On 21 October 2026, the most cited AI papers first published in 2025 were the Qwen3 Technical Report, the Qwen2.5-VL Technical Report, DeepSeek-R1, Gemini 2.5, DAPO, Wan, the Qwen3-VL Technical Report, YOLOv12, π0.5 and the Gemma 3 Technical Report, ranked by Semantic Scholar citation counts. Qwen3 leads with 9,442 citations. OpenAlex, the other large open citation database, gives the same paper 147.
That gap is the most important fact on this page. Ask a different database and you get a different list — not a reshuffled order, but mostly different papers. Every figure below links to the record it was read from, so you can check any of them yourself.
What are the most cited AI papers of 2025?
| Rank | Paper | Semantic Scholar citations | Influential citations | OpenAlex citations | First posted |
|---|---|---|---|---|---|
| 1 | Qwen3 Technical Report | 9,442 | 1,381 | 147 | 14 May 2025 |
| 2 | Qwen2.5-VL Technical Report | 6,307 | 1,420 | 92 | 19 Feb 2025 |
| 3 | DeepSeek-R1 | 6,092 | 983 | 539 | 22 Jan 2025 |
| 4 | Gemini 2.5 | 4,415 | 1,128 | 64 | 7 Jul 2025 |
| 5 | DAPO | 3,007 | 617 | 16 | 18 Mar 2025 |
| 6 | Wan | 2,951 | 882 | 14 | 26 Mar 2025 |
| 7 | Qwen3-VL Technical Report | 2,880 | 593 | 26 | 26 Nov 2025 |
| 8 | YOLOv12 | 2,760 | 249 | 801 | 18 Feb 2025 |
| 9 | π0.5 | 2,450 | 472 | 3 | 22 Apr 2025 |
| 10 | Gemma 3 Technical Report | 2,088 | 354 | 72 | 25 Mar 2025 |
All counts were read on 21 October 2026. "First posted" is the date of the paper's first arXiv version. Citation counts rise every week, so treat the table as a dated snapshot, not a standing fact.
Where can you verify each paper's numbers?
Each paper links to its arXiv page, its Semantic Scholar record and its OpenAlex record.
- Qwen3 Technical Report — arXiv · Semantic Scholar · OpenAlex
- Qwen2.5-VL Technical Report — arXiv · Semantic Scholar · OpenAlex
- DeepSeek-R1 — arXiv · Semantic Scholar, preprint record · Semantic Scholar, Nature record · OpenAlex, preprint · OpenAlex, Nature
- Gemini 2.5 — arXiv · Semantic Scholar · OpenAlex
- DAPO — arXiv · Semantic Scholar · OpenAlex
- Wan — arXiv · Semantic Scholar · OpenAlex
- Qwen3-VL Technical Report — arXiv · Semantic Scholar · OpenAlex
- YOLOv12 — arXiv · Semantic Scholar · OpenAlex
- π0.5 — arXiv · Semantic Scholar · OpenAlex
- Gemma 3 Technical Report — arXiv · Semantic Scholar · OpenAlex
Why does OpenAlex count these papers so differently?
Because almost all of them were published on arXiv rather than in a journal, and the two databases treat arXiv very differently.
Semantic Scholar dates a paper by its first version and merges a preprint with its later published version into one count. OpenAlex counts each version as its own record, and its preprint records collect few of the citations. For a field where the most important work appears on arXiv first and often only there, that one design choice decides the list.
The size of the gap follows from where each paper was published:
| Paper | Semantic Scholar | OpenAlex | Ratio |
|---|---|---|---|
| π0.5 | 2,450 | 3 | 816.7x |
| Wan | 2,951 | 14 | 210.8x |
| DAPO | 3,007 | 16 | 187.9x |
| Qwen3 Technical Report | 9,442 | 147 | 64.2x |
| DeepSeek-R1 | 6,092 | 539 | 11.3x |
| YOLOv12 | 2,760 | 801 | 3.4x |
YOLOv12 is the closest the two databases get, at 3.4 times. π0.5 is the furthest: 2,450 citations in Semantic Scholar against 3 in OpenAlex.
The practical consequence is stark. Rank 2025 AI papers by OpenAlex and the frontier technical reports drop out of the top ten almost entirely, replaced by journal surveys and reviews — because journals are what OpenAlex counts well. Neither database is wrong about what it counts. They count different things, and only one of them counts where AI research is actually published.
How were these papers selected?
The ranking starts from Semantic Scholar's own list of 2025 Computer Science papers sorted by citations — the top 60 records — and applies three mechanical rules, so that no place depends on a judgement call.
- First published in 2025. Semantic Scholar dates a paper by its first version, preprints included, so a 2023 preprint published in a journal in 2025 does not count as a 2025 paper.
- AI research, by the authors' own classification. For a paper on arXiv, the test is the primary subject its authors filed it under: cs.AI, cs.LG, cs.CL, cs.CV, cs.RO, cs.IR, cs.MA, cs.NE or stat.ML. Only a paper with no arXiv version falls back to OpenAlex's subfield.
- One entry per paper. DeepSeek-R1 appears in Semantic Scholar twice — as the arXiv preprint (6,092 citations) and as the Nature paper (5,918). Their citations overlap, so they are not added together; the entry carries the larger count.
Of the 60 records, 59 distinct works remained after merging, 48 passed the subject test and 11 did not. The excluded ones are a good illustration of why the test is needed: Semantic Scholar's "Computer Science" field includes the TITAN guideline for reporting AI in medicine (2,348 citations), a surgical case-report guideline (1,733), a statistics textbook and a chemistry software release. None of them is AI research.
The subject test also corrects OpenAlex the other way. OpenAlex files π0.5 under Control and Systems Engineering; its authors filed it under Machine Learning (cs.LG) on arXiv.
What do influential citations change?
Semantic Scholar marks a citation as influential when the citing paper builds on the work substantially, rather than mentioning it in passing. Ranked that way, the top of the list changes:
- Qwen2.5-VL moves to first, with 1,420 influential citations against Qwen3's 1,381 — despite having about two thirds of Qwen3's total.
- Gemini 2.5 has 1,128, ahead of DeepSeek-R1's 983.
- Wan has the highest influential share in the top ten, at 29.9% of its citations.
- YOLOv12 has the lowest, at 9.0% — widely cited, less often built upon.
A paper cited mainly as a baseline to beat collects citations quickly but influential citations slowly. Read the two columns together.
What is each of the top ten?
1. Qwen3 Technical Report — 9,442 citations
The latest version of the Qwen family of large language models. First posted 14 May 2025.
2. Qwen2.5-VL Technical Report — 6,307 citations
The flagship model of the Qwen vision-language series. The highest influential-citation count of any 2025 AI paper: 1,420.
3. DeepSeek-R1 — 6,092 citations
Reasoning in large language models elicited through reinforcement learning. The earliest paper in the top ten, first posted 22 January 2025, and the only one also published in Nature.
4. Gemini 2.5 — 4,415 citations
The technical report for the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, alongside the earlier Gemini 2.0 Flash and Flash-Lite.
5. DAPO — 3,007 citations
An open-source system for large-scale reinforcement learning of large language models, aimed at eliciting reasoning.
6. Wan — 2,951 citations
An open suite of video foundation models for video generation.
7. Qwen3-VL Technical Report — 2,880 citations
The most capable vision-language model in the Qwen series at the time of its report. First posted 26 November 2025 — the newest paper in the top ten, and the fastest to this many citations.
8. YOLOv12 — 2,760 citations
A real-time object detector that moves the YOLO framework from convolution-based designs to attention.
9. π0.5 — 2,450 citations
A vision-language-action model for robots, aimed at performing tasks in real-world settings outside the lab.
10. Gemma 3 Technical Report — 2,088 citations
A multimodal addition to the Gemma family of lightweight open models, ranging from 1 to 27 billion parameters.
Which paper just missed the top ten?
VGGT: Visual Geometry Grounded Transformer, at 2,051 citations — 37 behind Gemma 3. A feed-forward network that infers a scene's 3D attributes directly. It has 522 influential citations, more than three of the papers above it, and 473 in OpenAlex, one of the smallest gaps between the two databases.
What do these measures count?
| Measure | What it counts |
|---|---|
| Citation count | Every paper that cites this one, however it cites it. |
| Influential citation | Semantic Scholar's estimate of a citation that builds on the work substantially rather than mentioning it. |
| Preprint | A version posted publicly, usually to arXiv, before or instead of journal publication. |
| Version merging | Counting a preprint and its published version as one paper. Semantic Scholar does; OpenAlex keeps them as separate records. |
What are the limits of this ranking?
- It is one database's ranking. Semantic Scholar is the primary count because it handles arXiv properly. Google Scholar, which is not openly queryable, would give different numbers again.
- Recency is a handicap. A paper from January has had ten more months to collect citations than one from November. Qwen3-VL's seventh place is more impressive than its rank suggests.
- The subject boundary is a rule, not a fact. The arXiv categories used here include robotics (cs.RO) and information retrieval (cs.IR). A narrower definition of AI would give a slightly different list.
- Citation counts do not measure quality. A paper cited mainly as a baseline to beat scores the same as one that changed how others work. That is what the influential-citation column is for.
How is this different from being cited by an AI search engine?
Academic citations count how often researchers reference a paper. AI-search citation is a different measure: whether an assistant names or links your page when someone asks a question. Asked for the most cited AI papers of 2025, AI search engines mostly cite the papers' own arXiv pages — which is a reminder that the most authoritative source on a topic tends to be the one that gets cited. If that is the citation you care about, start with what AI visibility is or how to track brand mentions in AI search.
Frequently asked questions
By Semantic Scholar on 21 October 2026, the Qwen3 Technical Report, with 9,442 citations. It was first posted to arXiv on 14 May 2025. OpenAlex gives the same paper 147 citations, because it barely counts arXiv preprints.
By Semantic Scholar citation counts on 21 October 2026: Qwen3 Technical Report (9,442), Qwen2.5-VL Technical Report (6,307), DeepSeek-R1 (6,092), Gemini 2.5 (4,415), DAPO (3,007), Wan (2,951), Qwen3-VL Technical Report (2,880), YOLOv12 (2,760), π0.5 (2,450) and Gemma 3 Technical Report (2,088).
Semantic Scholar merges a paper's arXiv preprint and published versions and dates it by the first version. OpenAlex keeps each version as a separate record, and its preprint records collect few citations. For AI, where most important work is published on arXiv first, the gap is large: π0.5 has 2,450 citations in Semantic Scholar and 3 in OpenAlex.
Semantic Scholar holds it under two records: the arXiv preprint with 6,092 citations and the Nature paper with 5,918. Their citations overlap, so they should not be added. OpenAlex gives 539 for the preprint and 1,122 for the Nature paper.
Semantic Scholar's estimate of citations that build on a paper substantially, rather than mention it in passing. By that measure the Qwen2.5-VL Technical Report leads 2025 AI papers, with 1,420, ahead of Qwen3's 1,381.
Academic citations count researchers referencing a paper. AI search visibility is whether an AI assistant names or links your page when someone asks a question. They are measured differently: the second is tracked across a set of real queries, by which sources and brands get cited.
The short version
On 21 October 2026 the most cited AI paper of 2025 was the Qwen3 Technical Report, with 9,442 citations in Semantic Scholar, followed by Qwen2.5-VL, DeepSeek-R1 and Gemini 2.5. The ranking depends heavily on the database: OpenAlex, which barely counts arXiv preprints, gives Qwen3 just 147 citations, and an OpenAlex-ranked list would leave out most of these papers entirely. Every figure here links to its record.