World Tech Rankings
All technologies

Large language models

AI systems trained on very large amounts of text that can write, answer questions and write code.

China leads large language models research, with 27.4% of the 10% most cited papers published 2021–2025; the United States is second with 19.4%.

This field publishes mainly at conferences and as preprints, so this one technology counts journals, conference papers and preprints from any source, not only core journals. A paper posted as a preprint and later published can appear twice.

CN
FirstChina27.4%
US
SecondUnited States19.4%
1.4×lead, by share of the 10% most cited papers, 2021–2025

Standings

1CNChina27.4%2USUnited States19.4%3GBUnited Kingdom3.7%4DEGermany2.4%5–6KRSouth Korea2.3%5–6AUAustralia2.3%7INIndia2.2%8–9HKHong Kong2.1%8–9ITItaly2.1%10CACanada1.7%

Share of the 10% most cited papers, 2021–2025; 3,671 papers. Source: OpenAlex, pulled 18 October 2026. Method

Year by year

About 734 papers a year. Single-year moves are noise.
010203040% 20212022202320242025 China 32.6%United States 15.4%United Kingdom 3.7%Australia 2.4%South Korea 2.4%Germany 2.3%

Share of the 10% most cited papers published in each year. Source: OpenAlex, pulled 18 October 2026. Method

Inside the field

Always the top 10%Pilot

Each paper sorted into one sub-area by an AI model reading its title and abstract, then credited as usual. On 200 hand-labelled papers it chose the same sub-area, or a defensible second, 92% of the time.

Sub-areaPapersFirstSecondThird
Domain applications (medicine, science, finance, industry)687USUnited States24.0%CNChina22.5%GBUnited Kingdom4.1%
Evaluation, benchmarks and hallucination473USUnited States22.5%CNChina16.8%GBUnited Kingdom3.8%
Code generation and software engineering381CNChina36.1%USUnited States18.2%SGSingapore3.8%
Retrieval-augmented generation and knowledge graphs339CNChina35.7%USUnited States14.0%GBUnited Kingdom3.7%
Fine-tuning, prompting and reasoning methods308CNChina33.0%USUnited States19.6%INIndia3.2%
ChatGPT overviews, education and societal impact298USUnited States18.1%CNChina9.9%INIndia5.7%
Multimodal and vision-language models283CNChina49.0%USUnited States10.0%GBUnited Kingdom4.2%
Model architecture, pre-training and scaling253CNChina26.9%USUnited States23.2%ITItaly2.5%
Security, privacy and safety (incl. LLMs for cybersecurity)191CNChina24.3%USUnited States19.1%AUAustralia4.1%
LLM agents and tool use179CNChina30.5%USUnited States21.0%GBUnited Kingdom5.3%

Share of each sub-area's papers in the top 10%. Source: OpenAlex, pulled 18 October 2026. Method

Earlier and recent

The same top 10%, split into 2021–2023 and 2024–2025.

FallingRising2021–20232024–2025Change
CNChina 15.3%29.1%+13.9 pts
USUnited States 29.1%18.0%−11.0 pts
GBUnited Kingdom 4.4%3.6%−0.8 pts
DEGermany 1.9%2.5%+0.5 pts
KRSouth Korea 1.2%2.5%+1.3 pts
AUAustralia 2.0%2.4%+0.3 pts
INIndia 2.8%2.1%−0.6 pts
HKHong Kong 2.3%2.1%−0.2 pts

Share of the 10% most cited papers: 452 papers in 2021–2023, 3,219 in 2024–2025. Source: OpenAlex, pulled 18 October 2026. Method

2026 so far: a different count

Not the most cited papers: every 2026 paper matched so far (37,747), and a paper counts in full for each country on it. This year's papers are too new to rank by citations, so these shares cannot be compared with the ones above.

  1. CNChina22.7%of 2026 papers have an author here
  2. USUnited States17.7%of 2026 papers have an author here
  3. GBUnited Kingdom4.0%of 2026 papers have an author here
  4. HKHong Kong3.3%of 2026 papers have an author here
  5. DEGermany2.9%of 2026 papers have an author here

Who cites the top papers

How much more often a country's top papers are cited from home than its size predicts. Expected is the country's share of all papers in the field; the bar is the excess, in percentage points. Home citers include authors citing their own earlier work.

HomeExpectedExcess
CNChina 42.8%20.8%+22.0
USUnited States 27.0%20.5%+6.5
GBUnited Kingdom 10.3%4.5%+5.8
DEGermany 11.0%3.0%+8.0
AUAustralia 11.8%2.1%+9.7

Based on up to 100 of each country's most cited single-country papers in the top 10%. Source: OpenAlex, pulled 18 October 2026. Method

Behind the share

Institutions with the most credit

  1. 1CNTsinghua University1.1%
  2. 2CNZhejiang University1.1%
  3. 3CNPeking University1.1%
  4. 4CNShanghai Jiao Tong University0.8%
  5. 5USMicrosoft (United States)Company0.7%
  6. 6USStanford University0.7%
  7. 7CNHarbin Institute of Technology0.7%
  8. 8CNWuhan University0.6%
  9. 9HKHong Kong Polytechnic University0.6%
  10. 10SGNanyang Technological University0.6%

Share of the top 10%. Institutions across all 45 technologies

Who does the work

SectorAllCNUSGB
Universities84.4%
83.0%
79.7%
91.4%
Companies6.0%
7.5%
11.0%
3.8%
Research facilities4.9%
5.9%
2.7%
2.1%
Hospitals1.6%
0.7%
3.2%
0.1%

Share of credit by type of institution, overall and for the three leading countries.

Across borders

32.5% of the top 10% have authors in more than one country.

  1. CNUSChina with United States187
  2. CNHKChina with Hong Kong134
  3. CNSGChina with Singapore110
  4. CNGBChina with United Kingdom85
  5. GBUSUnited Kingdom with United States81

Most frequent pairings, by number of shared papers.

The papers behind the numbers

The five most cited papers this definition catches. If one looks out of place, the definition needs work.

  1. A Survey on Evaluation of Large Language Models
    ACM Transactions on Intelligent Systems and Technology, 2024CNHKUS
    2,996citations
  2. 2,368citations
  3. GPT-4 Technical Report
    arXiv (Cornell University), 2023
    2,334citations
  4. PaLM: Scaling Language Modeling with Pathways
    arXiv (Cornell University), 2022US
    2,051citations
  5. 1,831citations

More in computing and ai

The most technologically advanced countries, and is China ahead of the US?

How this technology is defined

A paper counts if its title or abstract matches the search below. Articles and preprints from any source; retracted papers are excluded.

"large language model" OR "large language models" OR LLM OR ChatGPT OR "generative pre-trained transformer" OR "foundation model"

and its main topic in OpenAlex is one of: Topic Modeling; Natural Language Processing Techniques; Multimodal Machine Learning Applications; Software Engineering Research; Sentiment Analysis and Opinion Mining; Adversarial Robustness in Machine Learning; Advanced Neural Network Applications; Speech and dialogue systems; Software Testing and Debugging Techniques; Domain Adaptation and Few-Shot Learning; Explainable Artificial Intelligence (XAI); Recommender Systems and Techniques; Biomedical Text Mining and Ontologies; Artificial Intelligence Applications; Robot Manipulation and Learning. The topic check removes papers that only mention the technology.

Papers matched, 2021–2025
36,675
In the top 10%
3,671
In the top 1%
368
Countries with credit
96
Credit with no known country
12.8%
Every one of the 3,987 most cited papers was checked for relevance; 233 were left out as not about large language models

The most cited of those left out:

  • DINOv2: Learning Robust Visual Features without Supervision1,046 citations
  • LAION-5B: An open large-scale dataset for training next generation image-text models1,022 citations
  • A visual-language foundation model for computational pathology928 citations
  • CoCa: Contrastive Captioners are Image-Text Foundation Models503 citations
  • A Survey on Deep Neural Network Pruning: Taxonomy, Comparison, Analysis, and Recommendations493 citations
  • AI for life: Trends in artificial intelligence for biotechnology437 citations
  • Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model414 citations
  • A Comprehensive Review of Deep Learning: Architectures, Recent Advances, and Applications327 citations
  • Florence: A New Foundation Model for Computer Vision326 citations
  • Advancing Plain Vision Transformer Toward Remote Sensing Foundation Model285 citations
  • Deep learning based multimodal biomedical data fusion: An overview and comparative review278 citations
  • A Multimodal Biomedical Foundation Model Trained from Fifteen Million Image–Text Pairs268 citations
17 catalogue records left out: their titles or citation dates do not match the paper
  • BNAI, NO-TOKEN, and MIND-UNITY: Pillars of a Systemic Revolution in Artificial Intelligence4,201 citations. The arXiv number on this record, 2201.11903, belongs to a different paper.
  • LoRA Fine-Tuning of a 3B Code LLM for Algorithmic Efficiency2,488 citations. The arXiv number on this record, 2106.09685, belongs to a different paper.
  • A Direction-Aware Study of LLM-Based Code Translation Across Eight Programming Languages1,430 citations. The arXiv number on this record, 2107.03374, belongs to a different paper.
  • Can Open Large Language Models Catch Vulnerabilities?520 citations. The arXiv number on this record, 2501.12948, belongs to a different paper.
  • Code of "Sirens' Whisper: Inaudible Near-Ultrasonic Jailbreaks of Speech-Driven LLMs"195 citations. The arXiv number on this record, 2307.15043, belongs to a different paper.
  • A Note on Normalized Emergence Timing (in Pythia Language Model Evaluations)160 citations. The arXiv number on this record, 2304.01373, belongs to a different paper.
  • CareerX: A Retrieval-Augmented Generation Framework for Personalized AI-Driven Career Guidance99 citations. The arXiv number on this record, 2310.11511, belongs to a different paper.
  • Counterfactually Auditable Lifecycle Certification for Autonomous Agents70 citations. The arXiv number on this record, 2307.16789, belongs to a different paper.
  • GardenBench: A lightweight, daily evaluation of LLM capabilities69 citations. The arXiv number on this record, 2310.06770, belongs to a different paper.
  • Its Alive: AI Independence Without Human Prompting64 citations. The arXiv number on this record, 2310.01405, belongs to a different paper.
  • Guardrail Blind Spots on Turkish Prompt Injection: A Same-Set Miss-Rate Comparison of Four Guard Models55 citations. The arXiv number on this record, 2312.06674, belongs to a different paper.
  • CWEFT: CWE-Aware Evaluation of Free-Text vs. Typed Prompts51 citations. The arXiv number on this record, 2310.11324, belongs to a different paper.
  • Lexical Semantic Change through Large Language Models: a Survey48 citations. The arXiv number on this record, 2304.01666, belongs to a different paper.
  • Embedded prompt tuning: Towards enhanced calibration of pretrained models for medical images25 citations. The arXiv number on this record, 2407.01003, belongs to a different paper.
  • AI Benchmark Half-Life in Recursive Corpora: A Theory of Validity Decay under Semantic Leakage and Regeneration19 citations. The arXiv number on this record, 2406.19314, belongs to a different paper.
  • Reasoning Beyond Words ? Exploring framework for hidden state reasoning18 citations. The arXiv number on this record, 2412.06769, belongs to a different paper.
  • Automating quantum computing laboratory experiments with an agent-based AI framework18 citations. The arXiv number on this record, 2412.07978, belongs to a different paper.

Definition version 0.5, pilot. It is a plain text file: copy it, change it, rerun it.

Read the method