Methodology
How We Rank Papers
Scholar Feed uses empirically validated signals to surface high-quality research. Here's how.
Predicted Impact
When a paper is new — before real citations have arrived — its badge is a forecast: a 0–100 predicted-impact score (the teal half of the badge, explained under Reading the Score below). A gradient-boosted model forecasts how often the paper will be cited by other papers in our corpus over its first year, using only what is knowable at publication: the paper's own signals, the citation track record of the works it cites, and the prior track record of its authors. Venue and present-day citation counts are deliberately excluded — both would leak information from the future.
We validate it the hard way. The model is trained on papers published through 2023, then asked to rank papers from 2024–2025 — and checked against what those papers actually went on to accrue. On that held-out test the ranking tracks realized first-year citations at a Spearman correlation of 0.59, and the papers it places in the top 10% collect roughly 4× the average paper's citations — far above the novelty-only score it replaced (0.10).
Ranking Formula
Papers in the main feed are scored by a composite rank signal with five weighted components. This score determines which papers appear at the top of the feed and which papers are selected as the hero paper each day.
For personalized feeds, papers citing papers you have liked receive an additional citation-chain boost. Papers you have already seen are deprioritized but remain accessible via the “Show seen” toggle.
Reading the Score
Every paper carries one 0–100 number — its impact badge, the score shown on every card. It is field-normalized: a paper is measured against the typical citation accrual for its own field (and, once it has a citation history, its year), so a 90 means the same thing whether it's an NLP paper or a control-theory one, this week or from 2015. You can compare two papers from different areas or eras directly.
The badge reads on one honest curve with two regimes, signalled by its color:
Predicted — the teal badge
For newer work, the number is the forecast described above, mapped onto this scale: an estimate of how much the paper will be cited, before the citations have arrived.
Established — the gold badge
Once real citations accumulate, the number switches to realized impact — what the paper actually earned, relative to its field and year. A paper earns its way from a teal forecast up to a gold, established score as the citations come in, so a landmark you already trust now wears a gold badge.
The scale is absolute — a fixed curve, not a percentile or a daily cohort — so it tracks the real size of a paper's impact: most papers score low and only standout work approaches 100. A fresh paper reads modest and climbs as it earns citations; a typical day's best new paper sits in the middle of the range, and only a genuine landmark reaches the high 90s. That's deliberate — the number tells you whether something is actually remarkable, instead of crowning a “99” every day. On a paper's own page we also show its standing within its field(“top 4% in cs.RO”).
Originality is a separate axis. A paper's novelty (how new the method is) is largely independent of how much it will be cited, so we don't fold it into the impact number; instead the paper page shows it on its own line (Highly original / Incremental advance / Synthesis-oriented).
It is a measure of citation impact — predicted, then realized — not a verdict on correctness, and it counts only in-corpus citations (which undercount real-world citations). The recency-first ranking above powers the “Recent” feed; Trending ranks separately by what is rising right now, not by this badge.
Citation Intelligence Pipeline
Scholar Feed maintains an internal citation graph built from ~1 million citation edges extracted from arXiv paper metadata and source files. This graph powers three distinct signals:
Internal Citation Score
A quality-weighted in-degree score: each citation a paper receives contributes proportionally to the citing paper's quality. Papers heavily cited by high-quality work score higher than those cited by low-novelty papers. Recomputed nightly.
Citation Velocity
The count of new internal citations received in the past 7 days, normalized via LN(1+count)/LN(21) — capping the scale at ~20 new citations/week, which is extraordinary for a preprint. Citation velocity is one of the inputs to the unified paper score; the Trending section ranks by that unified score.
Bibliographic Coupling
Two papers that share many references are likely related in a way that semantic similarity alone may miss. The similar papers engine blends Jaccard overlap of shared references with embedding cosine similarity (65% embedding, 25% Jaccard, 10% direct-citation bonus). This surfaces methodologically adjacent papers that use different vocabulary.
The citation graph is updated nightly via the ArXiv source reference scraper, which downloads LaTeX source packages, parses bibliography files, and matches references against the Scholar Feed corpus by title and arXiv ID.
Venue Validation
The model never sees venue — it is trained only on citations. That makes venue an independent check: do the papers we score highest actually clear peer review at top-tier venues? On a mature cohort where venue is now known, they clearly do.
Among papers that reached any venue, the share published at a top-tier venue (NeurIPS, ICML, ICLR, ACL, CVPR and peers) rises steadily with the score:
The score ranks top-venue papers above the rest at an AUC of 0.79 — versus 0.64 for the novelty signal and 0.68 for the older quality score. Venue stays a sanity check, never an input: because the model targets in-corpus citations, it favors high-citation fields (NLP, vision) over theory, so we never read venue prestige back into the number.
Novelty Assessment
Each paper's novelty is assessed using a decomposed rubric with three independent components, reducing the bias inherent in single-judgment scoring:
Method Novelty
Does this introduce a new paradigm, a novel technique, or an incremental improvement?
Demonstrated Impact
Does it achieve broad state-of-the-art results, strong gains on specific benchmarks, or competitive performance?
Scope of Contribution
Is this a foundational advance applicable across fields, or a domain-specific improvement?
Component scores are summed into an internal novelty score from 0 to 100. Novelty is aseparate axis from the impact score — how new a method is barely predicts how much it gets cited — so it is not folded into the headline number; the paper page shows it on its own line (Highly original / Incremental advance / Synthesis-oriented), and it still drives the originality filter on Watches. This decomposition ensures that a paper with a truly novel method but narrow benchmarks is described differently from one with incremental methods but broad impact.
Personalized Feed
When you create an account, you can build a personalized “For You” feed that blends quality scores with your research interests.
Choose your research categories during onboarding
Like a few seed papers to calibrate your preferences
Your feed blends quality score with semantic similarity to your interests
Your personalized feed considers topic match, category affinity, and authors you follow. As you interact with papers, your preferences refine over time.
Privacy: Your preferences stay private. We don't share or sell interaction data.
What We Don't Do
No pay-to-rank — scores are purely algorithmic
No manual curation — all scoring is automated
No tracking across sites — we only use on-platform interactions
Open methodology — you're reading it