2026 edition
How the lists are made
Every paper's place comes from one scoring formula, applied the same way to every paper. A few things a formula can't judge, such as whether a paper is readable for a learner, are decided by hand; each of those decisions is shown, with its reason, on the paper's own page. None of them moves a paper up or down the Top 500.
How a paper is scored
Each paper gets a score between 0 and 1, made of five parts:
| What counts | Weight | How it's measured |
|---|---|---|
| Named by the expert lists | 30% | Two AI models, ChatGPT and Grok, each wrote ranked top-50 lists for 19 fields, plus a must-read list and a list of recent landmarks. The nearer the top a paper appears, the more it scores (first place 1.0, fiftieth 0.02, an honourable mention 0.15). Each model's best placement counts, and the two are averaged. |
| Must-read votes | 10% | Whether neither, one or both models put the paper on their lifetime reading list. |
| Citations for its era | 35% | How often the paper is cited per year since it came out, compared with papers from the same decade and field. A 1950s classic competes with other 1950s papers, not with 2020s ones. Papers before 1970 are compared together, and a decade with too few papers in one field is compared across all fields. |
| Test-of-time award | 15% | Whether the paper won a test-of-time or classic-paper award (CVPR, ICCV, NeurIPS, ICML, ICLR, ACL, ECCV, AAAI, KDD, SIGIR, IJCAI-JAIR). |
| On existing must-read lists | 10% | Whether it appears on Ilya Sutskever's reading list (as reconstructed by others) or on the list of most-cited deep learning papers. |
40% of the score comes from what the two AI models named as important. We say so openly; the rest comes from published records: citations, awards and existing lists. 77 of the top 100 papers were first released in 2010 or later.
Which papers can be ranked
- Research papers only. Books, surveys, reports and essays have their own shelves.
- First released in 2025 or earlier, counting from its first public version, often a preprint. The year it appeared in a journal or conference is shown beside it. Papers from 2026 go on the recent shelf.
- A permanent identifier (a DOI, an arXiv number or an OpenAlex record), so that it has its own page here.
- Named by both AI models; or named by one and also backed by an award, a place on an existing must-read list, or at least median citations for its era.
- Versions of the same paper (preprint, conference, journal) count once.
911 of the 1585 works we considered qualified. When two papers score the same, the more cited one ranks higher.
100 papers to read in your lifetime
A reading list, not simply the top of the Top 500. It starts from the papers the AI models flagged as must-reads, ordered by those votes and then by score. At least 20 places are kept for papers from before 1990 (this edition has 20). Every paper passed a readability check: does it teach something lasting, and can a learner with the listed prerequisites get through it? Each verdict and its reason are on the paper's page, along with a difficulty level: introductory, intermediate or advanced. 76 of the 100 are also in the Top 500's first 100.
Field pages
The 19 fields the AI models wrote about overlap, so they are merged into 12 pages. Each paper is ranked on the page where it was named most often, and shown as “also relevant” on up to 12 others. Bigger fields get longer lists: about 40% of a field's candidates, between 25 and 100 papers.
- Machine learning foundations — Learning theory & classic ML, Probabilistic ML · 194 candidates, top 80 listed
- Deep learning — Architectures & training, Efficient ML & systems · 159 candidates, top 65 listed
- Computer vision & multimodal — Computer vision, Multimodal & vision-language, Self-supervised learning · 210 candidates, top 85 listed
- Language, speech & LLMs — Large language models, Natural language processing, Speech & audio · 245 candidates, top 100 listed
- Reinforcement learning — Reinforcement learning · 83 candidates, top 35 listed
- Generative models — Generative models · 81 candidates, top 30 listed
- Graph learning — Graph learning · 78 candidates, top 30 listed
- Robotics — Robotics · 93 candidates, top 35 listed
- Classic AI: search, logic & planning — Classic AI · 95 candidates, top 40 listed
- Search, retrieval & recommendation — Retrieval & recommendation · 92 candidates, top 35 listed
- Safety, alignment & interpretability — Interpretability, Safety & alignment · 171 candidates, top 70 listed
- AI for science — AI for science · 84 candidates, top 35 listed
Editions and corrections
The Top 500, the lifetime list and the field pages are fixed once a year; this is the 2026 edition. Clear mistakes are corrected in place, with a note. The recent shelf is updated between editions. Votes or ratings from visitors never move a paper, and your “I've read it” ticks stay in your browser.
Where the data comes from
- OpenAlex: identifiers, release dates and citation counts (CC0).
- Crossref: DOIs and publication years (CC0). We never show abstracts.
- arXiv: identifiers and release months.
- Papers With Code: citation counts and arXiv links (CC BY-SA 4.0).
Semantic Scholar: citation counts, under its API licence, which allows non-commercial use with this credit. While Semantic Scholar data is shown, the reading lists stay non-commercial.- Award winners and existing must-read lists, taken from their official pages.
Where sources disagree on a citation count, we show the largest and name its source.
Data sources
Catalogue data from OpenAlex (CC0), Crossref (CC0 metadata; no abstracts are shown), arXiv, and Papers With Code (CC BY-SA 4.0), and Semantic Scholar (used under its non-commercial API licence). Each count names its source. See the method.