Academics who switch to hybrid working and remote collaboration do less impactful research. That’s according to an analysis of how scientists’ collaboration networks and academic outputs evolved before, during and after the COVID-19 pandemic (arXiv: 2511.18481). It involved studying author data from the arXiv preprint repository and the online bibliographic catalogue OpenAlex.
To explore the geographic spread of collaboration networks, Sara Venturini from the Massachusetts Institute of Technology and colleagues looked at the average distance between the institutions of co-authors. They found that while the average distance between team members on publications increased from 2000 to 2021, there was a particularly sharp rise after 2022.
This pattern, the researchers claim, suggests that the pandemic led to scientists collaborating more often with geographically distant colleagues. They found consistent patterns when they separated papers related to COVID-19 from those in unrelated areas, suggesting the trend was not solely driven by research on COVID-19.
The researchers also examined how the number of citations a paper received within a year of publication changed with distance between the co-authors’ institutions. In general, as the average distance between collaborators increases, citations fall, the authors found.
They suggest that remote and hybrid working hampers research quality by reducing spontaneous, serendipitous in-person interactions that can lead to deep discussions and idea exchange.
Despite what the authors say is a “concerning decline” in citation impact, there are, however, benefits to increasing remote interactions. In particular, as the geography of collaboration networks increases, so too does international partnerships and authorship diversity.
Remote tools
Lingfei Wu, a computational social scientist at the University of Pittsburgh, who was not involved in the study, told Physics World that he was surprised by the finding that remote teams produce less impactful work.
“In our earlier research, we found that historically, remote collaborations tended to produce more impactful but less innovative work,” notes Wu. “For example, the Human Genome Project published in 2001 shows how large, geographically distributed teams can also deliver highly impactful science. One would expect the pandemic-era shift toward remote collaboration to increase impact rather than diminish it.”
Wu says his work suggests that remote work is effective for implementing ideas but less effective for generating them, indicating that scientists need a balance between remote and in-person interactions. “Use remote tools for efficient execution, but reserve in-person time for discussion, brainstorming, and informal exchange,” he adds.
Special forces veteran Bryan Stern says he told US defence officials some of his planned route to reduce airstrike risk
The most dangerous moments came when salvation seemed finally assured.
Many miles from land, the small fishing skiff carrying the Venezuelan opposition leader and Nobel prize laureate María Corina Machado had been lost at sea for hours, tossed by strong winds and 10ft waves. A further hazard was the ever present risk of an inadvertent airstrike by US warplanes hunting alleged cocaine smugglers.
One problem with a lot of modern games is just how big they're getting, but HELLDIVERS 2 is about to massively shrink thanks to work behind the scenes.
Looking ahead to the future of machine learning: (clockwise from top left) Jay Lee, Jimeng Sun, Pierre Gentine and Kyle Cranmer.
IOP Publishing’s Machine Learning series is the world’s first open-access journal series dedicated to the application and development of machine learning (ML) and artificial intelligence (AI) for the sciences.
Part of the series is Machine Learning: Science and Technology, launched in 2019, which bridges the application and advances in machine learning across the sciences. Machine Learning: Earth is dedicated to the application of ML and AI across all areas of Earth, environmental and climate sciences while Machine Learning: Health covers healthcare, medical, biological, clinical and health sciences and Machine Learning: Engineeringfocuses on applied AI and non-traditional machine learning to the most complex engineering challenges.
Here, the editors-in-chief (EiC) of the four journals discuss the growing importance of machine learning and their plans for the future.
Kyle Cranmer is a particle physicist and data scientist at the University of Wisconsin-Madison and is EiC of Machine Learning: Science and Technology (MLST). Pierre Gentine is a geophysicist at Columbia University and is EiC of Machine Learning: Earth. Jimeng Sun is a biophysicist at the University of Illinois at Urbana-Champaign and is EiC of Machine Learning: Health. Mechanical engineer Jay Lee is from the University of Maryland and is EiC of Machine Learning: Engineering.
What do you attribute to the huge growth over the past decade in research into and using machine learning?
Kyle Cranmer (KC): It is due to a convergence of multiple factors. The initial success of deep learning was driven largely by benchmark datasets, advances in computing with graphics processing units, and some clever algorithmic tricks. Since then, we’ve seen a huge investment in powerful, easy-to-use tools that have dramatically lowered the barrier to entry and driven extraordinary progress.
Pierre Gentine (PG): Machine learning has been transforming many fields of physics, as it can accelerate physics simulation, better handle diverse sources of data (multimodality), help us better predict.
Jimeng Sun (JS): Over the past decade, we have seen machine learning models consistently reach — and in some cases surpass — human-level performance on real-world tasks. This is not just in benchmark datasets, but in areas that directly impact operational efficiency and accuracy, such as medical imaging interpretation, clinical documentation, and speech recognition. Once ML proved it could perform reliably at human levels, many domains recognized its potential to transform labour-intensive processes.
Jay Lee (JL): Traditionally, ML growth is based on the development of three elements: algorithms, big data, and computing. The past decade’s growth in ML research is due to the perfect storm of abundant data, powerful computing, open tools, commercial incentives, and groundbreaking discoveries—all occurring in a highly interconnected global ecosystem.
What areas of machine learning excite you the most and why?
KC: The advances in generative AI and self-supervised learning are very exciting. By generative AI, I don’t mean Large Language Models — though those are exciting too — but probabilistic ML models that can be useful in a huge number of scientific applications. The advances in self-supervised learning also allows us to engage our imagination of the potential uses of ML beyond well-understood supervised learning tasks.
PG: I am very interested in the use of ML for climate simulations and fluid dynamics simulations.
JS: The emergence of agentic systems in healthcare — AI systems that can reason, plan, and interact with humans to accomplish complex goals. A compelling example is in clinical trial workflow optimization. An agentic AI could help coordinate protocol development, automatically identify eligible patients, monitor recruitment progress, and even suggest adaptive changes to trial design based on interim data. This isn’t about replacing human judgment — it’s about creating intelligent collaborators that amplify expertise, improve efficiency, and ultimately accelerate the path from research to patient benefit.
JL: One area is generative and multimodal ML — integrating text, images, video, and more — are transforming human–AI interaction, robotics, and autonomous systems. Equally exciting is applying ML to nontraditional domains like semiconductor fabs, smart grids, and electric vehicles, where complex engineering systems demand new kinds of intelligence.
What vision do you have for your journal in the coming years?
KC: The need for a venue to propagate advances in AI/ML in the sciences is clear. The large AI conferences are under stress, and their review system is designed to be a filter not a mechanism to ensure quality, improve clarity and disseminate progress. The large AI conferences also aren’t very welcoming to user-inspired research, often casting that work as purely applied. Similarly, innovation in AI/ML often takes a back seat in physics journals, which slows the propagation of those ideas to other fields. My vision for MLST is to fill this gap and nurture the community that embraces AI/ML research inspired by the physical sciences.
PG: I hope we can demonstrate that machine learning is more than a nice tool but that it can play a fundamental role in physics and Earth sciences, especially when it comes to better simulating and understanding the world.
JS: I see Machine Learning: Health becoming the premier venue for rigorous ML–health research — a place where technical novelty and genuine clinical impact go hand in hand. We want to publish work that not only advances algorithms but also demonstrates clear value in improving health outcomes and healthcare delivery. Equally important, we aim to champion open and reproducible science. That means encouraging authors to share code, data, and benchmarks whenever possible, and setting high standards for transparency in methods and reporting. By doing so, we can accelerate the pace of discovery, foster trust in AI systems, and ensure that our field’s breakthroughs are accessible to — and verifiable by — the global community.
JL: Machine Learning: Engineering envisions becoming the global platform where ML meets engineering. By fostering collaboration, ensuring rigour and interpretability, and focusing on real-world impact, we aim to redefine how AI addresses humanity’s most complex engineering challenges.
Connexion
.