Dan Kitwood/Getty
Artificial intelligence (AI) remains one of the biggest citation magnets in scientific research, as highlighted by the latest Google Scholar metrics ranking.
Here we look at some of the most highly cited academic articles that feature in this ranking, which covers articles published in 2020 to 2024 and includes citations from all articles that were indexed in Google Scholar as of July 2025.
To create our list, we isolated papers with more than 1,000 citations that were published in 2024 by one of the top ten journals or conference proceedings in the 2025 Google Scholar Metrics ranking. From these, we ranked the six papers with the most citations.
These papers have made a rapid impact, attracting more citations in one year than many other older papers in their respective publications. Popular topics include AI that can read both images and text and systems that are providing new ways to understand the human body.
1. Accurate structure prediction of biomolecular interactions with AlphaFold 3
5,961 citations
Proteins are key parts of every living cell, and their shape determines their function, from building tissues to driving chemical reactions. Mapping these structures using technologies such as X-ray crystallography and nuclear magnetic imaging is often slow and costly. So, in 2018, when UK machine-learning company Google DeepMind debuted AlphaFold, an AI system that could predict the three-dimensional shapes of proteins based on their amino-acid sequences, there was widespread excitement within the research community.
Since then, DeepMind has released AlphaFold 2 — a more accurate system for generating the configuration of single proteins or simple protein complexes — and AlphaFold 3, its most advanced iteration to date. Described in this 2024 Nature paper, AlphaFold3 can capture the structure and interactions of large protein groups, including those bound to DNA, RNA and various ions.
Developing bottom-up models of cellular components is an important step in unravelling the complexity of molecular regulation within a cell, the DeepMind authors write in the paper.
This paper is among the top 10 most cited papers published by Nature between 2020 and 2024. The top Nature paper for this period is its predecessor: "Highly accurate protein structure prediction with AlphaFold", which has garnered more than 34,000 citations since its release in 2021.
2. MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Models
International Conference on Learning Representations
3,356 citations
This preprint describes the inner workings of MiniGPT-4, an open-source AI model that can process both text and images. It was developed by a team of PhD students at King Abdullah University of Science and Technology (KAUST) in Saudi Arabia, who authored the paper.
MiniGPT-4 is a type of large multimodal model (LMM), which can work with different types of information at once, such as text, images, audio and video. It was built by combining a vision model (for reading images) with a language model, and was designed to mimic some of the abilities of GPT-4, one of the systems that powers ChatGPT, developed by OpenAI, in San Francisco.
In this paper, the KAUST researchers describe how MiniGPT-4 can carry out surprisingly advanced tasks, even with limited training. The system can connect pictures and words in similar ways to GPT-4, they report, and in some cases, showed abilities that at the time had yet to be demonstrated in GPT-4. For example, MiniGPT-4 could generate a detailed recipe just by looking at a photo of a specific meal.
The paper is one of two from the International Conference on Learning Representations (ICLR) in this list.
3. Improved Baselines with Visual Instruction Tuning
Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition
2,998 citations
Researchers are racing to turn large language models (LLMs) into ever-better general-purpose assistants that can answer user questions about different types of images.
In this paper, researchers from the University of Wisconsin–Madison and Microsoft Research Lab – Redmond in Washington investigated ways to improve their LMM, called LLaVA-1.5. Changes included upgrading the system’s vision-language connector (the part that links visual processing with text processing) and increasing the scope of the training data to include academic-style questions and answers.
The team writes that new approaches to LMM design could pave the way for more capable open-source systems.
In the 2025 Google Scholar Metrics ranking, Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition is the second-ranked publication overall, after Nature.
4. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis
International Conference on Learning Representations
2,807 citations
The second ICLR paper in this list, this preprint is focused on SDXL (Stable Diffusion XL), a text-to-image AI model based on an existing tool called Stable Diffusion. This model is designed to transform written prompts, such as “a cosy village in springtime” or “a black cat in a kitchen”, into a realistic image.
Compared to previous versions of Stable Diffusion, SDXL uses a three-times-larger UNet-backbone — part of the system’s neural network that is designed to look at images and highlight specific areas of interest, such as tumours in medical scans.
SDXL is described as achieving a “remarkable level of realism” by authors, who have since released an open-source version of the model.
5. YOLOv10: real-time end-to-end object detection
Neural Information Processing Systems
2,455 citations
This paper explores the newest iteration in the YOLO (You Only Look Once) series, a model named for its ability to detect objects in real time through the single scan of an image array. The authors, from Tsinghua University in Beijing, describe YOLOv10 (the 10th iteration) as faster and more accurate than previous models.
A key change in this iteration is the elimination of a ‘clean-up’ process called non-maximum suppression (NMS). In previous models, the system would register an object by drawing multiple detection boxes around it, before using the NMS algorithm to select the box that most accurately labelled the object. YOLOv10 uses a system trained to predict a single detection box per object.
Optimising the YOLO model greatly reduces the computer processing power needed to run it, while also enhancing its capability, the authors say. They expect YOLOv10 will be used in various real-world scenarios such as medical-image analysis.
6. Segment anything in medical images
2,067 citations
From uncovering tumours to determining the efficacy of treatments, medical-image segmentation – the process of dividing an image into regions that correspond to organs, tissues and lesions – is a key part of evaluating a scan. But there is still no universal model that can perform this function across different diseases, body parts and imaging systems.
The paper presents MedSAM, a new segmentation model developed by researchers at the University of Toronto. Unlike conventional segmentation models, which have been tailored to specific diseases or imaging types, MedSAM is designed to work with a wide range of medical images and imaging technologies.
The authors describe how they tested the new model using data from more than 30 different cancer types captured by 10 different medical-imaging technologies, including magnetic resonance imaging and ultrasound. MedSAM achieved a performance “on par with, or even surpassing” other more specialized models, they report.
This paper is the 7th most cited article published by Nature Communications in 2020 to 2024, according to the 2025 Google Scholar Metrics ranking.
Since the paper’s release, the team has expanded on the model, releasing MedSAM2, a system for three-dimensional imagery, earlier this year.
