How Salient ranks papers
Salient surfaces good preprints early using signals that do not simply proxy journal prestige, author fame, or institution. Every page is built from one transparent, reweightable signal score plus, on some pages, a time or personalization factor. Published journal articles are read in only as a relevance reference — they are not shown in the feeds.
The pages
| Rising | The signal score over a window you pick — Today, This week, or This month — each with a window-scaled time decay: Today is the raw score (no decay), This week rewards momentum (fresh things with early signal float up and fade within days), and This month applies only a gentle freshness tilt so the month's best surface rather than just the last few days. Filterable by content type. The content types (and their exemplar papers) are set in the site configuration; each paper is classified by embedding similarity to prototypes built from those exemplars, but only when it's clearly closest to one type and far from the others — ambiguous papers are left unclassified rather than mislabeled. An LLM pass can sharpen this further. |
| For you | Personalized: your interest embedding — the average of your ORCID publications, plus the papers you upvote — is compared with each preprint and blended with the signal score, then decayed by recency. Upvoted papers count as extra “your papers” when measuring closeness, which helps if you have few publications; their combined influence is capped so it never outweighs your own ORCID work. Cards may say “because you wrote X” for your publications, but upvotes stay private — they quietly shape your feed without being announced. Two controls let you steer it: a matching style — compare a preprint to the single average of your work (centroid), to your single closest paper (best match), to your three closest (top-3), or deliberately to the papers furthest from all your work for on-topic exploration (frontier) — and a balance slider trading personal fit against the overall signal score. |
| Monthly Issues | Monthly, citable, frozen selections of the period's best — the overlay-journal layer. Once published, an issue never changes. |
The signal score & current weights
A weighted blend of prestige-independent signals. The last two apply only when a paper has that signal; a paper with no discussion or reads is scored on the others, never penalized for being quiet. And they are weighted by how much evidence backs them — a single post or a single reader barely moves the score; influence grows only as independent posts or distinct readers accumulate — so thin, early signals can neither unfairly lift a paper nor drag it down. Cards may also carry a top X% badge — where a paper ranks among the last month's preprints on a single signal (novelty, discussion, reads, openness), shown only when it lands in the top tenth.
| Relevance | Similarity to the topic region (the seed journals), as a soft prior. | 0.40 |
| Openness | Open code, data-availability, preregistration, benchmarking — objective and checkable. | 0.05 |
| Novelty | On-topic distinctness: how different a paper is from its nearest neighbours across a long history of embeddings — but credited only insofar as it's on-topic, so “novel” means within the field yet different, never an off-topic outlier. | 0.25 |
| Discussion (💬) | Independent, substantive social-media discussion beyond what the discussing accounts' reach predicts — the “attention residual”. Author self-promotion is excluded; critique counts. | 0.40 |
| Readers (👁) | On-platform engagement: opening an abstract is a soft signal, clicking through to the paper stronger, upvoting stronger still. Each reader counts once (their strongest action), and the score is shrunk toward zero until enough distinct readers accrue — so a few clicks can't dominate. | 0.35 |
Tune it yourself
The weights above are only a default. On Rising, the “⚙ Adjust feed” panel lets you slide each signal up or down and re-rank the feed live — upweight novelty, ignore openness, whatever fits how you read. The sliders are independent: because the score is renormalized over whatever weights you set, only their ratios matter, so you never have to trade one off against another to keep a total fixed. Your choices ride in the page URL, so a tuned feed is shareable and needs no account; “Reset” restores the defaults.
Each card shows a small freshness dot with a relative age — how long ago the preprint first entered Salient: under 2 days, this week, 1–2 weeks, older. The age label is always shown, so the colour is only a scan aid.
Every surfaced paper carries a plain-language explanation of why it appeared, and the 💬 / 👁 counts on each card show its discussion and distinct readers. The ranking function and weights are published here deliberately — transparency is what earns credibility for an algorithmic overlay journal.