%41%52
-
KAIST Reveals How the Light-Responsive Molecular Switch Azobenzene Changes Shape
Azobenzene is one of the best-known molecules that can be switched between two forms by light. However, the question of how the molecule moves in the first few picoseconds (trillionths of a second) after it absorbs light has remained unresolved for nearly 50 years. A Korean research team has now shown that, during the initial stage of the reaction, the molecule changes shape through a coordinated motion of the two nitrogen atoms at its center, while its two benzene rings stay nearly in place. The finding is expected to provide useful information for designing materials and molecular-scale machines that operate with light.
KAIST (President Choongsik Bae) announced on October 1 that a research team led by Hyotcherl Ihee, Professor in the Department of Chemistry at KAIST and Director of the Center for Advanced Reaction Dynamics (CARD) at the Institute for Basic Science (IBS, President Sukbok Chang), has identified the process by which the light-responsive molecule azobenzene changes its structure.
Azobenzene consists of two benzene rings connected by a central linkage of two nitrogen atoms. When it absorbs light, the molecule can change from a form in which the two rings are on opposite sides of the linkage (the trans form) to a form in which they are on the same side (the cis form). The molecule keeps the same atoms, and only their positions change. Researchers have used this property to control the activity of drugs with light and to develop light-responsive materials and molecular machines.
However, the structures of the molecule before and after the change do not show how it moves between the two forms. Researchers have proposed different explanations. Some suggested that the two rings rotate significantly, some that the central linkage straightens, and others that several parts of the molecule twist together. The question has been difficult to resolve because the structures that form during the reaction exist for too short a time to be observed directly.
The research team used the X-ray free-electron laser at the Pohang Accelerator Laboratory (PAL-XFEL) to observe these structures. The researchers first started the reaction by irradiating azobenzene dissolved in methanol with a laser. They then measured how its structure changed over time using ultrafast X-ray pulses.
The main challenge was to detect the weak signal from azobenzene, which was obscured by the much stronger signal from the surrounding solvent. The team separated the two signals using an analysis method that mathematically removes the contribution of the solvent. Based on the measurements taken at each point in time, the researchers then reconstructed the motion of the molecule as a movie.
The results showed that azobenzene does not change shape by rotating its two large benzene rings significantly at the same time. Instead, the reaction begins with torsion about the carbon–nitrogen bonds. During this initial motion, the central nitrogen linkage moves in a coordinated motion, similar to the way the two pedals of a bicycle move, and this changes the overall shape of the molecule.
This finding helps explain why the reaction rate does not change significantly even when the surrounding liquid becomes more viscous. If the molecule had to rotate its two large rings, it would need to displace a large volume of the surrounding liquid. Because the motion occurs mainly in the central part of the molecule, much less liquid needs to be displaced.
"This study shows the pathway by which azobenzene changes its shape after it absorbs light," said Professor Hyotcherl Ihee. He added that the team expects the work to help researchers understand how a wide range of light-responsive molecules work, because it improves methods for observing the motion of organic molecules that react rapidly.
The study did not directly improve the performance of drugs or materials that use azobenzene. Its significance is that it identifies the pathway the molecule actually follows as it moves. This provides basic data that researchers can refer to when they design light-responsive materials and molecular machines.
Dr. Jungmin Kim and Dr. Hosung Ki, who both received their degrees from KAIST and are now at IBS, are co-first authors of the study. The findings were published online in the international journal Nature on September 30 (UK time).
Paper title: X-ray liquidography decodes complex motions in azobenzene isomerization,
DOI: 10.1038/s41586-026-11068-4
This research was supported by the Institute for Basic Science (IBS) Research Center Program of the Ministry of Science and ICT.
2026.10.02 View 811 -
KAIST’s World-First Full-Course Robot Marathon Completion Published in Nature
A quadruped robot ran 42.195 km, a distance that is not easy even for humans to complete, without replacing its battery. KAIST researchers have developed this world-first achievement into design principles for long-distance operation, published the results in Nature, and are now moving toward commercializing robots that can be used in real industrial environments.
KAIST (President Choongsik Bae) announced on the September 24 that the long-distance running technology of RAIBO2, a quadruped robot developed by a research team led by Professor Jemin Hwangbo from the Department of Mechanical Engineering, and the results of its actual marathon completion were published in the international journal Nature.
This is the first time that robotics research conducted in Korea has been published in the main journal of Nature. Even globally, it is extremely rare for robotics research to be published in Nature, making this achievement an important demonstration of both the technological capability and academic level of Korean robotics research.
RAIBO2 completed the full 42.195 km course of the Sangju Dried Persimmon Marathon held in November 2024 in 4 hours, 19 minutes, and 52 seconds without charging or replacing its battery. It was the world’s first case of a quadruped robot completing the full course of an actual marathon event.
The core significance of this paper does not lie simply in the record of a robot running 42.195 km. In the global robotics industry, competition has recently intensified over the running speeds and obstacle-negotiation capabilities of quadruped robots and humanoids. However, to operate robots for long periods in real industrial environments, it is necessary to secure not only momentary locomotion performance but also energy efficiency, thermal management, battery performance, and actuator stability.
The research team approached long-distance operation not as a performance issue involving individual components such as batteries or motors, but as a “whole-robot system problem” in which mechanical structure, electrical systems, and locomotion control are closely interconnected.
Quadruped robots lose energy in multiple areas – including actuators, electrical systems, and mechanical components – as they repeatedly move their legs. Improving the efficiency of only one component makes it difficult to achieve both fast and stable locomotion performance and long travel distances at the same time.
The research team systematically analyzed energy losses throughout the entire robot and jointly designed the mechanical, electrical, actuation, and control systems so that energy consumption could be reduced while maintaining high locomotion performance. Through this approach, the researchers presented specific design principles for which factors must be considered together to enable quadruped robots to move stably for extended periods.
In particular, the research team went beyond short-distance laboratory running tests or maximum travel distances estimated through calculations and validated the technology on an actual full marathon course.
During the marathon, the team recorded the robot’s internal voltage, current, temperature, and battery status in real time and synchronized the data with position, speed, and altitude information measured by GPS. RAIBO2 was officially registered as a participant, allowing formal split times to be recorded, while the researchers ran alongside the robot and filmed the entire journey, lasting more than four hours from the starting line to the finish.
While the 42.195 km completion record demonstrated RAIBO2’s performance, the data accumulated throughout the process provided scientific evidence showing how that performance was made possible and how it was maintained in a real-world environment. The research team developed a single impressive demonstration into research results based on system design principles and actual measurement data that can be analyzed and revalidated.
The research team’s next challenge is to transfer the performance achieved with a single research robot into a product that can be used repeatedly in industrial environments.
RAIBO2 is a research robot that was designed and built by hand by the researchers themselves. To develop it into a product that can be deployed in real industrial environments, the same performance must be achieved reliably across multiple robots while also reducing production costs. The product must also meet industrial reliability requirements, including water and dust resistance, operation across a wide temperature range, durability against shock and vibration, and electromagnetic interference and compatibility requirements (EMI/EMC).
Raion Robotics, a KAIST faculty startup that originated from Professor Hwang’s laboratory, is responsible for connecting the research results to actual products. During the development of RAIBO2, Raion Robotics’ manufacturing technology also contributed to maintaining the high energy efficiency and locomotion performance achieved in the laboratory for more than four hours during the actual marathon.
The research team and Raion Robotics developed not only the mechanical and electrical systems, software, and artificial intelligence but also the motor drivers that precisely control the robot’s motors, allowing them to optimize the entire system.
Currently, Raion Robotics is using this systems-design capability to develop commercialization and mass-production technologies for quadruped robots that can operate for extended periods in real environments, going beyond research prototypes. The goal is to reproduce the world-class performance demonstrated by a single laboratory robot across multiple products and enable stable operation in a wide variety of environments.
Professor Jemin Hwang said, “This study goes beyond demonstrating how far a quadruped robot can run. Using actual marathon data, we demonstrated how the entire robot must be designed to achieve both high locomotion performance and energy efficiency,” adding, “We will connect the world-class performance achieved in the laboratory to products that anyone can use reliably in industrial environments.”
Dr. Choongin Lee and students Donghoon Youm and Jeongsoo Park of the KAIST Department of Mechanical Engineering participated in the study as co-first authors, and the research results were published in the international journal Nature on September 23 (local time).
Paper title: A quadruped robot designed to complete a marathon on a single battery charge, DOI: 10.1038/s41586-026-11102-5
Related photos and videos:
Comprehensive photo/video collection on Google Drive: https://drive.google.com/drive/folders/1ye82tTr2dCK2Bt7Ih0hSgWqkgqlVXk6X?usp=sharing
Full video of the marathon completion on YouTube: https://www.youtube.com/watch?v=-HMFoa3g9nA
The entire process of RAIBO2 completing the marathon can be viewed in the YouTube video.
2026.09.28 View 2117 -
KAIST Launches Korea's First Artificial General Intelligence (AGI) Research Institute
Moving beyond today’s AI, which excels at solving problems similar to those encountered during training, KAIST has launched a dedicated research institute to develop foundational technologies for Artificial General Intelligence (AGI) that can make decisions and adapt independently when faced with unfamiliar problems and environments.
KAIST announced that it officially launched the AGI Research Institute under the KAIST AI Research Institute (AIRI) on September 18, following the completion of internal approval procedures.
The research team initially comprises 11 professors: Kim Tae-Kyun (T-K), Kim Hyunwoo, and Oh Tae-Hyun (School of Computing); Sung Minhyuk (Department of AI Computing); Shin Jinwoo and Ye Jong Chul (Kim Jaechul Graduate School of AI); Kim Junmo and Chung Joon Son (Department of Electrical Engineering); Ahn Sungjin (Department of AI Computing); and Park Noseong and Yoon Sung-Eui (School of Computing). Professor Kim Tae-Kyun (T-K) of the School of Computing was appointed as the institute’s inaugural director, with plans to gradually expand the team.
The AGI Research Institute has been established with the mission of advancing the science and engineering of AGI through fundamental research and returning the benefits to society. Its core research focus is “generalization”—the ability of AI to apply learned knowledge to previously unseen tasks, environments, and data domains.
While current AI systems have achieved remarkable results through the scaling of data and computational power, they still struggle to learn and reason beyond their learnt domains. They perform well on problems similar to their training data, but their performance can decline when conditions or environments change.
The institute plans to move beyond simply scaling data and computing resources by developing foundational technologies that enable AI to reason about previously unseen problems and adapt to new environments, even with limited resources.
To this end, the institute will investigate the fundamental principles of AI through research into novel and efficient neural network architectures, learning principles, understanding the physical world, and the implementation of intelligence. It will also cultivate the next generation of research leaders.
The institute’s key research areas include:
• Efficient Foundation Models: Achieving high performance with less computation.
• Novel Knowledge Discovery (AI Meta-Scientist): Enabling AI to generate its own hypotheses and discover new knowledge.
• World Understanding: Developing world models that understand the structure and dynamics of the physical world and predict the future.
• Physical AI: Enabling embodied systems, such as robots, to perceive and act within real-world environments.
• AI Safety: Ensuring that AI remains aligned with human intent.
In Efficient Foundation Models, the institute will develop neural network architectures and learning methods that go beyond Transformers, with the aim of enabling AGI even in resource-constrained environments. In AI Meta-Scientist research, it will develop AI systems that go beyond exploring existing knowledge to generate their own hypotheses, conduct scientific research, and discover new knowledge.
By integrating these interconnected research areas into a single research institute, KAIST aims to establish itself as a global hub for AGI research.
Director Kim Tae-Kyun said, “There are limits to how far AI can advance simply by continuing to expand AI data centers.” He added, “Sustained efforts to develop foundational AI technologies that can adapt to new problems even in resource-constrained environments will be critical to staying competitive in the next generation of AI.”
The institute will pursue its research through national R&D projects supported by the Ministry of Science and ICT. Funding includes a total of KRW 10 billion over five years for the development of lightweight, low-power next-generation neural network architectures and KRW 18.75 billion over five years for AI Meta-Scientist research. The institute is also participating in a project led by LG Electronics to develop world models and robotics-based Physical AI technologies. That project has a total budget of KRW 34 billion over two years. In addition, the institute is conducting research on physically grounded intelligence.
Choi Sung-Yool, Vice President of Research at KAIST, remarked, “I expect KAIST’s outstanding AI research teams to make significant contributions to developing core next-generation AGI technologies that lead the way toward sustainable human development.”
Hong Jin-bae, President of IITP, stated, “As creative ideas and challenging research drive AI technological innovation, I look forward to the new AGI research institute becoming a hub of innovation that unlocks new possibilities for the future of artificial intelligence.”
2026.09.22 View 1603 -
KAIST Names Former Samsung Electronics Chairman and CEO Oh Hyun Kwon as Chair of Its Board of Trustees
KAIST announced that its Board of Trustees elected Oh Hyun Kwon, former Chairman and CEO of Samsung Electronics and Chairman of the Orange Planet startup foundation, as its new chair at the board’s 298th extraordinary meeting in Seoul on September 21, 2026. Prior to the election, Kichul Cha, a KAIST trustee and CEO of InBody, had served as acting chair since the term of former Chair Myung Ja Kim ended in late June.
Kwon earned a bachelor’s degree in electrical engineering from Seoul National University in 1975, a master’s degree in the same field from KAIST in 1977, and a doctorate in electrical engineering from Stanford University in 1985. He joined Samsung in 1985 as a researcher at its Semiconductor Research Institute in the United States. In 1992, he led the development of the world’s first 64Mb DRAM, helping lay the foundation for Samsung’s strategy of building a decisive technological lead over its competitors.
An engineer who rose through the ranks to become a leading business executive, Kwon served as President and Head of Device Solutions Division, Samsung Electronics from 2008 to 2011 and as Vice Chairman and CEO from 2012 to 2017. From 2017 to 2020, he held the positions of Chairman at both Samsung Electronics and the Samsung Advanced Institute of Technology. He is credited with helping propel South Korea's semiconductor industry to global leadership.
After stepping back from corporate management, Kwon served as President of the Korea Invention Promotion Association from 2021 to 2022 and as the seventh Chair of Seoul National University’s Board of Trustees from 2023 to 2025. In September 2025, he became the first business executive to chair the Committee on Mid- to Long-Term Strategy Committee, now under the Ministry of Planning and Budget (then called the Ministry of Economy and Finance). He has also chaired Orange Planet since 2021, supporting young entrepreneurs.
His honors include the Gold Tower Order of Industrial Service Merit in 2009, the Korea Best Scientist and Engineer Award in 2014, the Inchon Award in Industrial Technology category in 2012, and the Dasan Technology Award in 1997. He is also the author of Super Gap (2018) and Again, Super Gap (2020).
Kwon has maintained close ties with his alma mater, KAIST. In 2018, for instance, he became the first KAIST alumnus to receive an honorary doctorate from the university, earning the distinction in business administration. He also helped establish the Kwon Oh Hyun Assistant Professor program, which provides annual support for promising faculty members pursuing creative and ambitious research in physics, mathematics, chemistry, biological sciences, and engineering.
Kwon joined the KAIST Board of Trustees at its 297th extraordinary meeting in August before being elected chair at the September meeting.
His term as chair runs from September 21, 2026, the date of the board resolution, through August 19, 2029, when his term as a trustee expires. Drawing on his experience in building Samsung’s technological leadership, Kwon is expected to help KAIST advance innovation in research and education and further strengthen its global competitiveness.(END)
2026.09.22 View 2965 -
KAIST Develops Smartphone AI That Uses Past Solutions to Tackle Similar Problems
A new AI technology has been developed that retains and reuses knowledge, much like writing down a solution in a notebook and applying it to similar problems rather than asking an expert for help each time. Researchers at KAIST have developed a way for small AI models running on smartphones to store and reuse knowledge from a large server model. The approach reduced server calls by an average of 55.61% compared with a non-cumulative approach while maintaining high accuracy, suggesting that mobile AI could make faster decisions with less server support as it encounters similar problems.
KAIST (President Choongsik Bae) announced on September 21 that a research team led by Professor Jae-Gil Lee from the School of Computing has developed CURE (Cumulative Knowledge Reuse), a technology that enables a small AI model on a device to work efficiently with a large model on a server.
Smartphones have limited processing power and memory, so they typically use small, lightweight AI models. These models can handle simple tasks quickly but may be less accurate when identifying complex or unfamiliar images.
Sending every input to a powerful server model can improve accuracy, but transferring data and waiting for a response takes time. It also adds to network traffic and the server’s computational workload, increasing delays for users and operating costs for service providers.
Researchers have explored a collaborative approach in which the on-device model assesses each input first and sends only difficult cases to the server. Without a way to retain the server’s knowledge, however, the device uses each answer once and may need the same help when a similar input appears. This is much like a student who fails to write down the solution to a difficult problem and has to ask for help again when faced with another of the same kind. The team set out to turn these one-time answers into knowledge the device could continue to use.
CURE first checks whether the on-device model can handle an input reliably. If it cannot, the system consults knowledge previously obtained from the server and stored on the device. It contacts the server only if that knowledge is also insufficient. The process follows three steps, moving from solving a problem independently to consulting previous lessons and, when necessary, asking an expert.
For example, an on-device model that cannot identify a car model in a photo may turn to the server for help. CURE uses the server’s prediction and the image’s features to update the device’s knowledge store. When the device later receives a photo of a similar car model, it can use that knowledge to identify it locally.
CURE does not keep a collection of the original photos processed by the server. Instead, it stores a summary of their shared features and differences. This is like noting a car’s distinguishing features, such as its body shape or headlight design, rather than memorizing the entire photo. The stored knowledge can therefore help the system recognize not only images it has already seen but also similar images it encounters for the first time.
The team tested CURE using vision-language models, which connect images with text to understand visual information. These models link what they see to language, much as people do.
The researchers used MobileCLIP2 on the device and EVA-CLIP, which has 18 billion parameters, on the server. Parameters are numerical values within a model that encode information learned during training. They help the model distinguish objects and identify relationships between them.
In tests on a range of image classification datasets, CURE achieved accuracy close to that of an approach that sends every input to the large server model. It also made an average of 55.61 percent fewer server calls than a device-server collaboration baseline that does not retain knowledge from previous server responses.
In end-to-end tests that included communication time, CURE ran up to 2.80 times as fast as the non-cumulative device-server collaboration baseline and up to 3.67 times as fast as the approach that sends every input to the server. Making fewer server calls reduced the time spent transferring data and waiting for responses.
CURE requires no retraining of either the on-device model or the server model. It leaves both models unchanged and adds a separate store for knowledge obtained from the server, making it adaptable to a range of AI models and services.
The technology could reduce repeated data transfers and server computation image recognition. This could mean shorter waits for users and lower operating costs for service providers.
Potential applications also include robots and wearables that need to recognize their surroundings with limited computing resources. Robots operating over slow or unreliable networks, for example, could use previously acquired knowledge to make more decisions locally without waiting for a server response.
The benefits in practice will depend on the device’s processing power and storage capacity, network conditions, and the characteristics of the input data. Further testing across different devices and environments is therefore needed.
“CURE allows a small on-device AI model to remember and reuse knowledge it has already obtained, rather than repeatedly asking the server the same question,” said Professor Jae-Gil Lee. “By maintaining high accuracy while reducing communication demands and response times, we expect it to help smartphones, robots, and wearables use powerful AI models more efficiently.”
Dr. Youngjun Lee, a postdoctoral researcher from the KAIST Institute of Information Electronics, was the study’s first author, and Professor Jae-Gil Lee from the School of Computing was the corresponding author. Co-authors were Doyoung Kim from Amazon, Junhyeok Kang from LG AI Research, and Professor Hwanjun Song from the KAIST Department of Industrial and Systems Engineering. The findings were presented on September 10 at the European Conference on Computer Vision (ECCV 2026), a leading international computer vision conference held in Malmö, Sweden, from September 8 to 12.
Paper title “CURE: Cumulative Knowledge Reuse for Efficient Device-Server Hybrid Inference in Vision-Language Models”,
DOI: https://doi.org/10.1007/978-3-032-37627-5_21
Author information — Dr. Youngjun Lee (postdoctoral researcher, KAIST Institute of Information Electronics, first author), Doyoung Kim (KAIST graduate, now at Amazon, co-author), Junhyeok Kang (KAIST graduate, now at LG AI Research, co-author), Professor Hwanjun Song (KAIST Department of Industrial and Systems Engineering, co-author), Professor Jae-Gil Lee (KAIST School of Computing, corresponding author)
This work was supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2020-II200862, DB4DL: High-Usability and Performance In-Memory Distributed DBMS for Deep Learning, 50% and No. RS-2022-II220157, Robust, Fair, Extensible Data-Centric Continual Learning, 40%), the InnoCORE pro-gram of the Ministry of Science and ICT (AI Meta-Scientist, N10260110, 10%), and Samsung Electronics Co., Ltd.(IO251216-14634-01).
2026.09.21 View 2318 -
AI Takes the Wheel in 'Mario Kart' as Students Use Generative AI in Eight-Hour Hacking Contest
AI will take the wheel in "Mario Kart," while students use generative AI tools to tackle an eight-hour hacking contest, as KAIST and POSTECH bring their scientific rivalry to the playing field.
KAIST (President Choongsik Bae) and POSTECH (President Seong Keun Kim) announced on September 18 that the 23rd KAIST-POSTECH Science War, also known as the "KA-PO War," will take place at POSTECH on September 18 and 19.
First held in 2002, the annual event brings students from both universities together for scientific competitions, sports matches, and performances. The two universities take turns hosting the event in Daejeon and Pohang. Following convention, the visiting university's name comes first in the event title, making this year's edition the "KA-PO War." The name "Science War" reflects the event's distinctive combination of technical challenges and athletic competitions.
This year's games feature seven events. The four sports events are baseball, soccer, basketball, and esports matches featuring "League of Legends", while the three science events are a hacking competition, an AI competition, and a science quiz.
This year's science events highlight current technology trends, incorporating generative AI and concepts from autonomous driving into the competitions.
The hacking competition will run for eight hours, from 2:00 p.m. to 10:00 p.m. on the first day. The competition will present information security challenges, including cryptography problems, in rounds, with participants earning points for solving them within the allotted time. The use of generative AI tools is fully permitted. A screen will display live scores and participants' progress solving the challenges so spectators can follow the action.
Also beginning at 2:00 p.m. on the first day, the AI competition will take place at the POSTECH eSports Colosseum. AI programs developed and trained by students from both universities will control karts in the popular racing game "Mario Kart Wii" and compete head-to-head. With the AI used to drive the karts without human input, the event serves as a virtual autonomous driving competition.
Scores across four courses will determine the winner of each round. The first team to win two of three rounds will win the match.
The science quiz will begin at 2:00 p.m. on the second day on the stage in front of POSPLEX, combining questions on general science and recent research trends with 'rock-paper-scissors chess,' a chess variant that tests strategic thinking and teamwork alongside scientific knowledge.
The university that wins at least four of the seven events will be crowned this year's overall champion. KAIST currently leads the all-time standings with 13 overall victories to POSTECH's eight.
"We hope this year's KA-PO War goes beyond a friendly exchange to become a special festival where students from Korea's two leading science and technology universities compete fiercely with their pride on the line, while sharing laughter and enjoying themselves together," said Dong-hyeok Lee, chair of the KA-PO War organizing committee and a student in KAIST's Department of Materials Science and Engineering who entered KAIST in 2024.
"The KA-PO War is more than a competition to determine winners and losers. It is a special festival where KAIST and POSTECH students share their passion and talents through science and sports," said KAIST President Choongsik Bae. He added that he hopes the event inspires students to take pride in their role as future leaders of Korean science and technology and gives them a valuable opportunity to learn from one another and grow together through competition.
2026.09.18 View 1379 -
KAIST Researchers Uncover How Neurons Deliver RNA to Where It Is Needed
Much like a courier reading an address on a package, neurons must deliver RNA to the precise locations where it is needed. A KAIST research team has discovered how a small chemical modification on RNA acts as a “delivery tag,” helping selected RNAs travel to distant regions of neurons.
The findings provide a basis for investigating RNA “delivery errors” in brain disorders—cases in which RNA is produced normally but fails to reach its intended destination—and may ultimately inform the development of new therapeutic strategies.
KAIST (President Choongsik Bae) announced on September 14 that a research team led by Professor Ki-Jun Yoon of the Department of Biological Sciences has identified a mechanism through which N6-methyladenosine, or m6A, facilitates the transport of specific RNAs to distal axons, the long, cable-like projections of neurons.
Unlike most other cells, neurons have long, highly branched extensions. Axons transmit signals to other neurons, while dendrites receive incoming signals. For the brain to develop and function normally, these structures must grow properly and form appropriate connections.
Because neurons can extend over considerable distances, RNA produced in the cell body may be needed far away, including at the growing tip of an axon. Neurons must therefore identify particular RNAs among thousands of candidates and transport them to the correct subcellular locations.
Just as producing a package does not complete its delivery—the package must also reach the correct address—producing RNA is only part of the process. RNA must arrive at the right place at the right time. However, the mechanisms through which neurons select and transport RNAs over long distances have remained incompletely understood.
The research team found that m6A, a small chemical modification added to RNA, plays an important role in this process. m6A is a modification of adenosine, one of the building blocks of RNA, and is known to influence several aspects of RNA metabolism, including stability and translation. The new study shows that m6A can also function as a molecular “delivery tag” that promotes the transport of specific RNAs to distal regions of neurons.
The researchers first examined mice in which Mettl14, a component of the enzyme complex that installs m6A on RNA, was deleted from developing neurons. In the absence of normal m6A modification, the neurons exhibited impaired axon projection and neurite development. These findings confirmed that m6A plays an important role in the early development of neurons.
The team then used a high-resolution technique known as m6A-SAC-seq to map the precise locations of m6A modifications in RNA from the developing mouse brain at single-nucleotide resolution. The resulting map revealed that RNAs associated with axon development and synapse organization were among those marked by m6A.
The researchers also identified how this “delivery tag” is recognized and connected to the cellular transport machinery. They found that YTHDF2, a protein that recognizes m6A-modified RNA, works with the RNA-binding protein FMRP and the motor protein KIF5C to facilitate the transport of RNA along neuronal processes.
In the delivery analogy, m6A serves as the tag attached to the RNA, YTHDF2 reads the tag, and FMRP and KIF5C help connect the RNA to the cellular machinery that carries it toward its destination.
YTHDF2 has previously been studied primarily for its role in promoting the degradation of RNA that is no longer needed. This study reveals an additional function: in developing neurons, YTHDF2 also helps transport specific RNAs to distal locations rather than simply directing them toward degradation.
The findings offer a new perspective on the study of brain disorders. Research has traditionally focused on whether RNA is produced and maintained correctly. The new results suggest that researchers should also consider whether normally produced RNA is delivered to the correct location within a neuron.
In other words, neurological dysfunction may potentially arise not only from defects in RNA itself, but also from problems with the molecular tags or transport machinery that determine where the RNA travels.
Future studies could investigate whether such RNA delivery errors occur in neurodevelopmental or neurodegenerative disorders and whether they contribute to neuronal dysfunction or disease onset. If such links are established, the findings could eventually support the development of therapeutic approaches that regulate RNA transport and restore the delivery of specific RNAs to their intended locations.
The high-resolution m6A map of the developing nervous system generated in this study may also serve as a valuable resource for examining how RNA modification and intracellular transport change during brain aging and neurodegenerative disease.
“Our study shows that a small chemical modification on RNA does more than regulate RNA stability—it also plays an important role in delivering specific RNAs to the locations where they are needed within neurons,” said Professor Yoon. “We hope these findings will provide a foundation for understanding why RNA may fail to reach its intended destination in brain disorders and for exploring new therapeutic possibilities.”
Bonsang Koo, Dr. Ajeet Kumar, and Huiseon Hwang of the KAIST Department of Biological Sciences contributed equally to the study as co-first authors. The paper was published in Nature Communications on July 30.
Paper title: The m⁶A reader YTHDF2 regulates mRNA transport and axon outgrowth in developing neurons
DOI: 10.1038/s41467-026-76161-8
Author information: Co-first authors (Bonsang Koo, Ajeet Kumar, Huiseon Hwang); Contributing authors (Minyeong Park, Sung-Min Lee, Ki-Heon Lee, Ji-Hoon Moon, Jaewon Yang, Chan-Woo Park, Saebom Lee, Hee-Jung Jo, Woo-il Kim, Eunji Cho, Hagyeong Lee, Chang-Won Lee, Jae-Sung Woo, Jong Hyuk Yoon, Yoon Ki Kim); Corresponding author (Ki-Jun Yoon)
This work was supported by National Research Foundation of Korea (NRF) grants (2020M3A9E403967022, RS-2024-00332454, RS-2024-00440778, and RS-2026-25488918 to K.-J.Y.) funded by the Korean Ministry of Science and ICT, and Future Planning (MSIP), the POSCOScience Fellowship from the POSCO TJ Park Foundation (to K.-J.Y.), the Young Investigator Grant from the Suh Kyungbae Foundation (to K.-J.Y.), the Korea Brain Research Institute (KBRI) Basic Research Program through the Korea Brain Research Institute, funded by the Ministry of Science and ICT (MSIT) (24-BR-02-03 to J.H.Y).
2026.09.14 View 3805 -
KAIST and KIOM Develop a Wearable Microneedle Electroceutical for Personalized Pain Management
A "wearable electroceutical" has been developed that could reduce the need to visit a hospital or rely on painkillers every time pain occurs. When attached to the skin, it modulates pain through electrical stimulation and can be remotely controlled via smartphone even across long distances, such as between Korea and the United States.
KAIST (President Choongsik Bae) announced on September 13 that a joint research team led by Professor Jae-Woong Jeong from the School of Electrical Engineering at KAIST and Dr. Sanghun Lee from the KIOM (Korea Institute of Oriental Medicine, President Sung-Kyu Kho) has developed a wearable electroceutical platform that combines a wireless microneedle (fine needles that adhere to the skin) electroceutical with Internet of Things (IoT)-based remote control technology.
The key feature of this research is that a single small skin-attached device integrates stable electrical stimulation, smartphone-based remote control, and automatic stimulation based on the body's physiological state. Once its efficacy and safety are validated with actual patients, the technology may be used for personalized pain management at home or in daily life.
The need for such an approach is particularly relevant for chronic pain management. Painkillers are widely used to treat chronic pain, but long-term medication use raises concerns about side effects and dependency. In particular, opioid analgesics used for conditions such as cancer pain carry an increasing risk of tolerance and misuse with prolonged use, which has spurred research into "electroceuticals", devices that modulate nerves through electrical stimulation as an alternative to drugs.
However, existing electroceuticals have limitations. Implantable devices require surgery, and skin-attached electrodes may fail to deliver electrical current properly depending on skin conditions such as sweat or dead skin cells. A concentrated current at specific sites can also raise skin temperature or pose a risk of burns.
To address these issues, the research team developed a temperature-responsive, conductive, and adhesive microneedle electrode (a fine needle electrode that conducts electricity well while adhering to the skin).
Because the microneedles penetrate the highly resistive stratum corneum, the device can deliver stable electrical stimulation while reducing the influence of sweat and dead skin cells. The team also coated the electrode with a conductive hydrogel (a gel-like material that retains a large amount of water) so that current spreads evenly rather than concentrating at the needle tips.
In addition to improving electrical performance, the electrode was designed with a built-in thermal safety mechanism. When skin temperature rises abnormally, the electrode's adhesion weakens and it detaches from the skin on its own, helping reduce the risk of skin burns that could occur during electrical stimulation.
Beyond the skin interface itself, the research team integrated the device with IoT-based remote management. Using a smartphone and cloud server, a healthcare provider can control the device's operating time and electrical stimulation in real time or on a scheduled basis, even when located far from the patient. The team confirmed that the device could be remotely controlled even across long international distances, such as between Korea and the United States.
Following future clinical validation, this could develop into a home-based or remote pain management approach in which patients use the electroceutical under medical supervision without needing to visit a hospital.
The platform further extends beyond remote control by enabling automatic operation based on the body's physiological state. Using a photoplethysmography (PPG) sensor (a technology that measures pulse and blood flow changes using light), the team detected pain-related stress states and, based on this, implemented a closed-loop (a method that automatically adjusts treatment while continuously checking the body's condition) therapy function that automatically triggers electrical stimulation.
The researchers evaluated the performance of the platform in both animal experiments and a small-scale human study. In animal experiments, current was delivered more effectively than with conventional gel electrodes, and pain-relieving effects were also confirmed. In a small-scale study involving healthy adults, changes in skin sensory pain thresholds (the level of stimulation at which pain begins to be felt) following electrical stimulation were observed to assess the potential for application in humans.
However, direct analgesic effects in this study were confirmed only through animal experiments. The research team noted that further clinical studies are needed to confirm therapeutic efficacy and the safety of long-term use in actual chronic pain patients.
Professor Jae-Woong Jeong from the KAIST School of Electrical Engineering said, "By combining a stable skin interface with IoT-based remote management, we have expanded the potential for wearable electroceuticals in daily life. We hope that, following future clinical validation, this approach can evolve into a personalized digital healthcare platform that enables pain management tailored to each patient's condition."
Dr. Sanghun Lee from the KIOM said, "We hope this research will be integrated with future wearable acupuncture technologies to contribute to the development of a new non-pharmacological pain management approach that stimulates acupoints through electrical stimulation."
The study, co-first-authored by Heesoo Kim, a PhD student at KAIST, and Dr. Se Kyun Bang from the KIOM, was published in the international journal Nature Communications on August 28th.
Paper title: Wireless IoT-Enabled Microneedle Electroceutical for Personalized and Connected Pain Management, DOI: 10.1038/s41467-026-76527-y
Author information: Heesoo Kim (KAIST, co-first author), Se Kyun Bang (KIOM/UST, co-first author), Sanghun Lee (KIOM, co-corresponding author), Jae-Woong Jeong (KAIST, corresponding author), and 10 others
Demonstration video: https://www.dropbox.com/scl/fo/zrglys1t1c49l5u1xo08e/AJsbboH3sQMFag-6aO6sSSs?rlkey=dnux0jj80bdczws22u0dq7ugx&e=1&dl=0
This work was supported by the National Research Foundation of Korea (RS-2022-NR067853, RS-2025-02218624, RS-2024-00335066), and by the Korea Institute of Oriental Medicine (KSN2511012 and KSN2511013).
2026.09.14 View 3210 -
KAIST Develops AI That Can Sense “That’s Not What I Meant” Without Being Told
A future in which AI can recognize a person’s unspoken “that’s not what I meant” response from brain signals and adjust its behavior on its own is coming closer. KAIST researchers have developed a technology that detects cognitive mismatch between humans and AI through brainwaves, enabling AI systems to revise their actions in real time according to human goals. The achievement is expected to accelerate the shift from AI that follows explicit commands to AI that can infer human intent.
KAIST (President Choongsik Bae) announced on September 11 that a research team led by Endowed Chair Professor Sang Wan Lee of the Department of Brain and Cognitive Sciences (Director of the Center for Neuroscience-Inspired Artificial Intelligence), in collaboration with Microsoft Research Asia (MSRA), has developed Neural Value Alignment (NVA), a next-generation brain–computer interface (BCI) technology that uses human brainwaves to align AI behavior with human intentin real time.
For AI to collaborate naturally with people, it must accurately understand what a person actually wants. Until now, however, AI systems have largely inferred human intent from externally observable information such as speech, actions, or gestures.
The challenge is that the same action can reflect different goals, and the same goal can also be achieved through different actions. For example, when a person picks up a cup, it may be unclear whether they intend to drink from it or hand it to someone else. Conversely, if the person's goal is simply to quench their thirst, that same goal could be pursued through several different actions — reaching for the cup, picking up a water bottle, or asking someone else to bring a drink. Because of this two-sided “goal–action ambiguity,” when AI misunderstands human intent, users often have to correct it through additional commands or actions.
To address this problem, the research team focused on the “prediction error” signals that arise unconsciously in the brain when a person encounters an unexpected situation. In simple terms, the team used the brain’s instant “that’s not what I meant” response when AI performs the wrong action or pursues the wrong goal.
The researchers distinguished between two types of brain responses. The first is reward prediction error (RPE), which appears when AI misunderstands the person’s ultimate goal. The second is state prediction error (SPE), which appears when the goal is correct but the process or method of action differs from what the person expected.
The team measured real-time electroencephalography (EEG) signals from people as they observed AI performing tasks. They found that the brain produced different signals depending on whether the AI misunderstood the goal itself or chose the wrong method while pursuing the correct goal. The researchers also identified distinctive brainwave patterns that appeared when both types of errors occurred simultaneously.
By applying deep learning to these brain signals, the team developed a technology that can determine, from EEG alone, how a person is interpreting the AI’s behavior. In other words, even without a person saying “that’s wrong,” the AI can recognize whether the human brain is signaling that “the goal is wrong” or “the method is wrong.”
The team then proposed a Neural Value Alignment-based human–AI synergy algorithm, which feeds these decoded brain signals back to the AI in real time so that it can correct its own behavior.
When the AI detects an SPE signal, it interprets the situation as “the desired goal is correct, but the method is wrong” and adjusts its action strategy. When it detects an RPE signal, it understands that “the goal itself was misunderstood” and searches again for what the person truly intended.
Simulation results showed that the proposed method adapted more quickly than existing approaches even in uncertain situations, such as when a person’s goal suddenly changed or some human neural feedbacks were missing.
The key significance of this research is that it demonstrates the possibility of AI systems correcting themselves by reading a person’s unconscious “that’s not what I meant” brain response, without requiring the user to repeatedly say “do it this way” or “that’s not right.”
As the technology advances, it could be applied to physical AI robots in homes and industrial settings, allowing them to understand user intent more naturally and adjust their actions accordingly. It could also be extended to autonomous vehicles that quickly reflect driver judgment, medical and rehabilitation robots for patients who have difficulty speaking or moving, and educational AI systems that adapt to a student’s cognitive state.
Ultimately, the study presents a new model of human–AI collaboration, moving beyond AI that acts only when explicitly instructed toward AI that can sense human responses and adjust itself accordingly.
Professor Sang Wan Lee, who led the international collaboration, said, “This research is meaningful because it shows that AI can move beyond inferring human intent only from visible behavioral outcomes and instead directly use cognitive signals generated in the brain during collaboration with AI.” He added, “The technology can be expanded to a wide range of fields where human judgment and AI behavior must be closely connected, including physical AI, BCI, autonomous driving, precision personalized education, medical robotics, and human–computer interaction.”
Miran Lee, Director, Microsoft Research Accelerator at Microsoft Research, said, “This achievement is the result of the ongoing international collaboration between KAIST and Microsoft Research Asia. We look forward to continuing this partnership to develop world-class BCI technologies that enable humans and AI to communicate and collaborate more naturally.”
The study’s first author is Xin Xu, a Ph.D. student in KAIST’s Department of Brain and Cognitive Sciences. Researchers from Microsoft Research Asia, including Yansen Wang, Dongqi Han, and Dongsheng Li, also participated in the study. The research was published online in August 2026 in IEEE Transactions on Cybernetics, an international journal in the field of cybernetics.
Paper title: Neural Value Alignment: Human–AI Collaboration Under Goal–Action Ambiguity
DOI: 10.1109/TCYB.2026.3722605
Another KAIST–Microsoft Research Asia collaborative study on helping AI rapidly adapt to continuously changing environments was presented in June at ICML 2026, one of the leading international conferences in artificial intelligence. The study was led by Niklas Koeppe, a Ph.D. student in KAIST’s Program of Brain and Cognitive Engineering, as first author.
Paper title: Mitigating Plasticity Loss through Architectural Design in Continual Learning
Original paper: https://icml.cc/virtual/2026/poster/61534
This research was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP), funded by the Ministry of Science and ICT.
2026.09.10 View 3110 -
KAIST Signs MOU with Université Paris-Saclay on Behalf of K-STAR
KAIST (President Choongsik Bae) announced on September 9 that, representing K-STAR, a consortium comprising KAIST, GIST, DGIST, UNIST, and POSTECH, it signed a memorandum of understanding (MOU) with Université Paris-Saclay on September 7 to strengthen research cooperation. The signing ceremony took place at Bâtiment Bréguet on the university’s campus.
The agreement was concluded on the occasion of President Lee Jae-myung’s state visit to France. It follows up on plans discussed at the ninth Korea-France Joint Committee on Science and Technology, held in Seoul in April, to expand scientific, technological, and academic exchanges between K-STAR and French universities.
Université Paris-Saclay is one of France’s leading research universities. It is home to a large research and education ecosystem encompassing universities, grandes écoles, and major national research organizations, including the French National Centre for Scientific Research (CNRS) and the French Alternative Energies and Atomic Energy Commission (CEA). The university has internationally recognized strengths across a broad range of fields, from mathematics and physics to artificial intelligence, quantum technology, biology, and engineering.
Under the agreement, K-STAR and Université Paris-Saclay will establish a practical framework for cooperation encompassing student exchanges, mobility programs for early-career researchers, and joint research. They plan to begin by offering research internships through which undergraduate and graduate students can gain experience in laboratories at partner institutions. They will also expand opportunities for exchanges and collaboration among postdoctoral researchers and early-career faculty members.
The two sides will promote joint research in key scientific and technological fields, including artificial intelligence, quantum technology, advanced biotechnology, and advanced materials. They will also facilitate shared access to research infrastructure and organize regular workshops and symposia, enabling researchers to maintain sustained exchanges and develop new collaborative research projects.
The agreement strengthens “Global Connect,” KAIST’s strategy for international cooperation. Through this initiative, KAIST seeks to move beyond conventional university exchanges by closely integrating research, education, and talent mobility. It also aims to expand its global research network in partnership with Korea’s other science and technology-focused universities.
“Global cooperation must move beyond conventional exchanges and evolve into partnerships that combine complementary strengths to create new knowledge and cultivate talent together. By connecting the research capabilities of Université Paris-Saclay and K-STAR, we expect to generate new research achievements in critical fields such as AI, quantum technology, and biotechnology, while further strengthening Korea’s global network in science and technology,” said KAIST President Choongsik Bae.
The agreement is also significant because it expands the educational and research cooperation that KAIST has developed with France over the past 15 years to the K-STAR level.
KAIST also signed a student exchange agreement with CentraleSupélec, a grande école affiliated with Université Paris-Saclay, in 2011. The two institutions subsequently established a dual-degree program in mechanical engineering in 2017 and have continued to develop their partnership. Building on the experience and mutual trust accumulated through this cooperation, the new agreement broadens the partnership to encompass K-STAR and Université Paris-Saclay.
Université Paris-Saclay President Camille Galap said, “Université Paris-Saclay and K-STAR have broad research capabilities spanning basic science and advanced technology, giving this partnership considerable potential. I hope the agreement will promote more active exchanges among students and researchers from both countries and enable us to achieve new scientific advances through joint research.”
Jaemin Jung, KAIST Senior Vice President for Planning and Budget, said, “This agreement marks the beginning of our efforts to translate the university-level cooperation discussed at the ninth Korea-France Joint Committee on Science and Technology in April into concrete research collaboration. KAIST will support exchanges among students and researchers so that they can work in one another’s laboratories and produce tangible outcomes, including joint research projects, publications, and the shared use of research infrastructure.”
The signing ceremony was also attended by Stéphane Nonnenmacher, Director of the Graduate School of Mathematics at Université Paris-Saclay; Divya Madhavan, Director of International Relations at CentraleSupélec; and a delegation from the Korea Institute for Advanced Study (KIAS), a KAIST-affiliated institute.
2026.09.09 View 2231 -
KAIST Identifies Cause of Artifacts in Battery Nanoscale Analysis, Paving the Way for More Reliable Measurements
A signal that appears to show ions moving inside a battery may, in fact, be an illusion caused by an uneven surface. A KAIST research team has identified the origin of this type of artifacts, which can lead researchers to misinterpret what is happening inside a battery, and has developed a method to reduce it. The findings are expected to enable more accurate analysis of ion movement and improve the reliability of next-generation battery-material development, including that of solid-state and sodium-ion batteries.
KAIST (President Choongsik Bae) announced on September 7 that a research team led by Professor Seungbum Hong from the Department of Materials Science and Engineering, in collaboration with the research groups of Professor Jong Min Yuk from the same department and Professor Nam-Soon Choi from the Department of Chemical and Biomolecular Engineering, has identified the cause of a measurement artifact in nanoscale battery analysis that can be mistaken for actual ion transport. The team also proposed a method for effectively reducing this artifact.
During charging and discharging, lithium or sodium ions move back and forth within a battery. The speed and ease with which these ions move affect the battery’s performance and lifespan. Developing better batteries therefore requires researchers to precisely determine where ions can move freely and where their movement is hindered.
One technique used for this type of analysis is Electrochemical Strain Microscopy (ESM), which is based on Atomic Force Microscopy (AFM). ESM scans the surface of a battery material with an extremely fine tip and measures nanoscale changes in the material associated with ion movement, allowing researchers to indirectly track ion transport.
The problem is that when the surface of a battery material is rough, similar signals can appear even in the absence of actual ion movement. If these signals are interpreted as evidence of ion transport, researchers may incorrectly identify where ions are moving within the material.
To investigate the origin of theseartifactss, the team created fine trenches on the surface of an ionically inactive single-crystal silicon sample. This provided an experimental environment in which no ions were moving, while the sample surface remained uneven.
The results quantitatively demonstrated that variations in surface height alone can alter the degree of contact between the microscope tip and the sample, producing signals similar to those generated by actual ion movement.
The same phenomenon was also observed in actual battery materials. When the team analyzed a graphite anode and the sodium solid electrolyte Na₂Zn₂TeO₆, the ESM signals likewise varied according to surface topography. This confirmed that the issue is not limited to a particular material but is a phenomenon that researchers must account for when conducting nanoscale analyses of a wide range of battery materials.
As a solution, the team proposed making the surfaces of battery materials as smooth and flat as possible. To achieve this, the researchers used a cooling cross-section polisher (CCP), which employs an argon (Ar) ion beam to precisely polish sample cross sections. Because argon is chemically inert under most conditions, this technique allows the surface to be processed precisely without significantly altering the properties of the sample.
This treatment substantially reduced surface roughness and, in turn, decreased measurement artifacts caused by uneven surfaces. The mechanism is comparable to a car moving up and down while traveling over a bumpy road: as the scanning tip passes over height variations on the surface of a battery material, the degree of contact between the tip and the sample changes. These changes can generate signals resembling those produced by actual ion movement.
In particular, the team examined signals detected at grain boundaries—the interfaces at which the small crystals that make up a battery material meet, much like the seams between adjacent tiles.
Before the surface was smoothed, strong ESM signals appeared at these grain boundaries. After the surface was polished, however, the enhanced signals disappeared. This finding indicates that some signals previously interpreted as evidence of “pathways that facilitate ion transport” may actually have resulted from variations in surface height rather than genuine ion movement.
This study is significant because it experimentally demonstrates how this type of measurement artifact arises in nanoscale battery analysis and shows that it can be reduced using the practical approach of smoothing battery-material surfaces.
The findings are expected to provide a more accurate understanding of where ions move freely and where their movement is hindered within a battery. Such insights could provide an important foundation for designing battery materials that facilitate ion transport, thereby enabling faster charging and longer battery life.
The team expects this analytical approach to be applicable not only to widely used lithium-ion batteries but also to next-generation battery systems. These include solid-state batteries, which use solid rather than liquid electrolytes, and sodium-ion batteries, which use sodium ions in place of lithium ions. The approach could help researchers more accurately understand how these batteries operate and support the design of new materials.
Furthermore, the accumulation of reliable nanoscale analysis data could provide high-quality training datasets for artificial intelligence (AI) and machine-learning research aimed at designing new battery materials and predicting their performance.
“This research clearly demonstrates how variations in surface height affect the results of nanoscale battery-material analysis,” said Professor Hong. “We expect our findings to enable more accurate tracking of ion movement within batteries and contribute to understanding the operating mechanisms of next-generation battery materials and designing improved materials.”
Dongyan Chen, a PhD student in the Department of Materials Science and Engineering, served as the first author of the study, which was published in Small Methods, an international journal specializing in materials science and nanotechnology.
Paper title: Quantitative Analysis of Topographic Crosstalk in DART-ESM Arising from Feedback-Loop-Delay-Induced Contact Stiffness Variations in Battery Materials
DOI: https://doi.org/10.1002/smtd.70763
This work was supported by National Research Foundation of Korea (NRF) grants funded by the Korean government’s Ministry of Science and ICT (MSIT) (Nos. RS-2026-25468150 and RS-2023-00247245).
2026.09.07 View 2904 -
KAIST Develops AI Technology That Fixes SQL Errors Without Starting Over
“Find the best-selling product from last year.” When an AI system attempts to answer a question like this by querying a company database, even a single reference to a nonexistent item can cause the query to fail. Until now, correcting such an error often required regenerating the entire SQL query from scratch. A KAIST research team has developed a technology that instead identifies and fixes only the erroneous part. The technology is expected to make AI-powered data retrieval faster and more accurate, accelerating the adoption of AI work assistants in enterprise environments.
KAIST (President Choongsik Bae) announced on September 4 that a research team led by Professor Min-Soo Kim from the School of Computing has developed SafeQL, a technology that detects and corrects errors that arise when natural-language questions are translated into Structured Query Language (SQL).
Text-to-SQL technology enables AI systems to convert everyday questions, such as “Which product saw the largest increase in sales last year?” or “Which items are running low in stock?”, into SQL queries. This allows users to retrieve sales, customer, and inventory data using natural language without having to understand complex database commands.
However, AI systems can make mistakes when generating SQL. For example, they may refer to a table or column that does not exist or join tables incorrectly. These errors can prevent the query from running and leave the AI unable to retrieve the requested data.
Conventional correction methods send the database error back to a large language model (LLM) and ask it to regenerate the entire query. This is similar to rewriting an entire report to correct a single word. In the process, parts that were already correct may be altered, new errors may be introduced, and repeated calls to the LLM increase both cost and processing time.
SafeQL takes a different approach. Instead of discarding and regenerating the entire query after an execution failure, it interprets feedback from the database management system to precisely locate the faulty component, such as a relation, attribute, function, or value. It then incrementally repairs that component while preserving the valid structure and logic of the original query.
To achieve this, the research team developed a “safe query space” approach. Among the candidate corrections that can be executed on the database, SafeQL searches for the one closest to the query originally generated by the AI. The system prioritizes the most promising candidates and filters out unsuitable ones in advance, reducing the time required for correction.
The team implemented SafeQL as a PostgreSQL extension and integrated it with the database system’s parser, binder, and type analyzer. This enables SafeQL to precisely locate errors even in complex SQL queries. The system calls the LLM again when search-based refinement cannot resolve the error within a predefined threshold, thereby reducing unnecessary AI use.
The research team evaluated SafeQL using BIRD and Spider, two widely used benchmarks for assessing the database querying capabilities of AI systems.
On the BIRD benchmark, SafeQL resolved execution errors in up to 87.4% of initially erroneous SQL queries and improved execution accuracy by up to 5.8 percentage points over the unrefined baseline. Compared with regenerating the entire query, SafeQL reduced token use by a factor of up to 15.1 and refinement latency by a factor of up to 29.6.
SafeQL is expected to be particularly useful for enterprises that handle large volumes of data requests. If an error occurs while an AI system is searching internal sales, customer, or inventory data, SafeQL can repair only the affected part instead of regenerating the entire query. This can reduce the cost and time required to operate enterprise AI systems and support reliable work automation powered by AI agents and corporate data.
Professor Min-Soo Kim said, “For AI to perform real-world tasks in enterprise environments, it must be able to accurately retrieve the data it needs.”
He added, “When AI makes an error during a database search, SafeQL fixes only the affected part instead of starting over from scratch. By reducing errors, costs, and processing time, we expect the technology to accelerate reliable AI-powered work automation.”
Geonho Lee, a PhD student in the KAIST School of Computing, participated in the study as first author, with Professor Min-Soo Kim serving as corresponding author. The findings will be presented at the International Conference on Very Large Data Bases (VLDB), a leading international database conference, to be held in Boston, USA, from September 1 to 5.
Paper title: SafeQL: Search-based Refinement for Safe and Efficient LLM-based Text-to-SQL
DOI: 10.14778/3819518.3819545
Authors: Geonho Lee (KAIST, first author) and Min-Soo Kim (KAIST, corresponding author)
This research was supported by the National Research Foundation of Korea (NRF) and the SW Star Lab program of the Institute of Information & Communications Technology Planning & Evaluation (IITP), both funded by the Ministry of Science and ICT (MSIT).
2026.09.04 View 2730