-
KAIST Develops ‘Chameleon AI Semiconductor’ with Programmable Response Speeds
AI semiconductors are becoming more programmable. KAIST researchers have developed a device whose response characteristics can be programmed to process data changing at different speeds. The technology reduced prediction errors for time-varying data by up to 40-fold and is expected to enhance real-time AI performance in autonomous vehicles, robots, and wearable devices.
KAIST (President Choongsik Bae) announced on August 7 that a research team led by Chair Professor Shinhyun Choi from the School of Electrical Engineering and the Graduate School of Semiconductor Technology has developed a programmable dynamic memtransistor (PDM), a semiconductor device whose time-response characteristics can be adjusted to multiple states and retained, as well as an integrated array based on the device.
A memtransistor is a next-generation semiconductor device that combines the information-storage function of memory with the computing function of a transistor. In the developed PDM, the ability to process data while retaining previous information allows its response characteristics to be adjusted and retained for incoming data.
Today’s computers and smartphones require complex software processing to analyze data that changes over time, resulting in large computational loads and high power consumption. To address this, researchers have been studying technologies that allow semiconductor hardware itself to process data directly. However, conventional devices have had fixed response speeds that cannot be changed once the device is fabricated.
The research team overcame this limitation by introducing a dual-layer structure inside the transistor, combining a charge storage layer that accumulates and processes data with an electron trapping layer that controls the response speed in a nonvolatile manner.
In the PDM developed by the research team, incoming data is processed in the charge storage layer, while the electron trapping layer controls, across multiple levels, the recovery speed at which the semiconductor returns to its original state. In experiments, the team succeeded in tuning the current recovery time over an approximately 5-fold range and the characteristic frequency over a range of more than 10-fold.
In particular, in experiments involving the prediction of data in which fast and slow changes are intricately mixed, the PDM reduced prediction errors by as much as 40 times compared with conventional fixed-response semiconductor devices. The PDM enables accurate information processing even when handwriting or object-movement speeds vary, by using response characteristics configured to match different input timescales. Once the response characteristics are set, the device remembers them without requiring a continuous external power supply, and it does not require complex preprocessing of input data. Because it is fully compatible with materials used in widely adopted commercial semiconductor processes, it is also highly advantageous for mass production and commercialization.
The research team fabricated a PDM array and used it to predict complex data, confirming that it achieved accuracy comparable to conventional software-based systems while consuming far less energy.
“This study demonstrates an AI semiconductor whose response characteristics can be programmed to efficiently process data changing at different speeds,” said Chair Professor Choi. “We expect it to become a core technology that improves the performance of AI devices such as autonomous vehicles, robots, and wearables while reducing their power consumption.”
This research was led by KAIST Graduate School of Semiconductor Technology Ph.D. candidate Dae-won Kim as the first author, with Yoonho Cho, Seokho Seo, Yujin Kim, See-On Park, Taehwan Jang, and Chaebin Park participating as co-authors. Young Taek Oh and Fellow Jae-Duk Lee of Samsung Electronics’ Semiconductor R&D Center also participated as co-authors, and Chair Professor Shinhyun Choi served as the corresponding author. The research was published in July in the internationally renowned journal Nature Communications on July 4.
Paper title: Programmable memtransistor array with temporal dynamics modulation for efficient time-series data processing,
DOI: https://doi.org/10.1038/s41467-026-75211-5
This research was supported by the National R&D Program through the National Research Foundation of Korea funded by the Ministry of Science and ICT, the ETRI R&D Support Program of the Institute of Information & Communications Technology Planning & Evaluation, the HRD Program for Industrial Innovation of the Korea Institute for Advancement of Technology funded by the Ministry of Trade, Industry and Energy, Samsung Electronics, and others.
2026.08.07 View 135 -
KAIST Reconstructs Transparent Structures Through Dynamic Scattering Layers in a Single Shot
A KAIST research team has developed a technology that reconstructs the shape, optical thickness, and position of a transparent object hidden between two dynamic scattering layers from a single shot. The technology could enable precision inspection of transparent semiconductor and display components, as well as biomedical imaging.
KAIST (President Choongsik Bae) announced on August 6 that a research team led by Professor Mooseok Jang from the Department of Bio and Brain Engineering has developed a single-shot phase imaging technique that reconstructs a phase object — a transparent object such as glass, plastic film, or a living cell, which produces almost no visible contrast under an ordinary camera but induces a subtle shift in light called a phase change — from a single measurement, even when the object is fully enclosed between two dynamic scattering layers.
Phase objects are difficult to see with conventional cameras because they show little brightness contrast with their surroundings. However, analyzing the minute phase shift can reveal an object's morphology and optical thickness, and can be used to determine its physical thickness or refractive-index variation when the other quantity is known. For this reason, phase imaging is widely used to observe living cells without staining and to inspect transparent components in semiconductors and displays.
The challenge is that when scattering layers positioned in front of and behind an object are in motion — much like the blurred view through a foggy window — the light path continually changes, making it difficult to obtain accurate information about the object. Conventional techniques have therefore required multiple exposures of the same target, prior calibration of the scattering environment, or training an AI model on large volumes of data.
To address this, the team tightly focused the illumination onto a small spot on the first scattering layer —much like concentrating light to a point with a magnifying glass— so that the light passing through it would carry the object's information as reliably as possible.
The researchers then combined an optical model, which computes how light changes as it passes through the object and scattering layers, with an AI framework. Rather than training on a large set of reference images as conventional AI approaches do, the framework works backward from physical laws to infer the path the light must have taken to produce the measured pattern.
The process is comparable to recovering a clear image from a single blurred photograph taken in fog. Using this approach, the team succeeded in simultaneously determining the shape and thickness of a transparent object, the scattering-induced blur characteristics, and the object's position — all from a single image measuring light intensity.
The technology is expected to have applications in a wide range of fields, including precision inspection for semiconductors and displays and biomedical imaging.
"This is the first demonstration of restoring the shape and position of a transparent object from a single measurement, even in environments where light is severely scattered, such as behind fog or a diffusive film," said Professor Jang. He added that the team plans to develop the technique further so that it operates reliably in more complex environments, with applications in semiconductor inspection and biomedical imaging.
The study was co-first-authored by Yoosun Kim, a master's student, and Gookho Song, a PhD candidate, both in the KAIST Department of Bio and Brain Engineering, with Professor Jang serving as corresponding author. The paper was published in the international optics journal Optica.
Paper title: Single-shot imaging of phase objects fully enclosed by dynamic scattering layers
DOI: https://doi.org/10.1364/OPTICA.593328
This research was supported by the National Research Foundation of Korea under the Ministry of Science and ICT (RS-2021-NR060086, RS-2023-00251628, RS-2026-25479811), and by a Samsung Electronics industry–academia strategic project (IO260313-15915-01).
2026.08.06 View 124 -
KAIST brings ‘giant batteries’ closer to commercialization in the AI data center era
The explosive growth of AI data centers has brought the commercialization of "giant batteries" one step closer. A KAIST research team has developed a process that cuts the production time for a core material used in large-capacity batteries by 67%, resolving the largest production bottleneck standing in the way of commercialization.
KAIST (President Choongsik Bae) announced on August 5 that a research team led by Professor Hee-Tak Kim from the Department of Chemical and Biomolecular Engineering has developed a process for producing the core electrolyte of vanadium redox flow batteries (VRFBs)—a leading candidate for large-capacity energy storage systems (ESS)—faster and more stably.
As AI data centers operate around the clock in growing numbers, large-capacity ESS that can store electricity generated from solar and wind power and supply it reliably when needed have become increasingly important.
Because VRFBs use nonflammable, water-based electrolytes, they have a lower fire risk than many conventional battery systems. And their energy-storage capacity can be scaled by increasing the amount of electrolyte stored in external tanks. This has drawn attention to VRFBs as ultra-large batteries suited to AI data centers and renewable energy storage. However, producing the vanadium electrolyte with an average oxidation state of 3.5+—the standard starting composition for VRFB operation— has been slow and costly, making it a critical obstacle to commercialization.
The conventional process first produces the electrolyte through chemical reduction—a reaction in which a chemical reducing agent causes vanadium ions to gain electrons—and then refines it through electrochemical reduction, which applies electric current to adjust the vanadium ions' electron state to the desired level. This final electrochemical step, however, relies on a costly VRFB stack and significant electrical energy, increasing both operational complexity and capital costs.Beyond the limitations of the electrochemical reduction process, the research team found, for the first time, that the alternative chemical reduction process also suffers from a distinct kinetic bottleneck. The reaction rate slows sharply at a specific point, much like highway traffic suddenly backing up at a bottleneck. This bottleneck occurs when the average vanadium oxidation state reaches approximately +4.1, an intermediate stage in the production of V3.5+ electrolyte.
In previous research, the team had replaced the conventional electrochemical adjustment step with a Pt/C-catalyzed reduction process, preventing the waste of leftover electrolyte. In the present study, it further extended the catalytic process into the bottleneck region of oxalic-acid-based chemical reduction. By switching from chemical to catalytic reduction at an average oxidation state of approximately +4.1, the team was able to bypass the slowest stage of the production process.
As a result, production time for V3.5+ electrolyte was cut by 67% compared to the conventional process. The switch also eliminated residual oxalic acid, an impurity that can degrade battery performance. The same catalyst was reused more than 2,500 times without a notable drop in performance, demonstrating the process's viability for industrial-scale production.
"This study combined reaction engineering principles with thermodynamic predictions to identify the rate-determining step in the chemical reduction and redesigned the electrolyte production process to overcome this major bottleneck to the commercialization of large-scale batteries," said Hee-Tak Kim, professor in the Department of Chemical and Biomolecular Engineering. He added, "By scientifically identifying the conditions under which the catalyst operates stably without degrading in the electrolyte environment, we resolved a production bottleneck relevant to industry, and we expect this to significantly accelerate the commercialization of large-capacity energy storage technology."
Kyunghwa Seok, a PhD candidate in the Department of Chemical and Biomolecular Engineering, led the research as first author. The findings were published online in Advanced Energy Materials—a leading international journal in the energy field—on May 7. In particular, in recognition of its academic significance, the study was selected as the cover article for Issue 34, which is scheduled to be published online in early September.
Paper title: Streamlined V3.5+ Electrolyte Production by Leveraging Chemical and Catalytic Reductions
DOI: https://doi.org/10.1002/aenm.71029
Authors: Kyunghwa Seok (KAIST, first author), Minseong Kang (KAIST, second author), and Hee-Tak Kim (KAIST, corresponding author).
This research was supported by Lotte Chemical.
2026.08.05 View 286 -
KAIST and Seoul National University Students Hold 100-Hour Robot Hackathon to Nurture Physical AI Talent
KAIST (President Choongsik Bae) announced on August 4 that RoboticUS, a joint student organization formed by students from KAIST and Seoul National University, is holding the inaugural Robot Hackathon at KAIST from August 3 to 8.
"In the era of Physical AI, we need convergence talent who can go beyond building good AI to design and implement robots and systems that move the real world based on AI," said Choongsik Bae, President of KAIST. "This hackathon, planned and run entirely by students, is a good example of KAIST's culture of challenge and collaboration, and we expect it to become a new educational model for turning future technologies into reality," he added.
The hackathon puts this educational philosophy directly into students' hands. It is Korea's first student-led Physical AI robot hackathon, planned and run by students from KAIST and Seoul National University across institutional boundaries. Participants experience the entire process of designing and building working robots, developing hands-on capabilities that integrate AI and hardware.
The event is hosted by RoboticUS, a nonprofit student organization formed jointly by MR, a robotics club in KAIST's Department of Mechanical Engineering, and Seoul National University's robotics clubs SHAPE and SIGMA. Students who share a passion for robotics from the two universities joined forces across institutional lines, handling every stage themselves — from recruiting participants to designing the mission, running the event, and setting up the presentation and judging format. KAIST's Department of Mechanical Engineering supports the event with facilities and operational assistance so that the students' initiative can translate into genuine educational value.
Ten teams — 30 students total — selected from the two universities will take part. After receiving training in power circuits and robot joint control on August 3 and 4, participants will begin building their robots when the mission is unveiled on the morning of August 5 and continue working until 4 p.m. on August 8. Starting from an idea, they will go through design, assembly, programming, and repeated testing to complete a working robot — experiencing the full roughly 100-hour cycle themselves.
Each team will be provided with Angel Robotics' "phact" actuator, which serves as the robot's joints and muscles, and NVIDIA's Jetson AGX, which functions as the robot's brain. Taejin Technology Co., Ltd will provide training on circuits and electronic components for supplying stable power to the robots, and Angel Robotics will support hands-on training in using the actuators and controlling the robots.
Participants will not simply assemble a finished kit — they will design the robot's shape and movement from the ground up and build it themselves.
For fairness, the mission will be revealed only at the start of the hackathon on August 5. There is no single correct answer or predetermined robot form. Each team will interpret the same mission differently, combining mechanical structure, circuitry, AI, and control software into a single robot. One of the highlights will be seeing the different solutions the ten teams develop in response to the same mission.
The final day, August 8, will be an open Physical AI festival that welcomes the general public. A public conference at the KI Building (E4) Fusion Hall will introduce the current state of Physical AI in an accessible and engaging way — from robotic skin that lets robots feel touch like humans, to humanoid robots that can see and hear people, to a quadrupedal robot that has completed a marathon.
Professors Jung Kim, Yong-Hwa Park, and Jemin Hwangbo of the Department of Mechanical Engineering, along with Joon-Ha Kim, CEO of Diden Robotics, will each give a talk on robotic skin and haptics, multimodal perception in humanoid robots, the quadrupedal robot Raibo, and the journey of developing Physical AI for industrial use, respectively.
After the conference, an open demo day for the ten participating teams will run from 4 to 5 p.m. at KAIST's Culture Complex (E9), 3rd floor. Members of the public will be able to visit each team's booth, watch the robots the students built over 100 hours in action, and submit their own evaluations via QR code.
Judging criteria include mission achievement, technical execution, creativity, and presentation and demonstration. The final score will weight faculty advisor evaluation at 30%, peer evaluation among hackathon participants at 30%, sponsor judging panel evaluation at 30%, and pre-registered public attendee evaluation at 10%.
Angel Robotics and Taejin Technology are taking part as core technology partners, providing equipment and training. Faculty members and industry experts are providing education and technical guidance so that students can safely handle equipment used in real research and industrial settings.
"Physical AI's competitiveness comes not just from software but from hardware and control technology," said Kyoungchul Kong, Professor from Mechanical Engineering at KAIST and Head of the Future Technology Institute at Angel Robotics. "This hackathon will show that with high-performance robot components and the right development environment in place, even undergraduates can turn their imagination into a working robot in 100 hours," he added.
"This hackathon is about learning and building together, rather than competing between schools," said Yeonsu An, President of RoboticUS (and President of MR, KAIST's robotics club). "We hope the general public will get to experience the robots students have built firsthand and take part in the judging, coming away with the sense that Physical AI is a technology anyone can understand and enjoy — not just something for experts," she added.
“Although we do not yet know what the challenge will be, I am most looking forward to gathering in one place and developing robots together,” said Hyeontae Jeon, a participating student from Seoul National University. “It will be even more meaningful to work through challenges across university boundaries and present the robots we built ourselves to the public.”
The public conference is open to anyone through pre-registration. Pre-registered attendees can watch the lectures, view the open demo day, and take part in on-site judging. Registration is available on the RoboticUS official website or on Event-us, under "First Robot Hackathon – Robot & AI Public Conference (KAIST × SNU)." Registration closes August 6.
2026.08.04 View 332 -
KAIST Develops Marine Carbon Removal Technology That Turns Carbon Dioxide in Seawater into “Stone” for Permanent Storage
A new pathway has opened to enhance the ocean’s natural ability to clean the planet. KAIST researchers have developed a technology that converts carbon dioxide dissolved in seawater into “stone,” or minerals, preventing it from returning to the atmosphere and enabling permanent storage. The achievement is expected to help the ocean absorb more carbon dioxide and accelerate the commercialization of next-generation marine carbon removal technologies.
KAIST (President Choongsik Bae) announced that a research team led by Professor Dong-Yeun Koh from the Department of Chemical and Biomolecular Engineering, in collaboration with Professor T. Alan Hatton’s group at the Massachusetts Institute of Technology (MIT), has developed an electrochemical dissolved ocean carbon removal (e-DOC) technology that converts carbon dioxide dissolved in seawater into calcium carbonate (CaCO₃), a stable mineral form, enabling virtually permanent carbon storage.
The ocean is the planet’s largest carbon reservoir, absorbing about 30% of the carbon dioxide emitted by human activity. Just as water naturally refills a large container when some is removed, removing carbon dioxide from seawater enables the ocean to absorb more carbon dioxide from the atmosphere.
The research team developed a technology that converts dissolved inorganic carbon (DIC), the carbon species dissolved in seawater, into a mineral form that does not return to the atmosphere. Once stored in this form, the carbon is effectively prevented from returning to the air, allowing the ocean to continue absorbing new carbon dioxide. Such technologies are gaining attention as key carbon dioxide removal (CDR) solutions for responding to climate change.
However, conventional technologies have faced a major challenge: mineral scaling. Much like limescale building up inside a kettle, minerals such as calcium carbonate adhere to electrode surfaces and clog the system. As operation continues, performance declines, requiring frequent cleaning or replacement of components and increasing both energy consumption and maintenance costs.
To overcome this issue, the research team developed a hollow fiber electrode assembly (HFEA), a device composed of bundled hollow, thread-like electrodes. In this structure, minerals form outside the electrode surface rather than directly on it, while hydrogen bubbles naturally generated during the reaction act like a brush, continuously cleaning the electrode surface and preventing mineral buildup.
In experiments using Jeju lava seawater, the team successfully operated the device continuously and stably for more than 120 hours. The system removed 80–90% of dissolved inorganic carbon from seawater and reduced electricity consumption by up to 54% compared with existing technologies. In addition, the process simultaneously produced high-purity hydrogen (H₂) and magnesium hydroxide (Mg(OH)₂), a material used in eco-friendly products and industrial applications, further improving its economic potential.
The newly developed device can be produced in a compact, modular form, making it suitable for installation on ships, offshore plants, and other marine industrial facilities. The research team expects the technology to be scaled up into large-scale marine carbon removal systems that can contribute to achieving carbon neutrality and responding to climate change.
Professor Dong-Yeun Koh said, “This technology converts carbon dioxide dissolved in seawater into a mineral form that does not return to the atmosphere, enabling permanent storage and helping the ocean continuously absorb new carbon dioxide,” adding, “We expect this work to accelerate the commercialization of marine carbon removal technologies and contribute to the realization of a carbon-neutral society.”
This study was co-led by KAIST Ph.D. candidate Inhwan Park of the Department of Chemical and Biomolecular Engineering and Dr. Young Hun Lee of MIT, who received his Ph.D. from KAIST in 2023 and is currently affiliated with the Department of Chemical Engineering at MIT, as co-first authors. The paper was published online on June 19, 2026, in the international journal Advanced Energy Materials.
Paper title: A Compact Hollow Fiber Electrode Assembly Architecture for Continuous Electrochemical Marine Carbon Dioxide Removal
DOI: https://doi.org/10.1002/aenm.71205
This research was supported by Hyundai Motor Company and Kia, as well as the Global C.L.E.A.N. Program of the National Research Foundation of Korea funded by the Ministry of Science and ICT.
2026.08.04 View 397 -
KAIST Develops AI That Generates Feasible Plans for Delivery, Production, and Workforce Scheduling
From parcel delivery routes and factory production schedules to hospital duty rosters, many real-world planning tasks require solutions that satisfy numerous operational constraints. KAIST researchers have developed an artificial intelligence technique that can independently generate feasible plans satisfying all constraints specified in a mathematical optimization problem.
KAIST (President Choongsik Bae) announced on August 3 that a research team led by Professor Min-Soo Kim from the School of Computing has developed RL-SPH (Reinforcement Learning-based Start Primal Heuristic), a reinforcement learning technique that trains AI to independently produce feasible plans without relying on an external solver.
The key feature of the technology is its ability to learn how to produce solutions that satisfy the multiple constraints encoded in an optimization problem. The research team expects the method to serve as an important foundation for AI-based decision-making in fields including logistics, manufacturing, semiconductor production, and workforce management.
Parcel delivery routing, vehicle routing, factory production scheduling, and hospital staff rostering are representative planning problems that can be formulated using integer linear programming, or ILP. ILP is a mathematical optimization technique for finding the most efficient solution while satisfying a set of linear constraints and requiring some or all decision variables to take integer values.
A parcel delivery plan, for example, must do more than simply minimize delivery time. It must also comply with vehicle capacity limits and driver working-hour requirements while ensuring that every destination is visited. A route that violates even one of these conditions cannot be used in practice, regardless of how short or inexpensive it may appear.
Existing learning-based approaches can rapidly generate approximate or partial solutions, but these predictions frequently violate constraints. Consequently, many approaches pass their outputs to specialized ILP solvers, such as Gurobi or SCIP, which are then responsible for obtaining a feasible solution. The paper notes that existing end-to-end learning-based primal heuristics generally struggle to generate feasible solutions independently.
RL-SPH addresses this limitation by iteratively revising a candidate solution rather than attempting to predict the final answer in a single step. At each stage, it selects multiple decision variables that are likely to improve feasibility and determines whether their values should be increased, decreased, or left unchanged. The model then learns from the resulting changes in constraint violations and solution quality.
Notably, the team designed the AI to first find a plan that is actually usable, rather than the single best plan. The overall procedure consists of two stages. In the first stage, the AI prioritizes finding an initial feasible solution that satisfies all constraints. In the second stage, it seeks a higher-quality solution by reducing the objective value, such as cost or processing time, while maintaining feasibility.
For example, in a factory production-planning problem, the method would first identify a schedule that satisfies requirements such as delivery deadlines, equipment capacity, and available labor. It would then attempt to reduce production cost and time without violating those conditions. The research therefore prioritizes finding a plan that can actually be implemented before attempting to optimize it further.
The team also introduced ILP-GT, a new AI model that learns the relationships between variables and constraints, along with a feasibility-aware search strategy that prioritizes revising the variables most effective for resolving the problem, substantially improving computational efficiency.
Across five representative benchmarks, RL-SPH achieved a 100% feasibility rate, successfully finding a usable plan for every problem. It maintained the same performance even on more complex problems involving general (non-binary) integer variables.
Compared with existing techniques, RL-SPH reduced the primal gap — the gap between a method's solution and the best-known solution — by an average of 28.6 times, and improved the primal integral — a measure of the speed and quality of the search process — by 2.6 times. The time needed to find the first feasible plan was also 2.5 times faster on average.
Among recent AI techniques such as PAS, DDIM, and DiffILO, RL-SPH was the only method to achieve a 100% feasibility rate across three benchmarks compared (SC, CA, IS). Its training also took an average of just 30 minutes — 14.7 times faster than existing techniques and roughly 34 times faster than the most recent unsupervised learning — an AI training method that finds patterns in data without being given the correct answers in advance — based technique.
The technique further demonstrated its generalization potential on MIPLIB, an international benchmark library for mixed-integer programming widely used in academia and industry. It reliably found feasible plans not only for problems up to 67 times larger than those it was trained on, but also for entirely new problem types it had never encountered during training.
“In real-world applications, a plan that can actually be implemented is often more important than a theoretically optimal answer that violates practical constraints,” said Professor Kim.
He added, “This research demonstrates that AI can learn to generate feasible solutions without relying on a specialized optimization solver to enforce feasibility. We expect the technology to provide an important foundation for AI-based decision-making in logistics, manufacturing, semiconductor production, workforce management, and other industrial fields.”
Tae-Hoon Lee, a doctoral student in the KAIST School of Computing, participated as the first author, and Professor Min-Soo Kim led the research.
The findings were presented at the 43rd International Conference on Machine Learning, or ICML 2026, held in Seoul from July 6 to 11. ICML is regarded as one of the world’s premier international conferences in machine learning.
Paper title: RL-SPH: Learning to Achieve Feasible Solutions for Integer Linear Programs
DOI: https://doi.org/10.48550/arXiv.2411.19517
Authors: Tae-Hoon Lee (KAIST, first author), Min-Soo Kim (KAIST, corresponding author)
This research was supported by the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation through related software research and Information Technology Research Center programs, as well as by the National Research Foundation of Korea. The paper’s acknowledgements specifically identify NRF and IITP support, including an ITRC grant.
2026.08.03 View 374 -
KAIST Develops a Way to Combat Cancer Cachexia, the Wasting Syndrome That Debilitates Patients
A new path has opened toward stopping cancer-associated cachexia, a devastating complication that gradually wastes patients away. A KAIST research team has developed an RNA-based therapeutic strategy that blocks a brain signal to prevent muscle loss and extend survival.
KAIST (President Choongsik Bae) announced on the 2nd of August that a joint research team led by Professor Minho Shong and Professor Jinkuk Kim from the Graduate School of Medical Science and Engineering, together with the KAIST faculty startup THOR Therapeutics (CEO Minho Shong), identified a new therapeutic principle for cancer cachexia.
Cancer cachexia affects 50 to 80 percent of all cancer patients, making it one of the most common complications of the disease. As cancer cells disrupt the body's metabolism, patients continue to lose weight and muscle mass even when eating sufficiently, leading to severe physical decline. This is a major reason chemotherapy often becomes less effective and treatment is discontinued, ultimately lowering survival rates. Existing drugs, however, have been limited to temporarily boosting appetite and have failed to fundamentally address the underlying muscle loss and metabolic dysfunction.
The research team focused on the idea that the root cause of cancer cachexia lies not in the body, but in the brain. The team noted that when GDF15 (Growth Differentiation Factor 15)—a signaling protein secreted in large amounts as cancer progresses—binds to GFRAL, a receptor protein in the brainstem, it triggers a signal instructing the body to stop eating and instead break down stored muscle and fat, driving progressive physical decline.
To block this process, the team worked with Professor Kim's group to develop a therapeutic using antisense oligonucleotides (ASO)—an RNA-based gene therapy technologies that selectively suppresses the activity of a specific gene—to prevent GFRAL from being produced in the first place.
In effect, the treatment switches off the receiver of the signal that cancer cells send instructing the body to waste away. By blocking GFRAL production at the RNA stage—the intermediate step in which genetic information is converted into protein—the therapy shuts down the cachexia-inducing signal at its source.
The team administered the treatment to mice in which cancer cachexia had already progressed. The result was a substantial reduction in muscle and fat loss, along with the restoration of the metabolic function that had previously broken down. Notably, even though treatment began after the disease had advanced significantly, survival at the study's endpoint (around day 50) was markedly higher in the treated group—90 percent—compared with just 20 percent in the untreated group, demonstrating the therapy's potential for treating cancer cachexia.
Unlike existing treatments that only stimulate appetite, this study is significant in that it blocks the underlying signal driving the wasting process itself. The therapy improved both muscle loss and metabolic dysfunction, and researchers expect that it could eventually be used alongside existing cancer treatments as a next-generation adjuvant therapy to improve patients' quality of life, treatment effectiveness, and survival rates.
"While existing therapies have only temporarily boosted appetite, this study is significant in that it directly targeted a key receptor in the brainstem at the RNA level to suppress the root cause of cancer cachexia," said Professor Song. He added that the team's goal is to move forward with follow-up preclinical research and drug manufacturing as well as quality-control systems without delay, begin clinical development in cancer patients by 2030, and develop the therapy into a treatment that improves patients' quality of life and survival rates.
Dr. Hyunjung Hong from the Graduate School of Medical Science and Engineering, Dr. Minhee Lee from THOR Therapeutics, and Dr. Minsung Park from the Graduate School of Medical Science and Engineering participated as co-first authors, with Professor Song and Professor Kim serving as co-corresponding authors. The findings were published in the international journal Cell Reports Medicine on July 27.
Paper title: Therapeutic Gfral silencing via antisense oligonucleotides ameliorates cancer-associated cachexia and extends survival in tumor-bearing mice
DOI : https://doi.org/10.1016/j.xcrm.2026.102939
This research was supported by the National Research Foundation of Korea, the Korea Health Industry Development Institute, and the Ministry of SMEs and Startups.
2026.08.03 View 287 -
KAIST Develops AI That Avoids Hallucinating Even at Night or in Smoke
Multimodal large language models (MLLMs), which process multiple types of sensory information such as text, images, and audio at the same time, are rapidly expanding the range of applications for artificial intelligence (AI). However, in real-world environments, these models can misinterpret the physical characteristics of sensors, mistakenly identify objects, or claim to hear sounds that are not actually present simply because a certain object appears in a video. These errors are known as hallucinations. A KAIST research team has developed a new technology that corrects such information confusion and physical misperceptions in AI.
KAIST (President Choongsik Bae) announced on the 31st of July that a research team led by Professor Yong Man Ro from the School of Electrical Engineering has developed two core technologies that overcome the tendency of existing large language models to rely too heavily on ordinary camera (RGB) images and enable AI to suppress cross-modal hallucinations that occur when different sensory inputs become mixed.
The first technology developed by the research team is the Diverse Negative Attributes (DNA) optimization method, which helps AI accurately understand the physical characteristics of special camera sensors such as thermal, depth, and X-ray sensors. Existing AI models often failed to understand the physical meaning of such images, for example by mistaking bright areas in thermal images for simple light reflection.
The research team built VS-TDX, the first comprehensive benchmark for evaluating diverse vision sensors, and used the types of wrong answers that AI frequently produces as learning signals to help the model internalize the characteristics of each sensor. As a result, the AI gained a “new eye” that allows it to accurately infer the state of objects even in darkness or smoke.
The second technology is Modality-Adaptive Decoding (MAD), a control method that blocks hallucinations caused by confusion between visual and auditory information at the source. This technology prevents AI from mistakenly claiming that it hears a sound that does not actually exist simply because a certain object appears in a video.
MAD works by having the AI self-assess whether vision or audio is more important for a given task, and then increasing the weight of the more relevant modality in real time. A key advantage of this technology is that it can immediately suppress hallucination errors without costly model retraining, as it is training-free.
Instead of retraining AI models at large scale with massive computing resources, the research team maximized cost efficiency by introducing the DNA method, which enables fine adjustment with only a small amount of data, and the
MAD plug-in approach, which requires no additional training at all.
These technologies can be applied to autonomous vehicles operating at night or in bad weather, robots performing missions in smoke-filled environments, and unmanned aerial vehicles using thermal cameras. They are also expected to be useful in fields that process multiple types of sensor information together, such as airport X-ray security screening and medical image analysis.
Professor Yong Man Ro said, “This research is significant because it reduces AI’s sensory bias and misperceptions without large-scale retraining,” adding, “It will serve as a foundation for building multimodal AI that can be trusted in real-life and industrial settings.”
This achievement was notable for its continuity, with Sangyun Chung, a doctoral student in KAIST’s School of Electrical Engineering, participating as first author in both studies. Dr. Youngjun Yoo also participated as co-first author in the DNA study.
Among the related papers, the MAD study was presented in June at the Conference on Computer Vision and Pattern Recognition (CVPR), the world’s leading international conference in AI and computer vision. The DNA study was published in IEEE Transactions on Image Processing, a leading international journal in the field of image processing.
Paper title: Enhanced Vision-Language Models for Diverse Sensor Understanding: Cost-Efficient Optimization and Benchmarking,
DOI: 10.48550/arXiv.2412.20750 Author information: Sangyun Chung (KAIST, co-first author), Youngjun Yoo (KAIST, co-first author), Se Yeon Kim (KAIST, third author), Youngchae Chee (KAIST, fourth author), Yong Man Ro (KAIST, corresponding author)
Paper title: MAD: Modality-Adaptive Decoding for Mitigating Cross-Modal Hallucinations in Multimodal Large Language Models,
DOI: 10.48550/arXiv.2601.21181
Author information: Sangyun Chung (KAIST, first author), Se Yeon Kim (KAIST, second author), Youngchae Chee (KAIST, third author), Yong Man Ro (KAIST, corresponding author)
Related demo video: https://youtu.be/VuP9i6Vfk8o
This research was supported by the Institute of Information & Communications Technology Planning & Evaluation’s (IITP’s) Human-Centered AI Core Technology Development Program and by a Center for Applied Research in Artificial Intelligence (CARAI) grant funded by the Defense Acquisition Program Administration (DAPA) and the Agency for Defense Development (ADD).
2026.08.03 View 316 -
KAIST Develops AI That Finds Its Own Hidden Weaknesses, Paving the Way for Safer Generative AI Models
KAIST researchers have developed a safety verification technology that uncovers roughly seven times more hidden vulnerabilities in AI than existing methods. The technology is expected to serve as a foundation for developing safer, more trustworthy AI.
KAIST (President Choongsik Bae) announced on the 30th of July that a research team led by Professor Junmo Kim from the School of Electrical Engineering has developed a new framework called Stable-GFlowNet (S-GFN), which overcomes the limitations of red-teaming—a safety verification process that deliberately attacks large language models (LLMs) to expose hidden weaknesses.
Red-teaming for generative AI is the process of crafting attack prompts designed to probe an AI's vulnerabilities and induce the AI to produce harmful or dangerous responses before the program is deployed. Since discovering a wider variety of attack methods allows more vulnerabilities to be addressed in advance, both the success rate and diversity of attacks are critical.
Previous approaches primarily relied on reinforcement learning—an AI technique trained to maximize reward—to generate attack prompts. However, these methods frequently suffered from mode collapse—a phenomenon where the model repeatedly converges on a narrow set of high-reward attack prompts rather than generating diverse outputs, thereby limiting its ability to uncover various vulnerabilities.
Generative Flow Networks (GFlowNets)—an AI generation technique trained to produce diverse outputs in proportion to their reward—were proposed as a solution. Yet GFlowNet training is computationally complex and unstable, and noisy reward signals can assign high rewards even to meaningless sentences, often causing training to collapse.
To address these issues, the research team developed three core techniques that help the model learn effective attacks more reliably while filtering out flawed ones.
First, much like comparing several paths to choose the best one, the team introduced Contrastive Trajectory Balance (CTB), which reduces computational complexity and stabilizes training by directly comparing pairs of generated attack trajectories.
Second, akin to filtering out background noise to focus on a single voice, the team applied Noise Gradient Pruning (NGP) to eliminate minor reward fluctuations and ensure the model learns exclusively from meaningful signals.
Third, the team applied the Min-K Fluency Stabilizer (MKS), which guides the model to generate attack prompts resembling text that a real user would write—just as a human reader naturally prefers coherent sentences to gibberish.
As a result, Stable-GFlowNet discovered 134 unique attack types—about seven times more than the 17 unique attack types found by the existing GFlowNet-based technique—while maintaining a high attack success rate of 92%.
Defense models trained using attacks generated by Stable-GFlowNet also demonstrated strong generalization, effectively defending against a wide range of attacks in cross-attack tests, which evaluate performance using attack techniques different from those used during training.
The team further demonstrated that CTB and NGP achieve faster and more stable performance than existing methods—not only in AI safety verification, but also in other distribution-matching tasks such as molecular generation for drug discovery.
Professor Kim said, "This technology is significant in that it can reliably uncover a wide range of AI vulnerabilities even in realistic conditions with limited data and high noise." He added, "Because it allows a broader range of risk factors to be identified and defended against before generative AI is deployed in real-world services, we expect it to become a core foundational technology for developing safer, more trustworthy AI."
The study was led by first author Minchan Kwon, a Ph.D. candidate from the School of Electrical Engineering, and was selected as a Spotlight paper—placing it in the top 2.2% of submissions—at the International Conference on Machine Learning (ICML) 2026, one of the world's most prestigious AI conferences.
※ Paper title: Stable-GFlowNet: Toward Diverse and Robust LLM Red-Teaming via Contrastive Trajectory Balance
arXiv: https://arxiv.org/abs/2605.00553
This research was supported by the Institute of Information & Communications Technology Planning & Evaluation’s (IITP) SW Star Lab program, funded by the Ministry of Science and ICT.
2026.07.30 View 654 -
KAIST Finds Algal Blooms Make Plastic More Prone to Breaking Apart, Revealing a Mechanism of Microplastic Formation
Every summer, algal blooms turn rivers and lakes green. Although they are widely known as a major form of water pollution that makes the water murky, a KAIST research team has now shown for the first time that algal bloom conditions can make discarded plastics, such as plastic bags, more prone to breaking apart, potentially accelerating the formation of microplastics. The study points to a new direction for the era of climate change: water pollution and plastic pollution need to be managed together, rather than as separate problems.
KAIST (President Choongsik Bae) announced on July 29 that a research team led by Professor Jaewook Myung from the Department of Civil and Department of Civil and Environmental Engineering has found, through a microcosm experiment using water collected from Duck Pond, a pond on the KAIST campus, that algal blooms alter the microbial ecosystem on the surface of low-density polyethylene (LDPE) — a common plastic used in plastic bags — and accelerate its early-stage weathering, in which the surface oxidizes and develops microscopic cracks.
The study is significant because it suggests that, in natural environments, plastic pollution and the eutrophication that drives algal blooms can interact and lead to new ecological changes.
Over time, plastic debris discarded in rivers and lakes becomes colonized by a wide variety of microorganisms, creating a small ecosystem of its own on the plastic surface. This ecosystem is known as the “plastisphere.” The plastisphere is known to influence the spread of pathogens and the transport of microplastics, but little has been known about how algal blooms, a serious form of water pollution, affect this microbial ecosystem.
To investigate this question, the research team constructed microcosms — small-scale experimental systems that recreate natural environments in the laboratory — in which algal blooms were artificially induced by controlling light exposure and nutrient concentrations. Over the following six weeks, the researchers closely analyzed the biofilms forming on the plastic surface, the succession of microbial communities, and changes in their functional gene profiles.
The analysis showed that under eutrophic conditions in which excessive nutrients trigger algal blooms that cyanobacteria, a major group of photosynthetic bacteria, proliferated alongside a variety of other bacteria, forming a thicker biofilm on the plastic surface. Microorganisms capable of producing large amounts of extracellular polymeric substances (EPS) also became significantly more abundant. EPS is a sticky, mucilage-like material that binds microorganisms together and helps them adhere to plastic surfaces.
As the microbial ecosystem on the plastic surface changed, the early weathering of the plastic — including surface oxidation and the formation of microscopic cracks — also accelerated. The team found that microorganisms harboring genes encoding enzymes associated with plastic oxidation and early-stage degradation became more abundant under eutrophic conditions.
Analyses using Fourier-transform infrared spectroscopy (FT-IR) and scanning electron microscopy (SEM) directly confirmed these changes. Oxygen-containing functional groups associated with oxidation, including carbonyl and hydroxyl groups, increased on the plastic surface, while more fine, hairline cracks appeared. These changes indicate that the plastic had become more susceptible to further physical weathering and fragmentation. The results suggest that these changes were driven not by a single microbial species, but by the combined activity of a microbial ecosystem comprising photosynthetic and other bacteria.
Professor Myung said, “As algal blooms become more frequent because of climate change, we expect this work to provide an important scientific basis for integrated environmental management strategies that consider water quality management and plastic waste management together.”
The study was led by first author Youngju Kim, a doctoral student from the Department of Civil and Environmental Engineering, and was published online in the international environmental journal Water Research on May 25, 2026.
Paper title: Eutrophication drives taxonomic and functional trajectories in plastic-associated biofilms
DOI: 10.1016/j.watres.2026.126183
Author information: Youngju Kim, KAIST, first author; Yijin Wang, HKUST; Wenqian Xu, HKUST; Charmaine C.M. Yung, HKUST; and Jaewook Myung, KAIST, corresponding author — five authors in total
This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (RS- 2023–00209472, RS-2024–00437656, and RS-2026–25393325), by the Ministry of Oceans and Fisheries, Korea (20200104), and by the grant for the “KAIST Grand Challenge 30 Program” funded by the Korea Advanced Institute of Science and Technology (N11250072)
2026.07.29 View 556 -
KAIST Professor Sooel Son Selected for Microsoft Funding for AI Safety and Security Research
KAIST (President Choongsik Bae) announced on the 28th of July that Professor Sooel Son has been selected as the only researcher in Korea to receive funding from Microsoft for research on artificial intelligence safety and security.
The funding was awarded through the External Red Team Alliance (EXTRA), a new program established by Microsoft’s AI Red Team, which examines the safety and security vulnerabilities of AI systems. EXTRA is a global initiative designed to strengthen AI safety and security research capabilities by supporting researchers at universities and technology experts around the world.
Microsoft noted that most AI safety testing is still conducted internally by individual companies or organizations. However, assessing the major risks posed by increasingly advanced AI systems requires expertise across a broad range of fields, including cybersecurity, multilingual environments, regional and cultural contexts, AI alignment, and potential misuse. EXTRA was launched in recognition of the difficulty a single internal organization faces when comprehensively evaluating these diverse risks.
Through the program, Microsoft will provide KAIST with an unrestricted gift of USD 25,000, approximately KRW 37 million, with no prescribed project period, to support research related to AI safety, security, alignment, and responsible AI development. The funding will be used to support Professor Son’s research team in its work on AI security and safety.
More than a dozen universities across six continents are participating in EXTRA, with KAIST being the only university selected from Korea. Through the program, Microsoft plans to expand the ecosystem for independent AI safety research and strengthen collaboration between academia and industry.
Professor Son’s research team has been conducting research on the security and privacy of AI systems that use machine-learning models and large language models. In particular, the team analyzes adversarial attacks against deep neural networks and language models—including model extraction, membership inference, personal information extraction, model inversion, machine unlearning, and prompt injection—and develops defense methodologies to assess and improve model safety.
Building on these technologies, the team is also focusing on establishing systematic defense methodologies that enable the safe and trustworthy deployment of agentic AI systems operating in real-world service environments, including web agents and agentic browsers.
“As AI systems rapidly spread throughout society, research that verifies the security and reliability of increasingly advanced AI is becoming more important,” said Professor Son. “Our participation in Microsoft AI Red Team’s EXTRA program will provide an opportunity to further advance our research on the safe development and use of AI systems.”
“AI safety research has never been more important, and universities have a critical role to play in advancing the field,” said Ram Shankar Siva Kumar, who leads the Microsoft AI Red Team. “Through EXTRA, we aim to support researchers working to deepen our understanding of how increasingly powerful AI systems can be evaluated, protected, and governed responsibly.”
“Competition in AI technology is expanding beyond performance to encompass safety and trustworthiness,” said KAIST President Choongsik Bae. “KAIST’s participation as the only Korean research institution in Microsoft’s global AI safety research network is a meaningful achievement that demonstrates Korea’s competitiveness in AI research. We will continue to lead the advancement of responsible AI technologies that everyone can trust and use by pursuing world-class research in AI safety and security.”
2026.07.28 View 1029 -
KAIST Makes Cancer Cells Send Out Their Own Danger Signal — Delivering Immunotherapy and Gene Therapy in a Single Nanoparticle Platform
Cancer cells survive by hiding from the immune system's surveillance. A KAIST research team has developed a new anticancer platform that makes cancer cells send out their own danger signal—prompting immune cells to attack—while simultaneously delivering gene therapy. The approach is expected to offer a new treatment strategy that combines cancer immunotherapy and gene therapy in a single nanoparticle.
Immunogenic cell death (ICD) is a process in which dying cancer cells send danger signals to nearby immune cells, prompting them to attack. A polypeptide is a polymer made of a long chain of amino acids.
KAIST (President Choongsik Bae) announced on 28th of July that a team led by Professor Yeu-Chun Kim from the Department of Chemical and Biomolecular Engineering has developed a "helical polypeptide nanoparticle" platform that induces severe stress inside cancer cells to trigger immunogenic cell death, while also delivering a range of gene therapeutics into the cells.
The body's immune cells effectively eliminate external invaders such as viruses and bacteria, but cancer cells evade immune surveillance through a variety of immune-escape strategies. This failure of immune cells to recognize cancer cells as a threat has long been one of the biggest limitations in cancer treatment.
Recently, researchers have been actively exploring the use of nanoparticles to deliver drugs and genes to cancer cells and activate immune responses. However, it has not been clearly established which properties of nanomaterials actually induce cellular stress and activate antitumor immune responses.
By comparing and analyzing a range of nanoparticles, the team confirmed that not just the chemical composition, but the helical, coiled shape of the nanomaterial is a key factor determining therapeutic efficacy. In particular, when a positively charged quaternary amine—a chemical structure that binds readily to cell membranes—was combined with a helical structure, the particle could penetrate the cell membrane like a screw and enter cancer cells with ease. By contrast, particles with the same chemical composition but lacking the helical coil barely entered cells at all and failed to induce an immune response.
The helical nanoparticles developed by the team preferentially seek out and penetrate cancer cells, which have different membrane electrical properties from normal cells. Once inside, the particles disrupt the membranes of mitochondria—the cell's energy-producing organelles—and other organelles, subjecting the cancer cell to severe stress.
Under this extreme stress, the dying cancer cell releases damage-associated molecular patterns (DAMPs)—distress signals indicating "a dangerous cell is here"—into the surrounding environment. Immune cells that detect these signals recognize the previously hidden cancer cell as a threat and begin their attack. In effect, the cancer cell is made to broadcast its own location to the immune system.
The nanoparticle does more than trigger an immune response—it also functions as a carrier for gene therapeutics. Messenger RNA (mRNA), which carries the genetic information for protein synthesis, and small interfering RNA (siRNA), which suppresses the expression of specific genes, are both typically difficult to deliver into cells. The team's nanoparticles, however, delivered these molecules effectively into the cytoplasm.
The researchers also introduced guanidinium, a chemical functional group that binds strongly to genetic material, at an optimized ratio, enabling the particles to remain stable in the bloodstream while delivering gene therapeutics effectively.
In mouse models of melanoma and colorectal cancer, the team loaded the helical nanoparticles with siRNA targeting PD-L1 (Programmed Death-Ligand 1), an immune-evasion protein, and administered them. Tumor growth was suppressed by approximately 70–80%, and a marked increase was observed in cytotoxic T cells—which directly attack cancer cells—infiltrating the tumor, indicating a substantial boost in antitumor immune response.
"This study presents a new anticancer platform in which the nanomaterial does more than simply deliver a therapeutic agent—it drives cancer cells to trigger their own immune response," said Professor Yeu-Chun Kim. He added that the platform is expected to contribute to the development of next-generation treatments combining cancer immunotherapy and gene therapy.
Dr. Susam Lee, the paper's first author, added, "We showed that it is not just the composition of the nanomaterial but the helical structure itself that is the key factor determining therapeutic efficacy." He said he hopes the findings will serve as a new benchmark for designing next-generation immuno-oncology nanomaterials.
The study was published online in Biomaterials, a leading international journal in the field of biomaterials, on May 28, 2026.
Paper title: Helical quaternary amine polypeptide programs membrane stress to drive immunogenic cell death and cytosolic gene delivery for cancer immunotherapy
DOI: 10.1016/j.biomaterials.2026.124337
This work was supported by the National Research Foundation of Korea (NRF) grants funded by the Korean government (MSIT), the Biomedical Global Talent Nurturing Program of the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (RS-2025-25459605 to Susam Lee) and the Korea Basic Science Institute (National research Facilities and Equipment Center).
2026.07.28 View 664