedding ranks better, by 3 to 12 per cent in nDCG@25, depending on how many papers a reader has liked. It would add 25 to 35 MB and native code to the app, and has not been shipped.

What it learns from

Each thing you do with a paper counts for a fixed amount, from 0 to 1:

The strongest thing you did sets a paper's value, and anything else you did moves it part of the way towards 1, never past it. Less like this overrides everything else. The subjects you follow count as short example texts at 0.7.

A paper you scroll past is not used to train the model. It only counts towards how many places its subject gets in later digests.

Putting a digest together

Early on, the model is not trusted much. Its predictions are pulled towards a neutral value in proportion to how little evidence there is: with evidence from 6 papers the model carries about a fifth of the weight, and with 100 about four fifths.

Places in the digest are shared between your subjects by Thompson sampling, from how often you engage with each. Older evidence fades with a half-life of 30 days, so a change of project shows up within weeks. A subject the app knows little about still gets tried, because its uncertainty is wide.

Within a subject, papers are drawn at random with a strong preference for high scores, using Gumbel noise, rather than taken strictly from the top, and a paper too similar to one already chosen is marked down. Refreshing without anything new gives a different digest.

Acceptance at a conference or journal multiplies a paper's score by up to 1.35 at the default setting, so a well-published paper on a topic you dislike still ranks low. The venue comes from what authors write in arXiv's comments and journal fields, or from bioRxiv's and medRxiv's note that a paper has been published. Papers on the Hugging Face daily list get a smaller boost of the same kind.

The first three cards are always the best matches. After them, one card in four is a deliberate detour, each labelled: first a paper from a neighbouring field (math.OC, outside your usual), then near misses, papers that ranked just below the cut (testing whether this is for you). How you react to those shows the model where your line is. Digest size, weight on venue, exploration and variety can all be changed in the app's settings.

The reason on each card

For each card, the app finds the paper you kept that it most resembles, among those you liked, saved, downloaded, shared or read, and prints the words the two have in common, strongest first. A card reading matches sparse, autoencoder, diffusion means exactly that. If nothing you kept has a cosine similarity of at least 0.06, the card names its category instead.

The reason is not taken from the model's weights. Weights fitted on a few dozen papers are unreliable, and printing them produced reasons such as matches optimal, thereby, known.

Months later

If you passed over a paper three to twelve months ago and it has since been accepted at a conference or journal, it can come back, at most one per digest: You passed on this in March. It was accepted to NeurIPS 2026. Workshop papers do not count. Once a week, on Wi-Fi, the app asks arXiv for fresh details of those papers.

Citation counts would be the obvious signal, but open databases such as OpenAlex record almost none for preprints this recent, and asking Semantic Scholar or anyone else paper by paper would tell a third party what you read.

Sources and subjects

The app only asks the servers your subjects need, so a mathematician's phone never contacts the Law Archive.

Limits

• It is for Android. There is no iPhone or desktop version.
• It reads preprint servers, not journals, though a card shows when a preprint has since been accepted somewhere.
• Search and library import go through arXiv. Papers from the other servers can be searched once they are on the phone, and a library entry can only be matched to an arXiv paper.
• ChemRxiv papers open in the browser, because ChemRxiv does not let apps download them.
• Papers from PsyArXiv, SocArXiv, EdArXiv and the Law Archive are shown without authors. Asking OSF, which hosts them, for author names made fetching too slow.
• Popular leans heavily towards machine learning, as does the Hugging Face list it ranks by. For many fields it stays empty.
• The first days are general. Until you have reacted to a few papers, the digest leans on your subjects, recency and venue.

How far to trust the numbers

The measurements above come from the developer's own library and from a simulation that runs the app's own ranker for forty mornings with six made-up readers. They are evidence that the approach works, not proof that it works equally well in every field. The full design notes, with every measurement and what it changed, are part of the source code, along with the Python harness that made them.

Thank you to arXiv for use of its open access interoperability.