ACM RecSys 2026 Proceedings by the numbers: More Industry, Bigger Teams, More Generative Everything

RecSys, the ACM Conference on Recommender Systems, is back in Minneapolis for its 20th edition, and the proceedings are not online. The first RecSys took place here in 2007. According to the official RecSys statistics, it received 35 long-paper submissions and accepted 16. This year, 361 long papers were submitted and 64 accepted.

I used ChatGPT and Claude to parse and analyze their 250 main contributions, so the derived statistics below are based on that AI-assisted analysis. The contributions include long and short papers, reproducibility and resource papers, Industry Track papers, Research and Practice Notes, demos, and a category called Past, Present and Future (PPF). I left out workshops, tutorials, and the doctoral symposium. For readability, I use “paper” below as shorthand for these 250 proceedings contributions.

To measure who writes these papers, I count bylines. One byline is one author on one paper, so someone on three papers has three bylines. The 250 papers have 1,510 bylines in total.

The number that surprised me most is 51.1%. That is the Industry Track’s share of all bylines. Its 82 papers make up 32.8% of the proceedings but carry 771 of the 1,510 bylines. A third of the papers, half the bylines.

Industry submissions doubled in one year

The Industry Track received 306 submissions this year, up from 149 in 2025, an increase of 105%. Its acceptance rate fell from about 36% to 26.8%. Because so many more papers were submitted, the number of accepted Industry papers still grew by roughly half, from about 54 in 2025 to 82 this year.

Long-paper submissions also jumped, from 261 in 2025 to 361 this year, an increase of 38.3%. Accepted long papers rose from 49 to 64, and the acceptance rate fell slightly, from 18.8% to 17.7%.

Submissions and acceptances for all categories in 2026:

Category	Submissions	Accepted	Acceptance rate
Long papers	361	64	17.7%
Short papers	152	28	18.4%
Past, Present and Future	37	13	35.1%
Industry Track	306	82	26.8%
Reproducibility and Resources	36	11	30.6%
Research and Practice Notes	77	34	44.2%
Demonstrations	27	18	66.7%

The three research-paper categories (long, short, and PPF) received 550 submissions together and accepted 105. That is a combined acceptance rate of 19.1%, compared with 26.8% in the Industry Track.

Who writes RecSys papers?

Behind the 1,510 bylines are 1,307 distinct authors, after merging obvious name variants.

Industry papers have larger teams. An Industry Track paper has 9.4 authors on average, compared with 5.1 for a long paper and 4.3 for a short paper. The largest team, 28 authors, wrote the Industry Track paper UniTraj: Cross-Domain Long-Sequence Modeling for Commercial Recommendation. At the other extreme, eleven papers have a single author.

Most bylines carry an industry affiliation. Of all bylines, 1,003 list only an industry affiliation and 475 list only an academic one. Another 19 list both. Including those dual affiliations, 1,022 bylines, or 67.7%, have an industry affiliation. The few remaining bylines belong to independent researchers and other organizations.

Counted per paper, the split is more even. 102 papers have only industry authors, 94 have only academic authors, and 51 have both. About one paper in five is an academia–industry collaboration.

Paper composition	Papers	Share
Industry only	102	40.8%
Academia only	94	37.6%
Academia–industry collaboration	51	20.4%
Other	3	1.2%

Inside the Industry Track, 12 of the 82 papers have at least one university author. Together they include 22 academic authors from 14 universities. The other 70 Industry papers have no academic author, and none of the 82 is written by academics alone.

Publishing in both the Industry Track and the research tracks is rarer than collaborating across sectors. I found 18 clearly identifiable authors who have an Industry Track paper and also a long, short, or PPF paper. That is about 2.6% of the 695 authors in the Industry Track.

Google has the most papers; Meta has the most bylines

For company counts, I grouped subsidiaries and naming variants into company families. Google includes Google LLC, Google DeepMind, and YouTube. Meta includes Meta Platforms, and Alibaba includes Taobao and Amap. A company counts once per paper if at least one author lists it.

Companies with at least five papers:

Company family	Papers	Bylines
Google	14	116
Meta	13	128
Kuaishou	12	91
Amazon	10	39
Alibaba	9	67
Pinterest	8	93
Microsoft	5	5
Spotify	5	47

Pinterest’s eight papers carry 93 bylines, nearly twelve names per paper. Microsoft is the opposite case: five papers, each with exactly one Microsoft author.

Counting people instead of bylines, Meta has 122 distinct authors, Google 101, Pinterest 85, Kuaishou 70, and Alibaba 50. The largest universities are much smaller. Fudan University, the University of Illinois Urbana-Champaign, and the University of Science and Technology of China have 12 distinct authors each.

Gyeongsang National University shows how far bylines and people can differ. It appears 29 times in the author lists, but those bylines belong to only 10 people. Two of them, Gun-Woo Kim and Sang-Min Choi, are on six papers each.

Where are the authors based? Roughly.

This is the least precise part of the analysis. A byline that says “Google” does not show whether the author works in Mountain View, Zurich, London, or elsewhere. The full paper PDFs may list addresses, but they are not yet available for every paper.

As a provisional rule, I assigned universities to their country and multinational companies to their headquarters. Some authors list more than one affiliation, so I count author–affiliation records rather than people. Under this rule, 69.8% of all records map to the US or China. In the Industry Track, 90.3% of the records map to companies or institutions headquartered in these two countries.

Some of this comes from the rule itself. Sweden, for example, ranks third overall with 39 records, but 38 of them are Spotify authors assigned to Stockholm.

The other tracks depend less on the headquarters rule, because more of their authors are at universities. Without the Industry Track, China and the US are almost exactly tied, with 166 records for China and 163 for the US. Among long papers alone, China accounts for 36.1% of affiliations and the US for 24.5%.

More than 1,300 authors, but no giant collaboration cluster

Almost nine in ten authors appear on exactly one paper: 1,159 of the 1,307. At the other end, Ludovico Boratto appears on seven papers, spread over long and short papers, PPF papers, resource papers, and demos. Ladislav Peška and the two Gyeongsang authors, Gun-Woo Kim and Sang-Min Choi, appear on six each. Three of the eleven single-author papers are by Teresa Zhang from Stanford.

To see how authors are connected, I built a coauthorship graph. Every author is a node, and two authors are linked when they share a paper. Two different people can have the same name, so here I split identical names when the affiliations indicate different people. That gives 1,320 authors in the graph instead of 1,307.

The graph has 5,771 distinct coauthor links and 166 connected components. A component is a group of authors linked to each other through chains of shared papers, but to no one outside the group.

The largest component has only 56 people, about 4% of the graph. It is almost entirely one Pinterest cluster: 55 Pinterest authors and one OpenAI author, linked through several Industry papers. Dhruvil Deven Badani, part of this cluster, has the most distinct coauthors: 38.

Within the 2026 proceedings, RecSys consists of many separate company and academic groups rather than one large connected collaboration cluster.

Generative in Industry, sequential in long papers

Titles give a rough idea of what the papers are about. Thirty-two titles contain “LLM”, “LLMs”, or “large language model”, and 17 contain “agent”, “agents”, or “agentic”.

I also counted title words, leaving out stopwords and generic words such as “recommendation”, “recommender”, and “system”. The most frequent word is “generative”, with 28 occurrences, followed by “sequential” (25), “retrieval” (20), “learning” (18), and “ranking” (17). In this count, a hyphenated word such as “re-ranking” counts as its own word, not as “ranking”.

Long papers and Industry papers use different words. “Sequential” appears in 15 of the 64 long-paper titles but in only two of the 82 Industry titles. “Generative” appears in 12 Industry titles.

Ten words and a colon

RecSys authors like colons: 149 of the 250 titles contain one. In 83 titles, a one-word name of a method or system comes first, followed by a colon, as in UniTraj:, CoFiRec:, or Melo:. That format alone accounts for a third of the proceedings.

Counting hyphenated words as one word, the shortest title is Tasteprint: Cross-Platform Recommendation Agent. The longest is Do We Really Need LLMs to Augment All? A Selective Augmentation Framework with Lightweight Language Models for Multimodal CTR Prediction, with 20 words and 136 characters.

Fourteen titles ask a question, and seven begin with “Towards”. One follows the “All You Need” tradition: Tokens are All You Need: Dual-purpose Semantic IDs for Achieving LLM-Level I/O Efficiency in Recommendation Systems.

The most common word of all is the less glamorous “for”, with 136 occurrences. And at least in titles, RecSys talks about users twice as often as about items. 22 titles contain “user” or “users”, and 11 contain “item” or “items”.

All title statistics in one table:

Title statistic	Result
Contains a colon	149 / 250 (59.6%)
Starts with a one-word name and a colon	83 / 250 (33.2%)
Contains a question mark	14 / 250 (5.6%)
Average title length	10.1 words
Median title length	10 words
Titles mentioning LLMs	32
Titles mentioning agents or agentic systems	17
Titles containing user/users	22
Titles containing item/items	11

If RecSys 2026 has a house style, it is about ten words, a colon, and a decent chance that something is generative.

How I counted

The analysis covers the 250 main contributions of the RecSys 2026 proceedings. These are the long papers, short papers, PPF papers, reproducibility and resource papers, Industry Track papers, Research and Practice Notes, and demos. Workshops, tutorials, and doctoral-symposium papers are excluded.

A byline counts a person once for each paper on which their name appears. Distinct-author counts merge obvious spelling and diacritic variants. Company statistics group obvious subsidiaries and naming variants into company families. The coauthorship graph uses a stricter identity rule and separates identical names when the affiliations indicate different people.

The academia–industry classification uses the affiliations printed in the proceedings. Five papers contain the placeholder affiliation “AXXX”, which I kept as printed. The count of 19 explicitly dual-affiliated bylines may therefore change once those affiliations are resolved.

The country statistics map universities to their country and companies to their headquarters when the proceedings give no office location. They describe institutions, not the nationality or physical location of authors. I plan to rerun this part once all paper PDFs are available.

Some figures come from published sources rather than my own counts. The chairs’ welcome in the proceedings provides the 2026 submission numbers and acceptance rates, as well as the 2025 Industry Track figures. The 2025 Industry acceptances are an estimate based on the reported 36% acceptance rate. The official RecSys historical statistics provide the 2007 and 2025 long-paper figures.