Methodology & scope
How papers get into Awesome Computer Vision, how they are ranked, and what this collection does not cover.
What is in the collection
Awesome Computer Vision currently indexes 120,771 papers. Papers are harvested from arXiv and Hugging Face Papers, with metadata and citation counts from OpenAlex and Semantic Scholar.
What is not in it
Coverage is recent, not historical. These collections were built by harvesting forward from the date each one started, so foundational papers published before that point are generally absent. On the LLM collection, for example, more than 99% of indexed papers are from 2023 onward, and landmark earlier work — the original Transformer paper, BERT, GPT-3 — is not in the index. Use this as a guide to current work, not as a complete literature record.
Papers are admitted by a per-collection classifier with a published confidence threshold. It is tuned to favour precision, so borderline and cross-listed papers are excluded rather than admitted. Some relevant work will be missing as a result.
What “community signal” means
Ranking uses a heat score: a weighted blend of five signals, normalised on a log scale and decayed by paper age. It measures attention, not quality or correctness.
| Signal | Weight | Notes |
|---|---|---|
| Citations | 35.3% | OpenAlex citation counts, refreshed on a rolling schedule. |
| GitHub stars | 23.5% | Stars on repositories linked from the paper. |
| Hugging Face likes | 17.6% | Likes on an associated Hugging Face paper page. |
| Bluesky | 11.8% | Mentions found on Bluesky. |
| Hacker News | 11.8% | Points on a Hacker News submission. |
Reddit was previously a weighted signal. Its public search endpoint began refusing unauthenticated requests in June 2026, so the term scored zero for every paper; it was removed in August 2026 and its share redistributed proportionally across the signals above, which left the relative ranking of papers unchanged.
Whether the heat score predicts citations
Until August 2026 we published no evidence that these weights were better than any other weights. They have now been tested. The result is mostly negative, and it is on this page because a ranking that has not been checked is worth less than one that has been checked and found wanting in specific, stated ways.
The test used a daily snapshot of every ranked paper’s signal values. Papers were scored using only data recorded on 11–30 June 2026, then checked against whether they had gained a citation by 31 July – 8 August 2026. That is a median window of 45 days, across 215,691 papers in total. The figures below use the 76,260 of them published in 2022 or earlier, of which 2,794 gained at least one citation; a paper published weeks before the window began cannot realistically register a citation inside it, so including the newest papers measures the lag in citation records rather than the ranking.
Citations are 35.3% of the score, so comparing the score against a citation outcome is partly circular. Every figure below is measured within groups of papers that held the same citation count at the start of the window, which removes that: a predictor derived from current citations alone scores 0.500 under this measure, and the citations-only baseline does come back at 0.500. Chance is 0.500. Higher is better, and differences below 0.001 are inside the measure’s own noise.
| Ranking | Score | 95% interval |
|---|---|---|
| Heat score (current weights) | 0.502 | 0.500 – 0.504 |
| Equal weights across the five signals | 0.502 | 0.500 – 0.504 |
| Heat score with the citation term removed | 0.502 | 0.500 – 0.504 |
| Citations alone | 0.500 | chance, by construction |
| Recency alone | 0.392 | 0.373 – 0.410 |
Papers published 2022 or earlier, where citation records have had time to settle. Figures for the full corpus, for each collection separately, and for the alternative weightings are in the linked results file.
The weights are doing almost no work. Replacing them with equal weights across the same five signals changes the ranking by less than 0.001 — indistinguishable from no change. The reason is coverage: GitHub stars are non-zero for 0.31% of indexed papers, Hugging Face likes for 0.66%, Hacker News for 0.64%, and Bluesky for 4.93%. Those four signals carry 64.7% of the weight between them and are absent from almost every paper, so for most of the corpus the score is the citation term plus a recency floor.
Beyond citations and publication date, the score adds nothing measurable. Among papers of the same age holding the same citation count, it scores 0.502 against a chance value of 0.500. It does beat ranking by recency alone. It does not beat ranking by citations alone, and no alternative weighting we tried beat it either.
This does not show that the ranking is bad at surfacing papers worth reading. The score measures attention, and the test only asks whether attention predicts near-term citations. Nor does it cover long horizons: six weeks is a short window for citation behaviour, and 96% of the papers tested gained nothing at all in it. One thing we could not test at all is the stored score itself, because it draws on a citation column we do not snapshot daily. The figures above come from a reconstruction of the score, checked against the production code, rather than from the stored value.
The analysis script and the full results, including the per-collection breakdown and the limitations in longer form, are in the repository under analysis/heat_score_validation.py. It is read-only and re-runnable.
Citations
Citation counts come from OpenAlex and are a snapshot, not a live figure. They are incomplete: a substantial share of indexed papers have no citation count at all, either because the paper is too recent or because it has no OpenAlex record. A missing count is shown as blank rather than as zero, because those are different things.
What is generated by a model
- Topic tags. Assigned by a classifier, refined by GPT where available; falls back to heuristic tags when it is not.
- Off-topic filtering. A GPT judge decides borderline in-scope/out-of-scope calls; a heuristic filter handles the rest.
- Daily digest & weekly long-form. Written by an LLM from the abstracts and metadata shown on the site. Every claim links to its source paper, and citations that no retrieval returned are stripped before sending.
- Learn curriculum. Concept ordering and resource selection are LLM-assisted, grounded in the indexed corpus.
- Research assistant. Answers are generated by an LLM restricted to the papers it retrieves from this corpus.
Paper titles, authors, abstracts, years and citation counts are taken from the source record and are not model-generated.
Leaderboards
Benchmark tables are fetched from the benchmark’s own publisher where one exists, and each board records its source. Numbers self-reported in papers are labelled as such and are excluded from the default ranking unless they state the same evaluation protocol as the board. Where a benchmark stopped publishing machine-readable results, the board is retired rather than left to go stale silently.
Corrections
If something here is wrong, it is a bug and worth reporting. This page describes what the system actually does; where the two disagree, the code is the authority and this page needs fixing.