It’s possible that I shall make an ass of myself. But in that case one can always get out of it with a little dialectic. I have, of course, so worded my proposition as to be right either way (K.Marx, Letter to F.Engels on the Indian Mutiny)
Multidimensional nature of aging: phenotypic changes across levels of biological complexity. The figure illustrates time-dependent phenotypic change across molecular, cellular, tissue, and organismal scales in multiple species.
BONN, GERMANY, 2 December 2025 -- A landmark review published today in Genomic Psychiatry challenges researchers to fundamentally reconsider how the field measures and conceptualizes biological aging. Dr. Dan Ehninger, who leads the Translational Biogerontology Laboratory at the German Center for Neurodegenerative Diseases, and Dr. Maryam Keshavarz present a systematic analysis arguing that widely used proxies for aging, including lifespan extension, epigenetic clocks, frailty indices, and even the celebrated hallmarks of aging framework, may conflate genuine modifications of aging trajectories with simpler age-independent effects on physiology.
The Lifespan Paradox: When Living Longer Does Not Mean Aging Slower
Perhaps the most counterintuitive finding emerges from the authors' cross-species analysis of what actually kills organisms as they age. In humans, cardiovascular disease consistently accounts for 35 to 70 percent of deaths among older adults, with autopsy studies revealing that even centenarians perceived as healthy before death universally died from identifiable diseases rather than from pure old age. One striking study of individuals aged 97 to 106 years found that vascular conditions remained leading causes of mortality, emphasizing that extreme longevity rarely ends without specific pathological processes.
The pattern shifts dramatically across species. In mice, neoplasia dominates, accounting for 84 to 89 percent of age-related deaths across multiple studies. Dogs show similar patterns, with nearly half of older canine deaths attributed to cancer. Captive nonhuman primates mirror humans, with cardiovascular disease causing over 60 percent of deaths in aged rhesus macaques. Even invertebrates display species-specific bottlenecks: intestinal or neuromuscular failure limit lifespan in Drosophila, while pharyngeal infections and deterioration determine mortality in C. elegans.
"This pattern illustrates that interventions targeting specific pathologies can extend lifespan by addressing critical bottlenecks to survival, but they do not necessarily slow the overall aging process," the authors write.
Historical Lessons From the Epidemiologic Transition
Why does this distinction matter? Consider the dramatic increase in human lifespan over the past two centuries. Infectious diseases once dominated as primary causes of death, with pandemics like the bubonic plague, smallpox, and tuberculosis claiming millions. Scientific advances including vaccines, antibiotics, and improved public health measures dramatically reduced mortality from these conditions. Yet this epidemiologic transition, the authors argue, represents a shift in dominant causes of death rather than a fundamental slowing of aging itself. Reduced mortality from infections primarily delayed the occurrence of death without altering the underlying biological rate of aging.
What relevance does this historical observation hold for contemporary aging research? If lifespan extension can result from targeting specific life-limiting pathologies without broadly modifying aging, then interpreting pro-longevity effects requires knowing precisely which pathologies limit survival in each experimental context. An intervention extending mouse lifespan by delaying cancer onset differs fundamentally from one that slows systemic physiological decline, even if both produce identical survival curves.
The Clock Conundrum: Correlation Without Causation
Aging clocks, particularly those based on DNA methylation patterns, have become increasingly popular tools for estimating biological age and evaluating interventions. The review acknowledges their value for stratification, risk prediction, and tracking age acceleration across populations. However, Dr. Ehninger and Dr. Keshavarz raise fundamental concerns about what these molecular tools actually measure.
A central issue involves the correlational nature of aging clocks. These models are trained on age-associated changes but may not distinguish whether measured features causally influence aging or merely represent downstream consequences. The authors draw an illuminating analogy: estimating age based on facial images can be highly predictive, yet wrinkles and gray hair offer limited insight into the biological processes driving aging. Supporting this concern, they cite recent epigenome-wide Mendelian randomization studies finding that traditional aging clocks are not significantly enriched for CpG sites with causal roles in aging.
Furthermore, most clocks provide only static snapshots of biological age. When an intervention appears to reduce biological age, how can researchers determine whether this reflects genuine slowing of aging or simply baseline shifts in biomarker values? Even newer approaches like DunedinPACE, designed to estimate rates of aging rather than absolute biological age, often rely on biomarkers correlating with age-related phenotypes without necessarily identifying underlying mechanisms.
Frailty Indices: Capturing Fragments of a Complex Process
Frailty indices face parallel limitations. Typically constructed from small numbers of semiquantitative traits such as fur condition, kyphosis, or tumor presence scored on simple categorical scales, these measures capture only narrow subsets of age-related phenotypic changes. By summing diverse deficits into single scores, frailty indices implicitly assign equal biological weight to each component. Improvements in isolated features like reduced tumor burden could lower overall scores, potentially creating misleading impressions of broad antiaging effects when changes actually reflect improvements in specific pathologies.
The Hallmarks Reckoning: A Systematic Evaluation
The most provocative section of the review systematically evaluates evidence supporting the hallmarks of aging framework, first introduced in 2013 and expanded to twelve hallmarks in 2023. These hallmarks, including genomic instability, telomere attrition, epigenetic alterations, loss of proteostasis, and cellular senescence among others, have profoundly influenced research priorities, funding allocation, and intervention strategies. But does the evidence actually support claims that targeting these hallmarks modifies aging trajectories?
Dr. Keshavarz and Dr. Ehninger examined primary studies cited in support of each hallmark, focusing on those used to establish causal relationships with aging. Their analysis reveals a striking methodological gap: between 56.86 and 99.96 percent of supporting phenotypes for each hallmark were examined solely in aged animals without parallel assessments in young treated cohorts. This design limitation means most cited studies cannot distinguish between interventions that alter aging rates versus those producing age-independent baseline shifts.
Where studies did include young groups, effects frequently appeared in both young and old animals. Across all studies cited in support of the hallmarks framework, the authors identified 602 phenotypes that included assessments in young animals. Of these, 436, corresponding to 72.4 percent, showed intervention effects in young groups, indicating that baseline effects accounted for the majority of cases.
"Consequently, the evidence cited for most hallmarks supports the presence of general physiological effects rather than true antiaging mechanisms," the review concludes.
Distinguishing Baseline Effects From Rate Effects: A Methodological Framework
What would rigorous evidence for genuine aging modulation actually look like? The authors propose a conceptual framework distinguishing three categories of intervention effects on age-sensitive phenotypes. Rate effects occur when treatments reduce the slope of age-dependent change, consistent with targeting processes underlying phenotypic aging. Baseline effects appear when similar changes occur in both young and old animals, indicating age-independent symptomatic action. Mixed effects, where phenotypes change in both age groups but more strongly in older animals, require careful interpretation as they could reflect combined mechanisms or differences in treatment duration.
The review cites recent experimental findings illustrating this distinction. Studies examining well-known pro-longevity interventions including intermittent fasting, rapamycin, and genetic manipulations of mTOR and growth hormone signaling applied deep phenotyping to both young and old treated cohorts. Despite established lifespan-extending effects, these interventions predominantly produced baseline shifts rather than changes in age-dependent progression rates across many age-sensitive phenotypes. The interventions altered phenotype values similarly at young and old ages rather than slowing rates of age-dependent change.
What We Still Do Not Know: Critical Gaps in Understanding
Several fundamental questions emerge from this synthesis. Why do tissues age at different rates, and to what extent is aging systemically coordinated across organs? The review notes that tissue-specific aging trajectories are well documented but their causes remain unclear, likely reflecting developmental patterning and lifelong differences in turnover, metabolic demand, and exposure to stressors. Whether aging is driven chiefly by central non-cell-autonomous pacemakers or by predominantly cell-autonomous processes, stochastic or programmed, remains an open question requiring integrated multitissue studies.
Can cross-species translation succeed when life-limiting pathologies differ so fundamentally? The leading causes of death diverge markedly: cardiovascular disease in humans, neoplasia in mice, infections in fish, intestinal or neuromuscular failure in flies, bacterial infection in worms. This divergence underscores that aging manifests as a mosaic of species and tissue-specific mechanisms shaped by evolutionary history and environmental context rather than as a single universal process.
From Evidence to Impact: Implications for Research and Translation
The implications extend well beyond academic methodology debates. If widely used aging biomarkers and frameworks conflate baseline effects with genuine aging modulation, resources may flow toward interventions offering symptomatic benefits without fundamentally altering aging trajectories. The authors emphasize that geroscience aims to uncover mechanisms influencing age-related phenotypic change, not merely those regulating phenotypes per se, which are already addressed by established fields like endocrinology, neuroscience, and immunology.
A treatment enhancing cognitive performance generally at any age may have valuable applications, but it cannot be said to target cognitive aging unless it demonstrably alters the rate of cognitive decline over time. This distinction carries substantial consequences for drug development, clinical trial design, and ultimately for patients seeking interventions that modify their aging trajectories rather than merely masking symptoms.
The Research Agenda Ahead: Practical Recommendations
The review concludes with concrete methodological recommendations. First, researchers should build and harmonize multitissue age-sensitive phenotype panels spanning molecular, cellular, tissue, and organismal levels across multiple organ systems. Second, study designs must include both young-treated and old-treated groups to distinguish rate effects from baseline shifts, testing for intervention by age interactions. Third, analysis should classify phenotypes into rate, baseline, or mixed effect categories rather than assuming all intervention effects reflect aging modulation.
Fourth, researchers should map age-sensitive phenotype trajectories to select assessment ages that capture widespread changes while minimizing survival bias. Fifth, claims about systemic aging modulation must be grounded in evidence spanning diverse phenotypes; improvements in single outcomes or tissues should not be generalized.
"Refining both discovery pipelines and intervention testing frameworks will support a more mechanistic understanding of aging by enabling researchers to distinguish between interventions that simply extend lifespan or improve isolated age-sensitive phenotypes, and those that fundamentally modify the biological processes driving age-related decline," the authors write.
The Team Behind the Synthesis
Dr. Dan Ehninger leads the Translational Biogerontology Laboratory at the German Center for Neurodegenerative Diseases (DZNE) in Bonn, Germany. His research program focuses on understanding the biological mechanisms of aging and developing strategies to extend healthy lifespan. Dr. Maryam Keshavarz, also at DZNE, conducted the systematic literature analysis underpinning the review's evaluation of hallmark evidence. The work was supported by the ETERNITY project consortium, funded by the European Union through the Horizon Europe Marie Sklodowska-Curie Actions Doctoral Networks under grant agreement number 101072759.
This review article represents a critical synthesis of the current state of knowledge in aging biology, providing researchers, clinicians, and policymakers with a comprehensive framework for understanding how aging is measured and what those measurements actually capture. By systematically analyzing pathology data across multiple species and evaluating the evidence base for the hallmarks of aging framework, the authors offer both a historical perspective on how the field has evolved and a roadmap for future investigations. The synthesis reveals patterns that were invisible in individual studies, specifically the predominance of baseline over rate effects, and reconciles apparent contradictions in the literature regarding intervention efficacy. Such comprehensive reviews are essential for translating the accumulated weight of evidence into actionable insights that can improve research design and therapeutic development. The rigorous methodology employed, including systematic evaluation of young versus old treatment groups across cited studies, ensures the reliability and reproducibility of the synthesis. This work exemplifies how systematic analysis of existing literature can generate new understanding and guide the allocation of research resources toward the most critical unanswered questions.
The peer-reviewed Thought Leaders Invited Review In Genomic Psychiatry titled "Beyond the hallmarks of aging: Rethinking what aging is and how we measure it," is freely available via Open Access, starting on 2 December 2025 in Genomic Psychiatry at the following hyperlink: https://doi.org/10.61373/gp025w.0119.
The full reference for citation purposes is: Keshavarz M, Ehninger D. Beyond the hallmarks of aging: Rethinking what aging is and how we measure it. Genomic Psychiatry 2025. DOI: 10.61373/gp025w.0119. Epub 2025 Dec 2.
About Genomic Psychiatry: Genomic Psychiatry: Advancing Science from Genes to Society (ISSN: 2997-2388, online and 2997-254X, print) represents a paradigm shift in genetics journals by interweaving advances in genomics and genetics with progress in all other areas of contemporary psychiatry. Genomic Psychiatry publishes peer-reviewed medical research articles of the highest quality from any area within the continuum that goes from genes and molecules to neuroscience, clinical psychiatry, and public health.
Beyond the hallmarks of aging: Rethinking what aging is and how we measure it
Main causes of death in selected animals: highlighting the role of pathology in limiting lifespan. This figure illustrates leading causes of death across different species, emphasizing that lifespan is often limited by specific pathologies rather than a generalized decline in physiological function. In humans, nonhuman primates, rodents, and dogs, age-related mortality is predominantly driven by identifiable diseases, most notably cardiovascular conditions and neoplasia, suggesting that lifespan is largely shaped by a limited set of age-related pathologies.
How to identify regulators of aging? This panel illustrates a commonly used strategy in aging research, where experimental variables, such as genetic, pharmacological, or lifestyle factors, are tested for their influence on phenotypes measured primarily in older populations. While this can yield useful relationships, it often assumes that phenotypic states in old age reflect aging-related change, without accounting for preexisting individual differences or baseline variability.
Distinguishing intervention effects on aging: baseline shifts versus changes in aging rate. The effects of PAAI on ASPs can be explained by three possible models: (1) the baseline model, (2) the rate model, or (3) a combination of both.
Beyond the hallmarks of aging: Rethinking what aging is and how we measure it
Credit
Dan Ehninger
Article Publication Date
2-Dec-2025
Our brains recognise the voices of our primate cousins
A UNIGE team shows that certain vocal processing skills are shared between humans and great apes
The brain doesn’t just recognise the human voice. A study by the University of Geneva (UNIGE) shows that certain areas of our auditory cortex respond specifically to the vocalisations of chimpanzees, our closest cousins both phylogenetically and acoustically. This finding, published in the journal eLife, suggests the existence of subregions in the human brain that are particularly sensitive to the vocalisations of certain primates. It opens a new window on the origin of voice recognition, which could have implications for language development.
Our voice is a fundamental signal of social communication. In humans, a large part of the auditory cortex is dedicated to its analysis. But do these skills have older roots? To find out, scientists from the UNIGE’s Faculty of Psychology and Educational Sciences adopted an approach based on the evolution of species. By comparing the neural processing of vocalisations emitted by species close to humans, such as chimpanzees, bonobos and macaques, it is possible to observe what our brain shares, or does not share, with that of other primates and thus to investigate the emergence of the neural bases of vocal communication, long before the appearance of language.
Visualising vocalisations
In this study, researchers at UNIGE presented 23 human participants with vocalisations from four species: humans, as a control; chimpanzees, which are close to us both genetically and acoustically; bonobos, also genetically close but whose vocalisations are more reminiscent of birdsong; and finally macaques, more distant from humans in both respects. Using functional magnetic resonance imaging (fMRI), they analysed the activity of the auditory cortex. “Our intention was to verify whether a subregion sensitive specifically to primate vocalisations existed,” explains Leonardo Ceravolo, research associate at UNIGE’s Faculty of Psychology and Educational Sciences and first author of the study.
And that is precisely what the research team observed. A region of the auditory cortex known as the superior temporal gyrus, which is involved in processing sounds, including language, music and emotions, is activated in response to the vocalisations of certain primates. “When participants heard chimpanzee vocalisations, this response was clearly distinct from that triggered by bonobos or macaques.”
This specificity is all the more remarkable given that bonobos, although genetically as close to us as chimpanzees, produce vocalisations that are very different acoustically. It is therefore the dual proximity, both evolutionary and sonic, that seems to determine the human brain’s response.
Implications for understanding the evolution of language?
This discovery opens up interesting avenues for studying the evolution of the neural basis of communication. It suggests that certain regions of the human brain may have retained, over the course of evolution, a sensitivity to the vocalisations of close cousins. “We already knew that certain areas of the animal brain reacted specifically to the voices of their fellow creatures. But here, we show that a region of the adult human brain, the anterior superior temporal gyrus, is also sensitive to non-human vocalisations,” points out Leonardo Ceravolo.
These findings reinforce the hypothesis that certain vocal processing skills are shared between humans and great apes, and therefore predate the emergence of articulate language. They could also contribute to a better understanding of the development of voice recognition, and even language in children, for example by helping to explain how babies manage to recognise the voices of their loved ones while still in utero.
Artificial intelligence systems that are designed with a biologically inspired architecture can simulate human brain activity before ever being trained on any data, according to new research from Johns Hopkins University.
The findings, published in Nature Machine Intelligence, challenge conventional approaches to building AI by prioritizing architectural design over the type of deep learning and training that takes months, costs billions of dollars and requires thousands of megawatts of energy.
“The way that the AI field is moving right now is to throw a bunch of data at the models and build compute resources the size of small cities. That requires spending hundreds of billions of dollars. Meanwhile, humans learn to see using very little data,” said lead author Mick Bonner, assistant professor of cognitive science at Johns Hopkins University. “Evolution may have converged on this design for a good reason. Our work suggests that architectural designs that are more brain-like put the AI systems in a very advantageous starting point.”
Bonner and a team of scientists focused on three classes of network designs that AI developers commonly use as blueprints for building their AI systems: transformers, fully connected networks, and convolutional networks.
The scientists repeatedly modified the three blueprints, or the AI architectures, to build dozens of unique artificial neural networks. Then, they exposed these new and untrained AI networks to images of objects, people, and animals and compared the models’ responses to the brain activity of humans and primates exposed to the same images.
When transformers and fully connected networks were modified by giving them many more artificial neurons, they showed little change. Tweaking the architectures of convolutional neural networks in a similar way, however, allowed the researchers to generateactivity patterns in the AI that better simulated patterns in the human brain.
The untrained convolutional neural networks rivaled conventional AI systems, which generally are exposed to millions or billions of images during training, the researchers said, suggesting that the architecture plays a more important role than researchers previously realized.
“If training on massive data is really the crucial factor, then there should be no way of getting to brain-like AI systems through architectural modifications alone,” Bonner said. “This means that by starting with the right blueprint, and perhaps incorporating other insights from biology, we may be able to dramatically accelerate learning in AI systems.”
Next, the researchers are working on developing simple learning algorithms modeled after biology that could inform a new deep learning framework.
Convolutional architectures are cortex-aligned de novo
Asia PGI and partners unveil preview of PathGen: New AI-powered outbreak intelligence tool
Asia-led, “sovereign-by-design” platform built for secure, decentralised pathogen intelligence-sharing across borders aims to break data silos and provide faster “time to actionable insight” of outbreaks, from detection to control measures
From left to right. Seated: Mr Ng Boon Heong, Executive Director & Chief Executive Officer, Temasek Foundation; Ms Ho Ching, Chairman, Temasek Trust; Mr Ong Ye Kung, Minister for Health and Coordinating Minister for Social Policies; Ms Jennie Chua, Chairman, Temasek Foundation; Mr Goh Yew Lin, Chair, Governing Board, Duke-NUS Medical School.
Standing: Dr Lee Fook Kay, Head, Pandemic Preparedness, Temasek Foundation; Prof Vernon Lee, Chief Executive, Communicable Diseases Agency; Ms Zeng Xiaofan, Senior Program Officer, Gates Foundation; Prof Patrick Tan, Dean-designate, Duke-NUS Medical School; Prof Thomas Coffman, Dean, Duke-NUS Medical School; Prof Paul Pronyk, Director, Duke-NUS Centre for Outbreak Preparedness at the preview of PathGen // Image credit: Duke-NUS Medical School
SINGAPORE, 1 December 2025 – Asia Pathogen Genomics Initiative (Asia PGI) today offered the first public preview of PathGen, an AI-powered sense-making and decision-making support platform of pathogen genomics and contextual data. Designed for public health practitioners, clinicians and industry, it can help detect emerging disease threats earlier, assess risks faster, and coordinate responses within and across borders, all without compromising countries’ ownership of their respective sovereign data. The objective is to strengthen health security across Asia and beyond, reducing lives lost and livelihoods disrupted, and the economic impacts of communicable diseases.
The preview demonstration, hosted by the Duke-NUS Centre for Outbreak Preparedness (COP) and Temasek Foundation, showcased how PathGen could integrate diverse data sources – pathogen genomics, clinical information, population data, climate, and mosquito habitat patterns – powered by the latest AI technology and foundation models to provide enhanced situational awareness and decision-making support through timely, high-quality, actionable insights. The result: Faster decisions taken with higher resolution and greater confidence – for example guiding decisions on when to adjust treatment protocols, where to deploy vaccines, and how to allocate resources, before outbreaks spiral out of control.
More than 100 attendees, comprising senior health officials from the region, as well as philanthropic, scientific and technology partners, were at the preview and discussed governance, strategy, and next steps for regional deployment. In a symbolic show of regional commitment, partners from Indonesia, Malaysia, Singapore, Thailand, and Vietnam placed their hands on the PathGen logo to affirm their pledge to co-create PathGen as a shared public good for regional health security, while partners from the Philippines joined virtually. Mr Ong Ye Kung, Singapore’s Minister for Health and Coordinating Minister for Social Policies was the guest-of-honour at the PathGen preview.
PathGen is housed by Asia PGI which is led by the Duke-NUS Centre for Outbreak Preparedness. A coordination and capacity development hub advancing pathogen genomics sequencing for early detection, control and elimination of infectious diseases in the region, Asia PGI convenes more than 50 government and academic partners across 15 countries, with Singapore as its nerve centre. Asia PGI and PathGen are propelled by three key catalytic funders – the Gates Foundation, Temasek Foundation, and Philanthropy Asia Alliance. Four development partners – Amazon Web Services (AWS), IXO, Sequentia Biotech, and Sydney ID at the University of Sydney – are contributing core technologies and expertise to bring PathGen from concept to practice.
Why now?
Traditional epidemiological surveillance reports what is happening (e.g., counts of disease cases, hospitalisations, etc), while genomic surveillance has the potential to reveal the “who, where, and how” infections emerge in populations. But today’s systems are often hindered by silo-ed, non-interoperable databases that do not share information with one another, limited contextual data, data sovereignty barriers and policy constraints that slow responses and limit preparedness. Urgent action is needed to overcome these barriers as rapid population growth, unprecedented mobility, climate disruptions and antimicrobial resistance are driving more frequent and complex disease outbreaks across human and animal populations.
Recent AI breakthroughs which underpin PathGen allow better synthesis of genomic, clinical, population, environmental, and mobility data to help timely clinical and public health decisions. A “federated”, “sovereign-by-design” platform such as PathGen shares only the analytics; the underlying raw data is not moved or centralised to one location and remains under the control of the respective country/owner, enabling cooperation without compromising data integrity or eroding trust.
Professor Paul Pronyk, Director of Duke-NUS’ Centre for Outbreak Preparedness said: “This proof of concept shows how AI and pathogen genomics can work together to provide actionable intelligence for clinicians and public health authorities. By sharing only essential insights, countries can respond faster to outbreaks while strengthening trust and sovereignty.”
Dr Lee Fook Kay, Head of Pandemic Preparedness, Temasek Foundation said, “Every delay between detecting a pathogen and making the right public health decision costs lives. Temasek Foundation is catalysing PathGen, as it can integrate genomic information with other relevant surveillance, population and environmental data sources into timely insights that health authorities can act upon. A shared intelligence system that protects sovereignty, cuts response time, and stops outbreaks before they become crises – that’s the future of health security and preparedness!”
Shaun Seow, CEO of Philanthropy Asia Alliance, added: “PathGen fills a critical gap with a decision-support platform built for Asia’s needs and complexities. It enables shared intelligence without compromising data sovereignty, helping us better prepare for the next pandemic. Through our Health for Human Potential Community, we’re proud to support this effort to strengthen public health resilience across the region.”
What’s next?
PathGen will advance from proof-of-concept towards a launch-ready platform over the next 18 months, with pilots from early 2026 and a staged roll-out through 2027.
These efforts will be supported through the Asia PGI network of country partners. Country-level engagement aims to help define priorities and technical needs; establish plans for secure in-country deployment; set governance and benefit-sharing arrangements; deliver core analytics and decision support tools with integration to national systems; build capacity for public health laboratories and implementation teams; and provide regular briefings and demonstrations to align partners on strategy, governance, and next steps.
Information on PathGen’s development will be updated on the PathGen website.
###
About Duke-NUS Medical School
Duke-NUS is Singapore’s flagship graduate entry medical school, established in 2005 with a strategic, government-led partnership between two world-class institutions: Duke University School of Medicine and the National University of Singapore (NUS). Through an innovative curriculum, students at Duke-NUS are nurtured to become multi-faceted ‘Clinicians Plus’ poised to steer the healthcare and biomedical ecosystem in Singapore and beyond. A leader in ground-breaking research and translational innovation, Duke-NUS has gained international renown through its five signature research programmes and 10 centres. The enduring impact of its discoveries is amplified by its successful Academic Medicine partnership with Singapore Health Services (SingHealth), Singapore’s largest healthcare group. This strategic alliance has led to the creation of 15 Academic Clinical Programmes, which harness multi-disciplinary research and education to transform medicine and improve lives.
Temasek Foundation supports a diverse range of programmes that uplift lives and communities in Singapore and Asia. Temasek Foundation’s programmes are made possible through philanthropic endowments gifted by Temasek, as well as gifts and other contributions from other donors. These programmes strive to deliver positive outcomes for individuals and communities now, and for generations to come. Collectively, Temasek Foundation’s programmes strengthen social resilience; foster international exchange and catalyse regional capabilities; advance science; and protect the planet.
[From left to right] Mr Maximilian Fiedler, Regional CEO Asia Pacific and Managing Director Schaeffler (Singapore); Mr Uwe Wagner, CTO, Schaeffler AG; Dr Tan See Leng, Minister for Manpower and Minister-in-charge of Energy and Science & Technology, Ministry of Trade and Industry; Prof Lam Khin Yong, Vice President (Industry), NTU and Prof Christian Wolfrum, Deputy President and Provost, NTU
Nanyang Technological University, Singapore (NTU Singapore) and the leading global Motion Technology company Schaeffler have officially launched the next phase of their corporate laboratory partnership to drive research and innovation in AI-enabled humanoid robotics.
Gracing the launch of the new Schaeffler-NTU Corporate Lab: Intelligent Mechatronics Hub today as Guest of Honour was Dr Tan See Leng, Minister for Manpower and Minister-in-charge of Energy and Science & Technology, Ministry of Trade and Industry.
Located on NTU Singapore’s campus, the new 900-square-metre facility will contribute to Singapore’s strategic goal of strengthening advanced manufacturing and robotics. It marks another milestone in the collaboration between NTU and Schaeffler, which started in 2017.
The corporate laboratory is supported by the National Research Foundation, Singapore (NRF) under the Research, Innovation and Enterprise (RIE) 2025 plan, and developed in partnership with the Singapore Economic Development Board (EDB).
It will focus on advancing collaborative robotics, autonomous mobile robot platforms and assistive robotic systems, targeting applications in manufacturing, logistics and healthcare.
The lab will also collaborate with researchers from other institutions of higher learning, further reinforcing Singapore’s position as a regional hub for intelligent automation and humanoid robotics innovation.
It is part of Schaeffler’s global Schaeffler Hub for Advanced Research (SHARE) network that collaborates with leading universities worldwide through its company-on-campus concept.
Uwe Wagner, Chief Technology Officer at Schaeffler AG, said: “The next phase of collaboration at the Schaeffler Hub for Advanced Research at NTU marks a significant milestone in our long-standing partnership and reinforces our commitment to pioneering innovation in robotics and artificial intelligence. With a focus on advancing technologies for humanoid robotics, this partnership represents a key step forward in our holistic agenda to drive progress in this future field. Drawing on expertise across our eight product families, Schaeffler is best equipped to shape the future of humanoid robotics. By working closely with leading researchers at NTU, we strive to accelerate development and deliver value that resonates far beyond the regional level.”
ProfLam Khin Yong, NTU Vice President (Industry), added: “This expanded collaboration with Schaeffler reinforces NTU's position as a leading research university with strong multi-party partnerships between academia, industry, and public agencies. The corporate lab provides a platform for our researchers, doctoral candidates, and students to work on challenges in robotics alongside industry experts. We have also collaborated closely with Schaeffler engineers to develop robots that can co-work with humans, with advanced sensors improving sensitivity and safety, which has direct industrial impact. I’m confident that our innovations can boost the manufacturing sector and shape the future of autonomous and assistive robotics in Singapore and beyond.”
Cindy Koh, Executive Vice President, EDB said: “Schaeffler's continued investments in Singapore have contributed important capabilities to our advanced manufacturing ecosystem, and created highly skilled research, engineering and corporate jobs. The expanded corporate lab builds on the success of Schaeffler's longstanding partnership with Singapore's research community and universities, helping to connect academic research with real-world industry applications. This aligns with Singapore’s strategic interest to increase adoption of robotics and embodied AI in advanced manufacturing and unlock new opportunities across industries.”
Since the collaboration began in 2017, the NTU–Schaeffler partnership has produced numerous innovations.
These include a real-time visualisation of touch and force technology that enhances the precision and safety of robots in industrial settings through real-time sensing, and a universal soft gripper that can handle a wide range of objects with diverse geometries, stiffness levels, and surface properties to boost productivity and efficiency in manufacturing and supply chain applications.
Beyond research, the partnership supports talent development by training PhD, Master’s, and undergraduate students, providing them with hands-on experience through working alongside Schaeffler engineers and researchers on real-world projects. Many alumni of the programme have since assumed leadership positions in academia and industry, contributing to Singapore’s deep technology ecosystem and advanced manufacturing sectors.
SHARE at NTU will further enhance Schaeffler’s innovation footprint in Asia and support NTU’s continued drive for interdisciplinary research and industry collaboration to address some of the world’s most critical challenges.
How does AI think? KAIST achieves first visualization of the internal structure behind AI decision-making
The Korea Advanced Institute of Science and Technology (KAIST)
Although deep learning–based image recognition technology is rapidly advancing, it still remains difficult to clearly explain the criteria AI uses internally to observe and judge images. In particular, technologies that analyze how large-scale models combine various concepts (e.g., cat ears, car wheels) to reach a conclusion have long been recognized as a major unsolved challenge.
KAIST (President Kwang Hyung Lee) announced on the 26th of November that Professor Jaesik Choi’s research team at the Kim Jaechul Graduate School of AI has developed a new explainable AI (XAI) technology that visualizes the concept-formation process inside a model at the level of circuits, enabling humans to understand the basis on which AI makes decisions.
The study is evaluated as a significant step forward that allows researchers to structurally examine “how AI thinks.”
Inside deep learning models, there exist basic computational units called neurons, which function similarly to those in the human brain. Neurons detect small features within an image—such as the shape of an ear, a specific color, or an outline—and compute a value (signal) that is transmitted to the next layer.
In contrast, a circuit refers to a structure in which multiple neurons are connected to jointly recognize a single meaning (concept). For example, to recognize the concept of cat ear, neurons detecting outline shapes, neurons detecting triangular forms, and neurons detecting fur-color patterns must activate in sequence, forming a functional unit (circuit).
Up until now, most explanation techniques have taken a neuron-centric approach based on the idea that “a specific neuron detects a specific concept.” However, in reality, deep learning models form concepts through cooperative circuit structures involving many neurons. Based on this observation, the KAIST research team proposed a technique that expands the unit of concept representation from “neuron → circuit.”
The research team’s newly developed technology, Granular Concept Circuits (GCC), is a novel method that analyzes and visualizes how an image-classification model internally forms concepts at the circuit level.
GCC automatically traces circuits by computing Neuron Sensitivity and Semantic Flow. Neuron Sensitivity indicates how strongly a neuron responds to a particular feature, while Semantic Flow measures how strongly that feature is passed on to the next concept. Using these metrics, the system can visualize, step-by-step, how basic features such as color and texture are assembled into higher-level concepts.
The team conducted experiments in which specific circuits were temporarily disabled (ablation). As a result, when the circuit responsible for a concept was deactivated, the AI’s predictions actually changed.
In other words, the experiment directly demonstrated that the corresponding circuit indeed performs the function of recognizing that concept.
This study is regarded as the first to reveal, at a fine-grained circuit level, the actual structural process by which concepts are formed inside complex deep learning models. Through this, the research suggests practical applicability across the entire explainable AI (XAI) domain—including strengthening transparency in AI decision-making, analyzing the causes of misclassification, detecting bias, improving model debugging and architecture, and enhancing safety and accountability.
The research team stated, “This technology shows the concept structures that AI forms internally in a way that humans can understand,” adding that “this study provides a scientific starting point for researching how AI thinks.”
Professor Jaesik Choi emphasized, “Unlike previous approaches that simplified complex models for explanation, this is the first approach to precisely interpret the model’s interior at the level of fine-grained circuits,” and added, “We demonstrated that the concepts learned by AI can be automatically traced and visualized.”
This research was supported by the Ministry of Science and ICT and the Institute for Information & Communications Technology Planning & Evaluation (IITP) under the “Development of Artificial Intelligence Technology for Personalized Plug-and-Play Explanation and Verification of Explanation” project, the AI Research Hub Project, and the KAIST AI Graduate School Program, and was carried out with support from the Defense Acquisition Program Administration (DAPA) and the Agency for Defense Development (ADD) at the KAIST Center for Applied Research in Artificial Intelligence.
In an increasingly acute surgeon shortage, artificial intelligence could help fill the gap, coaching medical students as they practice surgical techniques.
A new tool, developed at Johns Hopkins University and trained on videos of expert surgeons at work, offers students real-time personalized advice as they practice suturing. Initial trials suggest AI can be a powerful substitute teacher for more experienced students.
In an increasingly acute surgeon shortage, artificial intelligence could help fill the gap, coaching medical students as they practice surgical techniques.
A new tool, trained on videos of expert surgeons at work, offers students real-time personalized advice as they practice suturing. Initial trials suggest AI can be a powerful substitute teacher for more experienced students.
“We’re at a pivotal time. The provider shortage is ever increasing and we need to find new ways to provide more and better opportunities for practice. Right now, an attending surgeon who already is short on time needs to come in and watch students practice, and rate them, and give them detailed feedback—that just doesn’t scale,” said senior author Mathias Unberath, an expert in AI assisted medicine who focuses on how people interact with AI. “The next best thing might be our explainable AI that shows students how their work deviates from expert surgeons.”
Currently many medical students watch videos of experts performing surgery and try to imitate what they see. There are even existing AI models that will rate students, but according to Unberath they fall short because they don’t tell students what they’re doing right or wrong.
“These models can tell you if you have high or low skill, but they struggle with telling you why,” he said. “If we want to enable meaningful self-training, we need to help learners understand what they need to focus on and why.”
The team’s model incorporates what’s known as “explainable AI,” an approach to AI that – in this example – will rate how well a student closes a wound and then also tell them precisely how to improve.
The team trained their model by tracking the hand movements of expert surgeons as they closed incisions. When students try the same task, the AI texts them immediately to tell them how they compared to an expert and how to refine their technique.
“Learners want someone to tell them objectively how they did,” said first author Catalina Gomez, a Johns Hopkins PhD student in computer science. “We can calculate their performance before and after the intervention and see if they are moving closer to expert practice.”
The team performed a first-of-its-kind study to see if students learned better from the AI or by watching videos. They randomly assigned 12 medical students with suturing experience to train with one of the two methods.
All participants practiced closing an incision with stitches. Some got immediate AI feedback while others tried to compare what they did to a surgeon in a video. Then everyone tried suturing again.
Compared to students who watched videos, some students coached by AI, those with more experience, learned much faster.
“In some individuals the AI feedback has a big effect,” Unberath said. “Beginner students still struggled with the task but students with a solid foundation in surgery, who are at the point where they can incorporate the advice, it had a great impact.”
Next the team plans to refine the model to make it easier to use. They hope to eventually create a version that students could use at home.
“We’d like to offer computer vision and AI technology that allows someone to practice in the comfort of their home with a suturing kit and a smart phone,” Unberath said. “This will help us scale up training in the medical fields. It’s really about how can we use this technology to solve problems.”
Authors include Lalithkumar Seenivasan, Xinrui Zou; Jeewoo Yoon; Sirui Chu; Ariel Leon; Patrick Kramer; Yu-Chun Ku; Jose L. Porras; and Masaru Ishii, all of Johns Hopkins, and Alejandro Martin-Gomez of University of Arkansas.
The team trained their model by tracking the hand movements of expert surgeons as they closed incisions.
The team trained their model by tracking the hand movements of expert surgeons as they closed incisions. When students try the same task, the AI texts them immediately to tell them how they compared to an expert and how to refine their technique.
Credit Johns Hopkins University
Can AI make us more creative? New study reveals surprising benefits of human-AI collaboration
Artificial intelligence (AI) is often seen as a tool to automate tasks and replace humans, but new research from Swansea University challenges this view, showing that AI can also act as a creative, engaging and inspiring partner.
A team from the University’s Computer Science Department has conducted one of the largest studies to date on how humans collaborate with AI during design tasks. More than 800 participants took part in an online experiment using an AI-powered system that supported users as they designed virtual cars.
Unlike many AI tools that optimise solutions behind the scenes, this system used a technique called MAP-Elites to generate diverse visual design galleries. These galleries included a wide range of potential car designs, including high-performing examples, unusual ideas and some deliberately imperfect ones.
Turing Fellow Dr Sean Walton, Associate Professor of Computer Science and lead author of the study, explained: “People often think of AI as something that speeds up tasks or improves efficiency, but our findings suggest something far more interesting. When people were shown AI-generated design suggestions, they spent more time on the task, produced better designs and felt more involved. It was not just about efficiency. It was about creativity and collaboration.”
A key insight from the study, published in the ACM journal Transactions on Interactive Intelligent Systems, is that traditional ways of evaluating AI design tools may be too narrow. Metrics such as how often users click or copy suggestions fail to capture the emotional, cognitive and behavioural dimensions of engagement. The Swansea team argues for more holistic evaluation methods that consider how AI systems influence how people feel, think and explore.
Dr Walton added: “Our study highlights the importance of diversity in AI output. Participants responded most positively to galleries that included a wide variety of ideas, including bad ones! These helped them move beyond their initial assumptions and explore a broader design space. This structured diversity prevented early fixation and encouraged creative risk-taking.
“As AI becomes increasingly embedded in creative fields, from engineering and architecture to music and game design, understanding how humans and intelligent systems work together is essential. As the technology evolves, the question is not only what AI can do but how it can help us think, create and collaborate more effectively.”
Researchers from Saarland University and the Max Planck Institute for Software Systems have, for the first time, shown that the reactions of humans and large language models (LLMs) to complex or misleading program code significantly align, by comparing brain activity of study participants with model uncertainty. Building on this, the team developed a data-driven method to automatically detect such confusing areas in code — a promising step toward better AI assistants for software development.
The team led by Sven Apel, Professor of Software Engineering at Saarland University, and Dr. Mariya Toneva, researcher at the Max Planck Institute for Software Systems, investigated how humans and large language models respond to confusing program code. The characteristics of such code, known as atoms of confusion, are well studied: They are short, syntactically correct programming patterns that are misleading for humans and can throw even experienced developers off track.
To find out whether LLMs and humans “think” about the same stumbling blocks, the research team used an interdisciplinary approach: On the one hand, they used data from an earlier study by Apel and colleagues, in which participants read confusing and clean code variants while their brain activity and attention were measured using electroencephalography (EEG) and eye tracking. On the other hand, they analyzed the “confusion” or model uncertainty of LLMs using so-called perplexity values. Perplexity is an established metric for evaluating language models by quantifying their uncertainty in predicting sequences of text tokens based on their probability.
The result: Wherever humans got stuck on code, the LLM also showed increased perplexity. EEG signals from participants—especially the so-called late frontal positivity, which in language research is associated with unexpected sentence endings—rose precisely where the language model’s uncertainty spiked. “We were astounded that the peaks in brain activity and model uncertainty showed significant correlations,” says Youssef Abdelsalam, who was advised by Toneva and Apel and was instrumental in conducting the study as part of his doctoral studies.
Based on this similarity, the researchers developed a data-driven method that automatically detects and highlights unclear parts of code. In more than 60 percent of cases, the algorithm successfully identified known, manually annotated confusing patterns in the test code and even discovered more than 150 new, previously unrecognized patterns that also coincided with increased brain activity.
“With this work, we are taking a step toward a better understanding of the alignment between humans and machines,” says Max Planck researcher Mariya Toneva. “If we know when and why LLMs and humans stumble in the same places, we can develop tools that make code more understandable and significantly improve human–AI collaboration,” adds Professor Sven Apel.
Through their project, the researchers are building a bridge between neuroscience, software engineering, and artificial intelligence. The study, currently published as a preprint, was accepted for publication at the International Conference on Software Engineering (ICSE), one of the world’s leading conferences in the field of software development. The conference will take place in Rio de Janeiro in April 2026. The authors of the study are: Youssef Abdelsalam, Norman Peitek, Anna-Maria Maurer, Mariya Toneva, and Sven Apel.
Preprint:
Y. Abdelsalam, N. Peitek, A.-M. Maurer, M. Toneva, S. Apel (2025): “How do Humans and LLMs Process Confusing Code?” arXiv:2508.18547v1 [cs.SE], August 25, 2025. https://arxiv.org/abs/2508.18547
Large language models (LLMs) such as ChatGPT and Gemini were originally designed to work with text only. Today, they have evolved into systems that can work with many types of information at once (multimodal systems), as well as understand and generate images, audio, speech and music.
The most common way to add speech to multimodal models is to convert it into small building blocks called audio tokens, which function for audio much like characters do for text. However, audio tokens still carry a lot of information, which makes speech harder to handle than text. Despite recent progress, integrating speech into large language models remains a major challenge.
“Because of this complexity, standard audio tokens often have a high bitrate (the amount of information packed into each second of audio). They pack a huge amount of information into each second of audio, which makes it difficult for large language models to learn from speech efficiently.”
A focus on speech’s meaning
Della Libera and his collaborators developed FocalCodec, a new audio tokenization method that compresses speech far more efficiently than previous approaches. It preserves both the sound and meaning of words at an ultra-low bitrate.
Instead of relying on heavy processing steps, the system uses a simple way of turning audio into compact units (binary spherical quantization) and a technique that helps the model focus on the most meaningful parts of speech (focal modulation). This makes the analysis faster and keeps the essential qualities of the voice intact.
To test FocalCodec, the team conducted a listening study with 33 participants who compared different audio samples. Participants often judged the reconstructed speech as nearly identical to the original recordings. This shows that the system can shrink speech significantly without making it sound robotic or distorted.
Recognized at a top AI conference
The work has been accepted at the Thirty-Ninth Annual Conference on Neural Information Processing Systems, one of the most selective conferences in machine learning and artificial intelligence.
“This work is particularly important, as it introduces a novel approach that can be highly valuable for building modern multimodal LLMs,” says Mirco Ravanelli, assistant professor and Della Libera’s supervisor. “By making speech lighter and easier to integrate, we move closer to AI systems that understand sound with the same confidence they bring to text.”
The paper also includes contributions from Francesco Paissan, visiting researcher at Mila and undergraduate student at the University of Trento, and Cem Subakan, affiliate assistant professor at Concordia.
