The feed has two orderings. Recent is the simple one - newest first, by real-world ship date. Trending is the interesting one: it surfaces what is gaining attention right now, not what is merely new or merely popular.
What “trending” means here
Something trends when attention around it is growing fast lately, and that signal fades as the thing ages. Every item gets one score between 0 and 1, blended from two parts:
- Velocity (about two-thirds of the score). How fast its attention metric is climbing over the trailing week. A year-old eval that a just-released model posts a score on can outrank a quiet brand-new one.
- Recency (a fading floor). Newer items get a head start that halves roughly every two weeks, so something fresh still ranks before any velocity has had time to build.
We deliberately do notrank by raw popularity or by quality. A strong model from last year is not “trending” just because its scores are high - that is what the catalog and leaderboards are for.
The attention signal, per type
“Attention” means something different for each kind of thing, so every type has its own metric. Each one is normalized within its own type, so a hot paper and a hot eval land on the same 0-to-1 scale and can share a single ranking.
Star velocity- the stars the official code repo gained in the last 24 hours, mirrored from Papers With Code's own trending feed.
Benchmarking activity - how many fresh evaluation runs are landing on it, read from the real evaluation dates and weighted by source credibility (an official result counts for more than a vendor self-report), plus growth in how many models report.
Evaluation activity- how actively it is being benchmarked and ranked after release, again weighted by how trustworthy each result's source is, plus growth in its benchmark coverage.
Adoption - how many new evals the tool (an RL environment, dataset, or scaffold) is shown to lift or implement.
How often it updates
Trending is recomputed once a day. Benchmarking activity is read straight from the evaluation log, so an eval or model that gets freshly tested registers right away. Adoption counts (models reporting, evals improved) use a daily snapshot diffed against the past week, so that part reflects this week's movement rather than an all-time total. Paper velocity is mirrored daily from Papers With Code's precomputed trending feed - a few summary reads, not a crawl of their API.
The snapshot-based part needs a little history: right after a count starts being tracked there is nothing to diff against, so those items lean on the live activity signal and the recency floor until a day or two of snapshots accumulate. New things never disappear - they just lead with freshness until their velocity is measurable.
Prefer the plain chronological view? Switch to Recent on the feed.