How the radar works
Coverage
Every day the pipeline asks the arXiv API for all papers first submitted in these categories (primary or cross-listed): cs.*, eess.SY, eess.SP, eess.AS, eess.IV. The last 8 days are re-fetched on every run, because arXiv publishes a day's submissions only after its evening announcement and updates (new versions, acceptance notes) arrive later. Older days that are missing or were stored incompletely are repaired automatically.
The score
Each paper gets up to eight signals in [0, 1]; the score is their weighted mean. Every paper page shows the raw numbers, the signal values, and the evidence behind each one.
| Signal | Weight | Source | What it measures |
|---|---|---|---|
| Citations | 20% | Semantic Scholar | Citations plus twice the influential citations (citations that build on the work, not just mention it). |
| Upvotes | 10% | Hugging Face daily papers | Community upvotes. Mostly covers ML, vision and NLP. |
| Code stars | 10% | GitHub | Stars of the repository linked from the paper. |
| Venue | 15% | Semantic Scholar, arXiv, CORE ranking | Where it is published, tiered by the CORE conference ranking (A* 1.0, A 0.85, B 0.65, C 0.45; journals by family). Verified venues count fully; "accepted at X" written by the authors counts 90%; workshops 60%; "under review" almost nothing. |
| Code/data | 10% | the paper | Code, data, models or a project page are released. |
| Novelty | 15% | embeddings | Distance to the closest paper of the previous 90 days. |
| Momentum | 15% | embeddings | Growth of the paper's research neighbourhood this week versus the previous 4 weeks. |
| Substance | 5% | the paper | Full paper (pages, figures, journal version) rather than a short note. |
Fairness rules
- Same age, same field. Citations, upvotes, stars, novelty and momentum are percentiles within the paper's cohort: papers of the same research area submitted in the same week. A theory paper is compared with theory papers of its week, never with ML papers that are cited ten times faster.
- Missing is not zero. A signal is used for a cohort only if it carries information there (at least 3% of the cohort has a non-zero value). Otherwise it is dropped for the whole cohort and the other weights are rescaled. Each score shows how much of the weight was actually measured.
- Verified over claimed. Venue from Semantic Scholar or a journal reference beats an acceptance note in the arXiv comment.
- No prestige shortcuts. Author h-index, team size and number of cross-listed categories are shown for context but are not part of the score: they reward famous groups and are easy to game.
- Scores mature. Evidence keeps arriving, so the last 13 weeks are re-scored on every update. Rankings of the current week are provisional; older weeks settle as citations accumulate.
Fairness check — 2026-W46
Share of the week's papers vs. share of its top 50, and how much of the score could be measured. Live data: /api/fairness.
| Area | Papers | Share | Share of top 50 | Measured weight | Signals measured (share of papers) |
|---|---|---|---|---|---|
| AI & Machine Learning | 1234 | 31% | 30% | 80% | Citations 0% |
| Language, Vision & Media | 1040 | 26% | 28% | 100% | all |
| Robotics, Control & Signals | 624 | 16% | 16% | 70% | Citations 0% · Upvotes 0% |
| Theory & Algorithms | 307 | 8% | 6% | 80% | Upvotes 0% · Code stars 0% |
| Systems & Networking | 246 | 6% | 6% | 80% | Upvotes 0% · Code stars 0% |
| People & Society | 188 | 5% | 6% | 60% | Citations 0% · Upvotes 0% · Code stars 0% |
| Software & Programming | 173 | 4% | 4% | 60% | Citations 0% · Upvotes 0% · Code stars 0% |
| Security & Privacy | 153 | 4% | 4% | 60% | Citations 0% · Upvotes 0% · Code stars 0% |
Weekly selection. Because every score is a percentile within its own area, a fair list should mirror the week: each research area gets slots in the weekly top 20 in proportion to its share of the week's papers (largest-remainder method, at least one slot per area), and the slots are filled by score. Near-duplicates are skipped; unfilled slots go to the best remaining papers. Weights are configurable (SCORE_WEIGHTS); TOP_SELECTION=cap switches to a plain ranking with at most 4 papers per category.
Topics
A topic query is matched two ways: BM25 keyword search over titles and abstracts (the query words close together, falling back to all words anywhere), and semantic search with sentence embeddings (sentence-transformers/all-MiniLM-L6-v2, run locally). Results are fused by reciprocal rank. For the matching papers the radar shows:
- What's new — the most recent matching papers.
- Frontier — the highest frontier scores among clearly relevant papers in the window.
- Emerging terms — phrases whose share of the topic's papers rose most compared with the preceding period (log-odds z-score).
- Research directions — clusters of the topic's papers (spherical k-means on embeddings), labelled by their most distinctive phrases.
- Trend — the topic's share of all CS/CE papers per week.
API
Everything on the site is available as JSON — see the interactive API docs. For example /api/weekly?n=20, /api/daily?date=YYYY-MM-DD&cat=cs.CR, /api/topic?q=network%20security.
Limits
- The score combines measurable evidence; it does not read the paper. In the first days after submission there are few citations, so the newest week leans on venue, released artifacts, novelty and momentum.
- Venue detection relies on what authors write in the arXiv comment field.
- Days are UTC submission days, not arXiv announcement days.