AI Research Statistics 2026

Dupple's research radar ingested 18,327 arXiv papers in 2026, and the single biggest theme was large language models, named in 4,152 of those titles (22.7%), just ahead of AI agents at 2,809 titles (15.3%). That is the shape of AI research this year: a firehose of papers, most of it orbiting language models and the agents built on top of them.

Most "AI research statistics" pages recycle the same third-party market-size guesses. This one does not. Every figure below comes from Dupple's own arXiv radar and newsletter pipeline, measured on August 19, 2026, and free to cite with attribution. For the tooling side of the story, see our companion AI tools statistics page.

The research radar in 2026

• 18,327 arXiv research papers were ingested by Dupple's radar in 2026, measured by ingestion date through August 19, 2026. Source: Dupple radar pipeline, RadarItem table, source platform arXiv.
• arXiv coverage began on November 17, 2025, and has logged 19,705 papers in total, of which 1,378 landed in the final six weeks of 2025. There is no full prior-year baseline to compare against, because the feed is younger than a year. Source: Dupple radar, RadarItem created dates.
• arXiv is the only pure-research feed among the six sources in Dupple's radar, which holds 214,623 items in all. The other five are community and product signals: Reddit (99,309), Lemmy (66,595), Hacker News (25,140), Product Hunt (3,506), and GitHub (368). arXiv accounts for 9.2% of every item the radar has ever pulled. Source: Dupple radar, source platform counts.
• The radar averaged about 80 new arXiv papers a day across 2026, roughly 558 a week, taken over the 230 days from January 1 to August 19. The real pace is now far higher than that yearly average. Source: Dupple radar, 18,327 papers over 230 days.
• Two months, July and August, produced 12,379 papers, or 67.5% of the entire year's intake, as the research radar widened its coverage. Monthly volume held near 900 to 1,200 from January through June, then jumped to 6,507 in July and 5,872 in the first 19 days of August. Source: Dupple radar, monthly ingestion counts.

What researchers published about in 2026

These are title-keyword matches across the 18,327 arXiv papers ingested in 2026. A single paper can match more than one theme, so shares do not sum to 100%.

• Large language models are the dominant theme, appearing in 4,152 paper titles (22.7%) when counting "LLM" and "language model" together. Source: Dupple radar, 2026 arXiv title keywords.
• AI agents are second, in 2,809 titles (15.3%), counting "agent" and "agentic". If you want the plain-English version of why that number is exploding, start with our explainer on agentic AI. Source: Dupple radar, 2026 arXiv title keywords.
• "Benchmark" appears in 851 titles (4.6%), a sign of how much 2026 research is about measuring models rather than building them. Source: Dupple radar, 2026 arXiv title keywords.
• "Reasoning" appears in 839 titles (4.6%), essentially tied with benchmarking. Source: Dupple radar, 2026 arXiv title keywords.
• Graph methods appear in 794 titles (4.3%), keeping graph learning among the busiest classical areas. Source: Dupple radar, 2026 arXiv title keywords.
• Safety and alignment together appear in 645 titles (3.5%), counting "safety" and "alignment". Source: Dupple radar, 2026 arXiv title keywords.
• "Multimodal" appears in 508 titles (2.8%), counting "multimodal" and "multi-modal". Source: Dupple radar, 2026 arXiv title keywords.
• Retrieval and RAG appear in 433 titles (2.4%), counting "retrieval" and "retrieval-augmented". Source: Dupple radar, 2026 arXiv title keywords.
• "Vision" appears in 414 titles (2.3%) and reinforcement learning in 371 (2.0%), with diffusion close behind at 318 (1.7%). Source: Dupple radar, 2026 arXiv title keywords.
• The open-weights wave shows up directly in the paper stream too. The same models driving open-weights releases like Alibaba's Qwen 3 are what most of those 4,152 language-model titles are studying. Source: Dupple radar, 2026 arXiv title keywords.

From arXiv to the newsletter

• 498 of the 18,327 arXiv papers, about 2.7%, were selected into a Dupple newsletter at least once in 2026, for 583 total selections across editions. Source: Dupple radar, RadarItem pick counts.
• The arXiv feed powers dedicated research sections across at least four newsletters, including Techpresso's "Latest Research", Cyberpresso's research papers, Devshot's papers and resources, and AIpresso. Source: Dupple radar watcher groups.
• The single most-surfaced paper of 2026 was picked into Dupple editions five times, the ceiling for any one arXiv paper this year. Source: Dupple radar, RadarItem pick counts.
• The research a model routes to matters as much as the model itself. The plumbing that connects these papers to real products is consolidating fast, as when Stripe acquired OpenRouter. Source: Dupple newsletter coverage.

How this data was measured

These figures are read directly from Dupple's radar pipeline and newsletter database on August 19, 2026. The paper count uses the ingestion date, not the arXiv submission date, so it reflects when the radar caught each paper. Theme counts are title-keyword matches, which undercount true topics, because a paper can discuss LLMs or agents without naming them in the title, and a single title can match several themes. Selection counts measure how many arXiv papers the editorial radar surfaced into newsletter drafts at least once.

You are free to cite any number on this page with a link back to it. For the daily read on which papers, models, and tools are moving, AIpresso sends a free AI brief every morning, and you can browse the products built on this research over at Toolradar.