Most cited AI papers of 2025: ranked by citation count
The Qwen3 Technical Report leads this ranking of AI papers first published in 2025, with 9,442 Semantic Scholar citations as of 21 October 2026. Qwen2.5-VL follows with 6,307, then DeepSeek-R1 with 6,092.
Here, “2025” means the paper’s first publication year—not the year in which its citations were received. The ranking starts with Semantic Scholar’s most-cited Computer Science records for 2025, filters for AI research and includes each work once.
Semantic Scholar determines the order below. OpenAlex counts are shown alongside it because the databases report substantially different totals. For example, OpenAlex reports 147 citations for Qwen3. Those differences matter, but they do not establish which database is more complete or why the counts diverge.
What are the top ten most cited AI papers of 2025?
| Rank | Paper | Semantic Scholar citations | Influential citations (share) | OpenAlex citations | Semantic Scholar / OpenAlex | First arXiv posting | arXiv primary subject |
|---|---|---|---|---|---|---|---|
| 1 | Qwen3 Technical Report | 9,442 | 1,381 (14.6%) | 147 | 64.2x | 14 May 2025 | cs.CL |
| 2 | Qwen2.5-VL Technical Report | 6,307 | 1,420 (22.5%) | 92 | 68.6x | 19 Feb 2025 | cs.CV |
| 3 | DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning | 6,092 | 983 (16.1%) | 539 | 11.3x | 22 Jan 2025 | cs.CL |
| 4 | Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities | 4,415 | 1,128 (25.5%) | 64 | 69.0x | 7 Jul 2025 | cs.CL |
| 5 | DAPO: An Open-Source LLM Reinforcement Learning System at Scale | 3,007 | 617 (20.5%) | 16 | 187.9x | 18 Mar 2025 | cs.LG |
| 6 | Wan: Open and Advanced Large-Scale Video Generative Models | 2,951 | 882 (29.9%) | 14 | 210.8x | 26 Mar 2025 | cs.CV |
| 7 | Qwen3-VL Technical Report | 2,880 | 593 (20.6%) | 26 | 110.8x | 26 Nov 2025 | cs.CV |
| 8 | YOLOv12: Attention-Centric Real-Time Object Detectors | 2,760 | 249 (9.0%) | 801 | 3.4x | 18 Feb 2025 | cs.CV |
| 9 | π0.5: a Vision-Language-Action Model with Open-World Generalization | 2,450 | 472 (19.3%) | 3 | 816.7x | 22 Apr 2025 | cs.LG |
| 10 | Gemma 3 Technical Report | 2,088 | 354 (17.0%) | 72 | 29.0x | 25 Mar 2025 | cs.CL |
Source and snapshot: Semantic Scholar and OpenAlex APIs, with first-posting dates and primary subjects from the papers’ arXiv pages, read on 21 October 2026. The linked records below allow readers to inspect the sources; their current values may differ from this snapshot.
Why is Qwen3 the leading answer?
Qwen3 has the largest Semantic Scholar citation count among the eligible works in this ranking. It qualifies because its first arXiv posting was in 2025 and its primary subject, cs.CL, falls within the AI categories used here.
That conclusion depends on the question’s metric:
- Total Semantic Scholar citations: Qwen3 leads with 9,442.
- Influential-citation count among these papers: Qwen2.5-VL leads with 1,420, ahead of Qwen3’s 1,381.
- Influential-citation share among these papers: Wan leads at 29.9%; YOLOv12 has the lowest share at 9.0%.
- OpenAlex counts: these do not reproduce the Semantic Scholar ordering. For example, YOLOv12 has 801, compared with Qwen3’s 147.
Thus, Qwen3 is the answer to the stated total-citation ranking, not a claim that it leads every measure of research impact.
How were the papers selected?
The starting pool was Semantic Scholar’s top 60 Computer Science records for 2025 by citation count. Merging duplicates left 59 distinct works; 48 were retained and 11 excluded.
The selection applies these rules:
- First published in 2025. Semantic Scholar’s year is based on the first version, rather than a later publication date.
- AI by primary subject. For works with an arXiv version, eligible primary categories are cs.AI, cs.LG, cs.CL, cs.CV, cs.RO, cs.IR, cs.MA, cs.NE and stat.ML. OpenAlex’s subfield is used only when no arXiv version exists.
- One entry per work. Separate records for the same work are not treated as separate ranked papers or added together.
The subject rule matters because the initial Computer Science pool also contains material outside the chosen AI-research scope. Exclusions included 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.
Conversely, π0.5 qualifies through its arXiv primary subject, cs.LG, even though OpenAlex files it under Control and Systems Engineering.
This is therefore a ranking drawn from a defined candidate pool—not a claim to have exhaustively checked every AI paper published in 2025.
Why do the database counts differ?
The verified records establish the differences, but not a general explanation for them. They do not justify claiming that either database always merges versions, always separates them or systematically misses a particular publication type.
Within the top ten, the Semantic Scholar-to-OpenAlex ratio ranges from 3.4x for YOLOv12 to 816.7x for π0.5. Because the discrepancies vary by paper, the counts should not be treated as interchangeable or converted using a common multiplier.
DeepSeek-R1: keep the records separate
DeepSeek-R1 demonstrates why record handling must be explicit:
- Semantic Scholar: the arXiv preprint has 6,092 citations; the Nature paper has 5,918.
- OpenAlex: the preprint has 539 citations; the Nature paper has 1,122.
The ranking uses the 6,092-citation Semantic Scholar preprint record and includes the work once. The separate record totals must not be added. Both records are linked below so the choice is inspectable.
Where can you verify each paper?
Each entry links to its arXiv page and the database records used for comparison.
- Qwen3 Technical Report — arXiv:2505.09388 · Semantic Scholar · OpenAlex
- Qwen2.5-VL Technical Report — arXiv:2502.13923 · Semantic Scholar · OpenAlex
- DeepSeek-R1 — arXiv:2501.12948 · Semantic Scholar, preprint record · Semantic Scholar, Nature record · OpenAlex, preprint · OpenAlex, Nature
- Gemini 2.5 — arXiv:2507.06261 · Semantic Scholar · OpenAlex
- DAPO — arXiv:2503.14476 · Semantic Scholar · OpenAlex
- Wan — arXiv:2503.20314 · Semantic Scholar · OpenAlex
- Qwen3-VL Technical Report — arXiv:2511.21631 · Semantic Scholar · OpenAlex
- YOLOv12 — arXiv:2502.12524 · Semantic Scholar · OpenAlex
- π0.5 — arXiv:2504.16054 · Semantic Scholar · OpenAlex
- Gemma 3 Technical Report — arXiv:2503.19786 · Semantic Scholar · OpenAlex
Which paper just missed the top ten?
VGGT: Visual Geometry Grounded Transformer had 2,051 Semantic Scholar citations, 37 behind Gemma 3. It also had 522 influential citations and 473 OpenAlex citations.
This illustrates why the columns answer different questions: VGGT sits outside the top ten by total Semantic Scholar citations despite having more influential citations than some included papers.
How should you interpret the ranking?
- Keep the date attached. These are counts read on 21 October 2026, not year-end 2025 totals.
- Keep the database attached. An OpenAlex count is not a substitute for the Semantic Scholar count used to assign rank.
- Keep the selection boundary attached. Eligibility follows first-publication dates and specified subject categories, rather than an informal assessment of whether a paper concerns AI.
- Do not infer citation speed from rank. The papers have different first-posting dates; this table does not calculate citation rates.
- Do not equate citation totals with quality. The table ranks recorded citations, not correctness, usefulness or research merit.
Frequently asked questions
The Qwen3 Technical Report leads this ranking with 9,442 Semantic Scholar citations as of 21 October 2026. It was first posted on 14 May 2025. The ranking covers eligible AI works drawn from Semantic Scholar's top 60 Computer Science records for 2025.
By Semantic Scholar citation count in this ranking: 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). Counts were read on 21 October 2026.
No. It means the works were first published in 2025. The citation totals are the database counts read on 21 October 2026, not citations received only during 2025.
The checked records report different totals, but these observations alone do not establish the cause. Qwen3 has 9,442 Semantic Scholar citations and 147 OpenAlex citations. DeepSeek-R1 also has separate preprint and Nature records in both databases, so the exact record must be specified.
On 21 October 2026, Semantic Scholar reported 6,092 for its arXiv preprint and 5,918 for its Nature paper. OpenAlex reported 539 for the preprint and 1,122 for the Nature paper. This ranking uses the 6,092-citation Semantic Scholar record; the separate totals must not be added.
Among the top ten listed here, Qwen2.5-VL has the most Semantic Scholar influential citations: 1,420, compared with Qwen3's 1,381. Wan has the highest influential-citation share at 29.9%.
Related reading: AI search visibility
This article ranks academic paper citations, not appearances in AI-generated search answers. For that separate topic, see what AI visibility is and how to track brand mentions in AI search.