%47%45
-
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 216 -
KAIST Develops a Soft 3D-Printed Robotic Hand that Gently Grips Everything from Eggs to a 1 kg Water Bottle
3D printers that once could only produce rigid objects can now create products as soft and stretchable as rubber. A team of Korean researchers used AI to identify the optimal "recipe" for a material that can be printed into complex shapes while stretching to more than six times its original length. The material is expected to expand the range of applications for 3D printing, from robotic hands to form-fitting wearable devices and custom medical devices.
KAIST (President Choong-Sik Bae) announced on September 1 that a research team led by Professor Seungchul Lee from the Department of Mechanical Engineering, working with Dr. Jongbeom Na's team at the Korea Institute of Science and Technology’s (KIST, President Sang-Rok Oh) Extreme Materials Research Center and Professor Bumsoo Park from the Department of Manufacturing Systems and Design Engineering (MSDE) at Seoul National University of Science and Technology (SEOULTECH, President Dong-Hwan Kim), had used AI to develop a material that is both 3D-printable and highly stretchable, like rubber.
The need for such materials — soft, stretchable, and capable of forming complex shapes — has been growing as soft robots that come into direct contact with people, wearable devices worn on the body, and medical devices custom-fitted to patients have drawn increasing attention.
The 3D printing technology the team used, Digital Light Processing (DLP), cures a liquid material into a desired shape by exposing it to light. While DLP can quickly produce complex structures, making a material more stretchable and durable tends to raise its viscosity to the point that it no longer flows well enough to be printed. Conversely, thinning the material to make it easier to print reduces its stretchability and strength. Thus, developing a material that is both easy to print and highly stretchable was the central challenge.
The team used AI to identify the optimal "material recipe" that satisfies both conditions. Notably, the training data included not only materials that print well, but also highly viscous materials that are difficult to print.
The researchers cured various liquid material formulations in small molds and measured how stretchable and hard they were, how quickly they cured under light, and how well they flowed. This produced a dataset linking a wide range of material formulations to their respective properties.
The team then used machine learning to examine the relationship between material formulation and performance. Based on this, the AI identified the optimal material combination that is both 3D-printable and highly stretchable.
The material identified by the AI printed reliably on a DLP 3D printer and showed high stretchability, extending to more than six times its original length when pulled, without easily tearing.
To verify its real-world potential, the team 3D-printed a "soft actuator" using the material. A soft actuator is a device that uses air pressure and other means to create gentle, muscle-like movement. When inflated with air, it expanded like a balloon and bent as naturally as a human finger.
A soft robotic hand made by combining several actuators lifted a 1 kg water bottle and successfully and stably grasped objects of varying shapes and rigidity, from fragile eggs to glass bottles, an egg carton, and a computer mouse.
Beyond developing a single highly stretchable material, this research is significant for presenting an AI-based method for more quickly identifying materials with desired properties.
Previously, researchers had to directly formulate and test countless materials to find the optimal combination. Going forward, AI can first identify promising material combinations based on experimental data, which researchers then verify through testing, thereby reducing trial and error and shortening material development time.
"This research is significant as it shows that combining researchers' experimental data with artificial intelligence can efficiently identify optimal material combinations that were previously difficult to find," explained Professor Seungchul Lee. "We expect it to be used to more rapidly develop 3D-printing materials with the performance needed across a range of fields, including soft robots, wearable devices, and custom medical devices."
The study, with Dr. Younghan Song and Professor Bumsoo Park as co-first authors, was published in the international journal Nature Communications on June 4.
Paper title: Machine learning guided formulation design of digital light processing printable elastomers beyond viscosity stretchability tradeoff
DOI: https://doi.org/10.1038/s41467-026-73735-4
This research was supported by the Ministry of Trade, Industry and Resource's Machinery and Equipment Industry Technology Development Program (20023762), and by the Ministry of Science and ICT's Nano & Material Technology Development Program (RS-2026-25534767) and Excellent New Researcher Program (RS-2024-00350423).
2026.09.01 View 510 -
KAIST Uses Light to Distinguish Real from Fake with Hard-to-Replicate “Artificial Fingerprints”
Shine a light, and the real can be distinguished from the fake. KAIST researchers have developed a security technology that uses unique “artificial fingerprints” created by the random assembly of nanoparticles. Although extremely difficult to replicate, these fingerprints can be conveniently authenticated using only a smartphone flashlight and a laser pointer, opening up potential applications in anti-counterfeiting and electronic device authentication.
KAIST announced on August 26 that a research team led by Professor Sang Ouk Kim from the Department of Materials Science and Engineering, in collaboration with a team led by Professor Seok Joon Kwon of Sungkyunkwan University, has developed a new foundational security technology based on randomly assembled colloidal nanopatterns—unique microscopic patterns formed by particles too small to be seen with the naked eye. The technology enables authentication using everyday light sources such as smartphone flashlights and laser pointers.
Recent advances in artificial intelligence have made cyberattacks increasingly sophisticated, while future quantum computers may pose a threat to conventional cryptographic systems. As a result, growing attention is being paid to technologies that use the unique physical characteristics of products or devices themselves for security, in addition to software-based encryption.
A physical unclonable function, or PUF, is a security technology that uses minute physical differences naturally generated during the manufacturing process as security information. Just as every person has a unique fingerprint, microscopic particles form a different arrangement each time they assemble. Even when the same materials and process are used, reproducing the exact positions and orientations of the particles is extremely difficult. The researchers used these differences as “artificial fingerprints” for authenticating products and devices.
However, conventional high-security PUFs typically require expensive microscopes, spectroscopic equipment, or imaging systems to read information from their tiny and complex structures, making them difficult to use conveniently in everyday settings.
The research team focused on solving this dilemma between high security and easy authentication. By using the self-assembly of spherical particles hundreds of nanometers in size on a water surface, the team created unique structures composed of many small crystalline domains with different sizes and orientations. These structures are different every time they are made, making them difficult to replicate, while also producing clear optical signals when illuminated.
The research team implemented an authentication method in which the two patterns generated by each product are registered in advance and subsequently compared with those observed from the actual product. In other words, a single “nanofingerprint” is authenticated using two different light sources: a flashlight and a laser. Much like identifying a person using both their face and fingerprint, verifying one nanostructure in two independent ways strengthens security.
When illuminated with ordinary light, such as a smartphone flashlight, the nanostructure produces a unique color and reflection pattern depending on the particle arrangement. When illuminated with a laser pointer, the microscopic particle structure scatters the light in multiple directions, producing a second distinctive optical pattern.
To create a counterfeit, a forger would have to reproduce not only the nanoparticle structure itself, but also the exact color and reflection pattern produced under a flashlight and the optical pattern generated under laser illumination—making replication extremely difficult.
The researchers also successfully transferred the nanostructures onto a variety of surfaces, including flexible plastics, metals, transparent films, and hydrogels—soft, gel-like materials capable of retaining large amounts of water.
The technology could be used to assign a unique “hardware ID” to electronic products and Internet of Things devices for product authentication. It could also serve as an anti-counterfeiting label for luxury goods, artworks, and pharmaceuticals. Because it can be applied to transparent films, it may also be developed into security stickers that do not obscure a product’s design or appearance.
Professor Sang Ouk Kim of KAIST’s Department of Materials Science and Engineering said, “The key achievement of this study is that it combines randomly formed structures that are extremely difficult to replicate with a simple authentication method using readily available tools such as a flashlight or laser pointer.” He added, “We expect the technology to develop into a next-generation security solution that can be readily used in everyday applications, including electronic device authentication and anti-counterfeiting labels.”
Dr. Geon Gug Yang of KAIST’s Department of Materials Science and Engineering and Ph.D. student Seong-Gyun Im of Sungkyunkwan University’s Department of Chemical Engineering contributed equally as co-first authors. Professors Sang Ouk Kim of KAIST and Seok Joon Kwon of Sungkyunkwan University served as co-corresponding authors. The results were published online on July 23 in the international journal Nature Communications.
Paper title: “Dual-space visible light authentication toward high security physical unclonable function”
DOI: https://doi.org/10.1038/s41467-026-75781-4
This research was supported by the Mid-Career Researcher Program and the InnoCORE Program funded by the Ministry of Science and ICT, as well as by the Samsung Research Funding & Incubation Center for Future Technology.
2026.08.26 View 548 -
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 1675 -
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 1383 -
KAIST Develops a Molecular Platform for the Selective Control of Oxygen Reaction Pathways
Controlling how oxygen reacts is important for improving technologies such as batteries, fuel cells, and environmentally sustainable chemical processes. A KAIST research team has developed a new molecular system that can selectively switch the pathway through which electrons are transferred during oxygen activation. The findings are expected to provide a fundamental design principle for next-generation catalysts and energy-conversion technologies.
KAIST (President Choongsik Bae) announced on the 22nd of July that a research team led by Professor Seung Jun Hwang from the Department of Chemistry has developed a molecular system capable of directing oxygen activation along a selected electron-transfer pathway. By combining germanium with a molecular framework that can store and transfer electrons, the team established a design principle for selectively switching oxygen activation between two- and four-electron pathways.
Catalysts for controlling oxygen reactions have traditionally been developed around transition-metal centers such as iron, cobalt, and nickel. Germanium, by contrast, is a main-group element in the same group of the periodic table as silicon and has generally been considered less suitable for reactions requiring the coordinated transfer of several electrons.
To overcome this limitation, the research team combined germanium with a redox-active ligand, a molecular framework capable of storing, accepting, and transferring electrons. The ligand serves as an electron reservoir and cooperates with the germanium center, allowing the entire molecular structure to participate in multielectron reactions.
When oxygen reacts, the products and reaction outcomes depend on whether two or four electrons are transferred. In general, two-electron oxygen reduction produces hydrogen peroxide, while four-electron reduction produces water. Selectively controlling these pathways is therefore an important challenge in the development of batteries, fuel cells, and greener chemical catalysts.
The study presents a rare example of a main-group molecular system in which two- and four-electron reactivity can be selectively accessed within the same underlying molecular framework. This approach broadens the range of elements that may be considered in catalyst design and provides an alternative strategy to relying exclusively on transition metals.
The team also succeeded in isolating and analyzing a germanium compound representing the two-electron stage of the reaction, which they stabilized by attaching a methyl group to the germanium complex. Remarkably, the germanium atom in this compound could both donate and accept electrons, providing an important clue to how the system controls different reaction pathways.
The team also confirmed the practical potential of the new system. Under mild, light-free conditions, the germanium complex removed halogen atoms such as bromine and chlorine from organic compounds and regenerated alkenes (organic compounds containing a carbon-carbon double bond), which are widely used as raw materials for pharmaceuticals, plastics, and other chemical products. These results suggest that useful chemical feedstocks could be produced through simpler and potentially more energy-efficient processes.
“We expect these findings to inform the development of next-generation catalysts for energy conversion and to contribute to more selective and efficient chemical processes.” said Professor Hwang.
The study was conducted by Sung Gyu Kim and Jinrok Oh, currently postdoctoral researchers in the KAIST Department of Chemistry, and Dae Eui Choi, a student in the combined master’s and doctoral program in the Department of Chemistry at POSTECH. The results were published online in the international journal Chem on July 6.
Paper title: Germanium Ligand Redox Cooperativity: A Key to Ambiphilicity and Switchable Two- and Four-Electron Transfer
DOI: 10.1016/j.chempr.2026.103127
This work was supported by National Research Foundation of Korea grants funded by the Korean government through the Ministry of Science and ICT (NRF-2021R1C1C1010220 and RS-2025-02216980), and by the Samsung Science and Technology Foundation under Project No. SSTF-BA2101-09. Sung Gyu Kim received research fellowship support from the Basic Science Research Program through the National Research Foundation of Korea, funded by the Ministry of Education (RS-2024-00415390).
2026.07.22 View 1295 -
KAIST Identifies the “Hidden Energy Cost” of AI Agents for the First Time
As the era of AI agents—systems that can reason and act autonomously—begins, the power consumption of data centers is emerging as a critical challenge. A KAIST research team has, for the first time, analyzed the computational cost and energy consumption of AI agents, finding that they can consume up to 136.5 times energy per query than conventional generative AI. The study shows that competitiveness in the AI era is expanding beyond model performance to include the efficiency of data centers and power infrastructure.
KAIST announced that a research team led by Professor Minsoo Rhu of the School of Electrical Engineering has systematically analyzed, for the first time, how much computational resources and power AI agents require in real-world service environments.
Large language model (LLMs) powered applications such as ChatGPT have rapidly evolved beyond simply answering questions. They are now developing into AI agents: next-generation AI systems that can plan, use external tools such as web search, calculators, and code execution environments, and solve complex tasks by coordinating multiple steps on their own.
Although AI agents are increasingly being adopted in areas such as software development, research, and workplace automation, little has been known about the amount of electricity and operational cost required to run them in practice.
The research team defined AI agents not merely as software programs, but as a new type of workload that must be continuously processed by data-center servers and graphics processing units, or GPUs—high-performance chips used for large-scale AI computation. The team then analyzed the computational load and energy consumption incurred during actual AI agent execution.
The analysis found that AI agents perform, far higher volumes of LLM invocations than conventional chain-of-thought reasoning. Chain-of-thought, or CoT, refers to a method in which an AI model breaks down its reasoning process step by step to reach an answer, while an LLM invocation refers to each computational request made to a language model to generate a new judgment or response.
Because AI agents repeatedly call language models during execution, their response latency also increases significantly. The team found that response time can increase by up to 153.7 times, while GPUs remain idle for as much as 54.5 percent of the total execution time as external tools perform their tasks. In other words, as AI systems take on more complex tasks, a new form of inefficiency emerges in which expensive GPUs cannot be fully utilized.
The research team also analyzed the power consumption of AI agents at data-center scale. An AI agent using a 70-billion-parameter LLM—a scale comparable to current commercial AI services—consumed an average of 348.41 watt-hours per query. This is 136.5 times higher than the energy consumed by a conventional generative AI system performing simple question answering.
In addition, the team projected a future scenario in which 13.7 billion AI agent requests are generated per day — a volume equivalent to current Google search traffic. Under this scenario, data-center power demand would reach approximately 198.9 gigawatts, a level far exceeding the scale of AI data centers currently under development (which are in the range of a few gigawatts) and equivalent to roughly half of the average power consumption of the United States.
This study demonstrates that the focus of competition in the AI era is shifting from “smarter AI” to “optimally efficient AI.” Going forward, it will be essential not only to advance AI models, but also to jointly optimize AI semiconductors, data centers, and power infrastructure through co-design. Such an approach is expected to become a key strategy for reducing the operating cost of AI services and building sustainable AI infrastructure.
“This study is the first to quantitatively show not only how AI is becoming more intelligent, but also how much electricity and cost are required to implement and sustain that intelligence,” said Professor Rhu. “As AI agents become widespread, it will become increasingly important to take an integrated co-design approach that optimizes not only AI data-center infrastructure, but also AI agent models and power infrastructure.” He added, “Research and investment in this direction will be essential to dramatically reduce the cost for end users to access AI services while building sustainable AI infrastructure.”
The study was conducted with Jiin Kim, a Ph.D. student in the KAIST School of Electrical Engineering, as the first author. The paper was presented in February at the 32nd IEEE International Symposium on High-Performance Computer Architecture, or HPCA, one of the most prestigious international conferences in computer system design. The research team has also released the AI agent implementations and benchmarks used in the paper as open source to support follow-up studies by researchers worldwide.
Paper title: “The Cost of Dynamic Reasoning: Demystifying AI Agents and Test-Time Scaling from an AI Infrastructure Perspective”
Open-source repository: 10.1109/HPCA68181.2026.11408569
This research was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) through the SW Starlab program, the K-Cloud Technology Development Program using AI semiconductors, and the Leading Technology Development Program for Advancing AI-Semiconductor-Based Data Centers, as well as by the Samsung Electronics Future Technology Incubation Center.
2026.07.07 View 2022 -
KAIST Global Entrepreneurship Summer School Marks Fifth Consecutive Year of Cultivating Future Entrepreneurs in Silicon Valley
The 2026 Global Entrepreneurship Summer School (GESS), organized by the KAIST Office of Global Initiatives, has successfully concluded its fifth annual program.
Now in its fifth year, GESS has become KAIST's flagship global entrepreneurship program, providing students with firsthand experience in Silicon Valley—the world's leading startup ecosystem—and equipping them with the entrepreneurial mindset and global competencies needed to launch ventures on the international stage.
Participants in the 2026 GESS program, including both undergraduate and graduate students, were selected through a competitive process consisting of document screening, interviews, team presentations, and peer evaluations.
Prior to traveling to Silicon Valley, the selected students completed a four-month preparatory program that included team building, customer discovery, business model development, and pitch preparation. Throughout the program, they received mentoring from entrepreneurs, venture investors, and industry experts, enabling them to refine their business ideas and evaluate their potential for entering global markets.
The Silicon Valley program, held in late June, was organized in collaboration with leading startup support organizations, including KOTRA Silicon Valley, IBK Changgong Silicon Valley, and Plug and Play. Through meetings with entrepreneurs, venture capitalists, and representatives from global technology companies, students gained firsthand insight into the Silicon Valley startup ecosystem while developing a deeper understanding of global markets.
For the fourth consecutive year, students from the KAIST College of Business Impact MBA program also participated in the Silicon Valley program, creating valuable opportunities for interdisciplinary collaboration and exchange among students with diverse academic backgrounds and professional experiences.
A highlight of this year's program was a startup storytelling workshop conducted in collaboration with educators from Stanford University. The workshop helped students strengthen their communication skills by learning how to present their ideas more persuasively—an essential competency for aspiring global entrepreneurs.
In partnership with KAIST alumni based in Silicon Valley, participants also visited leading global technology companies and unicorn startups, including Meta, NVIDIA, and Moloco. They attended networking events with local professionals and alumni, gaining firsthand exposure to the innovation culture and growth strategies of global technology companies while broadening their perspectives on international careers and entrepreneurship.
To put into practice one of the core values of entrepreneurship—creating positive social impact—GESS participants also organized "Let's Play AI+Tech," a community outreach program for elementary school students from underserved families in Sunnyvale, California. Designed and led entirely by KAIST students, the program introduced fundamental concepts in artificial intelligence through engaging, hands-on activities for children and their parents. The initiative also offered KAIST students a meaningful opportunity to give back to the local community while sharing their expertise in AI and technology.
The program concluded with the Final Pitch Competition, where each team presented the business models they had developed over several months to a panel of Silicon Valley investors and entrepreneurship experts. Through expert feedback and evaluation, participants had the opportunity to validate the global potential of their ventures.
Following a highly competitive final round, Team CUPID was named the overall winner. Team CUPID presented an AI-powered developer platform that automatically routes coding tasks to the most cost-effective AI model, significantly reducing developers' AI usage costs. The team received high praise from the judges for its clear problem definition, strong market potential, and scalability in the global market.
Gianidita Nurani Pertiwi, a member of Team CUPID and a student in the Department of Bio and Brain Engineering, said, "GESS provided an invaluable opportunity to experience Silicon Valley's entrepreneurial ecosystem firsthand. Through conversations with founders, investors, and industry experts, I learned how to refine our ideas from a global perspective. The experience has motivated me to continue pursuing innovation that can create meaningful impact beyond borders."
The 2026 GESS program has been organized for the fifth consecutive year by the Office of Global Initiative in collaboration with the Impact MBA program and the Startup KAIST. KAIST will continue strengthening partnerships with Silicon Valley and other global innovation hubs to nurture entrepreneurial talent capable of leading future industries worldwide.
2026.07.06 View 1555 -
KAIST: Dementia-Causing Substance Turns On a Therapeutic “Switch”
A substance that worsens dementia has become a “switch” that initiates treatment. KAIST researchers have developed a new therapeutic approach that uses hydrogen peroxide (H₂O₂), a reactive oxygen species that damages cells and increases in the brains of patients with Alzheimer’s disease, to activate a drug selectively in diseased brain tissue. The team also confirmed improvements in cognitive function through animal experiments, presenting a new possibility for next-generation dementia treatment.
KAIST announced on the 2nd that a research team led by Professor Mi Hee Lim of the Department of Chemistry, in collaboration with Professor Mingeun Kim of Chonnam National University, Dr. Chul-Ho Lee and Dr. Kyoung-Shim Kim of the Korea Research Institute of Bioscience and Biotechnology, and Dr. Young-Ho Lee of the Korea Basic Science Institute, has developed a prodrug that is activated selectively in the diseased brain in Alzheimer’s disease and confirmed its therapeutic effects through animal experiments.
A prodrug is a drug that initially has minimal therapeutic effect but is converted into an active therapeutic agent only under specific conditions inside the body. In this study, the prodrug was designed to be activated only when it encounters hydrogen peroxide, which increases in the brains of patients with Alzheimer’s disease, allowing it to function as a “smart therapeutic agent” that selectively acts in diseased brain tissue.
In the brains of Alzheimer’s disease patients, hydrogen peroxide, which damages cells, is elevated above normal levels. Until now, it has generally been regarded only as a harmful substance that should be removed. However, the research team devised a method to use it instead as a signal that activates a drug.
The prodrugs developed by the research team, BE-1 and BE-2, are designed to remain minimally reactive in a healthy brain. However, when they encounter hydrogen peroxide in a brain affected by dementia, they are converted into active therapeutic compounds, AP-1 and AP-2. Through this process, they reduce reactive oxygen species, including hydrogen peroxide, while also preventing amyloid beta (Aβ) peptides — peptides known as a major cause of dementia that accumulate in the brain and damage nerve cells — from aggregating into highly toxic clumps.
Using advanced analytical techniques, the research team confirmed that the activated drug alters the morphology of amyloid beta aggregates and suppresses their growth into large aggregates.
These effects were also confirmed in Alzheimer’s disease mouse models. The drug crossed the blood-brain barrier (BBB), a protective barrier that controls whether substances in the blood can enter the brain, and was converted into the therapeutic compound inside the diseased brain. In mice that received long-term drug administration, oxidative stress in the hippocampus, which is responsible for memory, was reduced, and amyloid beta accumulation in the brain also decreased. In behavioral experiments assessing the ability to recognize new objects and navigate mazes, cognitive function was also found to improve.
This study is significant in that the drug was designed to operate only where needed by using the environment of the diseased brain itself. This approach presents a new strategy for dementia treatment that can enhance therapeutic efficacy while reducing side effects, and it is expected to be applicable to the treatment of other neurodegenerative diseases, such as Parkinson’s disease.
Professor Mi Hee Lim of KAIST’s Department of Chemistry said, “This study is meaningful in that hydrogen peroxide, which had previously been regarded only as something to be eliminated, was used as a signal to activate a drug. We expect this strategy, which activates drugs in diseased tissue, to become a new platform for treating complex diseases such as Alzheimer’s disease more safely and effectively.”
This study was co-first-authored by Jimin Lee and Eunseo Hong, Ph.D. candidates in KAIST’s Department of Chemistry, and was published online on May 31, 2026, in the international journal Small (Impact Factor: 12.1, top 10% in the field of chemistry).
※ Paper title: A Prodrug Approach for Activity-Based Chemical Modulation toward Multiple Pathological Targets in Alzheimer’s Disease
DOI: 10.1002/smll.74013
This research was supported by the National Research Foundation of Korea’s Leader Researcher Program, Global Leading Research Center Program, Sejong Science Fellowship, Graduate Student Research Encouragement Program, and institutional programs of KRIBB and KBSI.
2026.07.03 View 2340 -
KAIST Develops AI That Reads Animal Behavior Like Language
An artificial intelligence model capable of reading and interpreting animal behavior like language has been developed by researchers at KAIST. The team created BehaVERT, an AI model that learns behavioral data in a manner similar to natural language and was able to independently identify social behavioral deficits in an autism mouse model, opening a new avenue for interpretable neuroscience.
KAIST (President Kwang-Hyung Lee) announced that a research team led by Professor Dae-Soo Kim from the Department of Brain and Cognitive Sciences has developed an AI model that interprets animal movements as a form of behavioral language.
The researchers transformed skeletal movements of mice into tokens, analogous to words in natural language, and trained a transformer-based model to learn behavioral meaning. The resulting model, named BehaVERT, successfully identified core social behavioral abnormalities in an autism mouse model without being provided any prior biological knowledge.
The study introduces a novel AI framework for analyzing animal behavior through language-based representations. Beyond simple behavior classification, the model demonstrates the ability to uncover biologically meaningful patterns and may serve as a foundation for next-generation behavioral foundation models applicable to drug discovery, psychiatric research, and behavioral genetics.
Inspired by the idea that animal behavior may possess structures similar to language, the researchers represented the positions of a mouse's nose, ears, spine, limbs, and tail as behavioral tokens and trained a BERT-based transformer architecture.
As a result, BehaVERT learned not only to classify behaviors but also to understand their contextual meaning over time, much like language models infer meaning from sequences of words.
The model achieved state-of-the-art performance across five international benchmark datasets covering social interaction, multi-animal behavior, three-dimensional motion analysis, and autism-related behavioral assessment.
Importantly, BehaVERT also provides interpretability, allowing researchers to visualize which behavioral cues influenced its decisions.
In experiments distinguishing Shank3B knockout autism-model mice from healthy controls, the AI consistently focused on oral-oral contact behavior. This finding aligns with previous biological studies showing that autism-model mice exhibit deficits in social interaction despite maintaining normal approach behavior.
In other words, the AI independently rediscovered a key biological characteristic solely from behavioral observations, without explicit biological instruction.
The researchers further found that the model's internal representation space organized behavioral features such as mobility, attention, and social engagement into structured patterns. This suggests that animal behavior, much like language, may possess an underlying semantic structure.
The study also highlights an unusual interdisciplinary achievement. The first author, Dr. Seungjae Shin, and other members of the research team were trained primarily in biology rather than artificial intelligence. By independently learning transformer architectures and deep learning techniques, they designed specialized models and training strategies tailored for behavioral analysis.
Professor Kim's laboratory has long pursued AI-driven behavioral analysis and previously developed AVATAR, a technology that reconstructs rodent behavior in virtual environments, leading to the founding of Actnova Inc.
"The project began with a simple question: Could animal movements contain a structure similar to language?" said Dr. Seungjae Shin, the first author of the study.
The team also adopted a self-supervised learning framework that enables AI to learn directly from behavioral data without manual annotations. Furthermore, a model trained on rat behavior successfully transferred to mouse behavior analysis, demonstrating the feasibility of a behavioral foundation model applicable across species.
"BehaVERT goes beyond behavior classification and enables the interpretation of behavioral meaning," said Professor Dae-Soo Kim. "We expect it to become a key research tool for discovering new insights in drug development, psychiatric disorders, behavioral genetics, and many other areas of life sciences."
The study was published on March 24, 2026, in the International Journal of Computer Vision (IJCV), one of the world's leading journals in computer vision.
Paper Information
• Title: BehaVERT: A Transformer-Based Motion Language Model for Decoding Behavioral Semantics in Mice
• Journal: International Journal of Computer Vision (IJCV)
• DOI: 10.1007/s11263-026-02834-y
Related Videos
• BehaVERT — Social Behavior Analysis Visualization (Investigation & Mount), https://youtu.be/JshCr-ZBQR0
• BehaVERT — Social Behavior Analysis Visualization (Investigation & Attack), https://youtu.be/p9RPhZM__Js
• BehaVERT — AI Discovers Core Social Behavioral Features in an Autism Mouse Model, https://youtu.be/D6zUyDu3t9I
Funding
This research was supported by the Mid-Career Researcher Program and the Brain Convergence Technology Development Program through the National Research Foundation of Korea (NRF), funded by the Ministry of Science and ICT (MSIT), Republic of Korea.
2026.07.01 View 1878 -
KAIST Identifies Hidden Age Bias in Artificial Intelligence
Do responses generated by artificial intelligence systems such as ChatGPT reflect social prejudice? A KAIST research team has quantitatively analyzed and identified age-related stereotypes embedded in the responses of generative artificial intelligence. The study sheds light on the potential impact of hidden AI biases on social perceptions and suggests directions for the development of more inclusive AI.
KAIST, led by President Kwang Hyung Lee, announced on the 28th that a research team led by Professor Moon Choi of the Graduate School of Science and Technology Policy quantitatively analyzed subtle stereotypes about older adults embedded in sentences generated by OpenAI’s generative AI model ChatGPT-4o.
Generative AI is now widely used in everyday information search and decision-making processes, but concerns have also been raised that it may reproduce social biases contained in its training data. While previous studies have primarily focused on biases related to gender or race, this study, conducted by Ph.D. student Wan Hong as the first author, is significant in that it examined ageism from the perspective of artificial intelligence at a time when the issue is becoming increasingly important amid global population aging. Ageism refers to discrimination against, or negative perceptions of, certain groups based on age.
The research team collected 900 text samples generated by GPT-4o using neutral prompts that asked the model to describe the characteristics of age groups from 10 to 90 in 10-year intervals. The team then analyzed the responses using the Stereotype Content Model, a major theory in social psychology that explains perceptions of people or groups along two dimensions: warmth and competence.
The analysis found that older adults, defined as those aged 60 and above, received high scores for “warmth,” a trait associated with kindness, trustworthiness, and consideration. However, their scores for “competence,” which refers to ability, expertise, and efficiency, tended to be relatively lower than those of younger age groups.
The generated responses also tended to portray the human life course as divided into three groups: youth, covering those in their teens and 20s; middle age, covering those in their 30s to 50s; and older adulthood, covering those in their 60s and above. In particular, descriptions of people aged 70 and older repeatedly showed relatively uniform characteristics.
The research team also focused on “assertiveness,” which refers to the tendency to actively express one’s opinions and act with confidence and initiative. The analysis showed that the frequency of expressions related to assertiveness decreased as age increased. This suggests that ChatGPT-4o tends to portray older adults as wise and caring, while representing their agency and active capacities as relatively lower.
This study is significant because it quantitatively identified subtle biases embedded in generative AI by combining social science theory with computational analysis techniques. The findings show that generative AI tends to portray older adults as a “warm but less competent” group, a pattern similar to typical stereotypes of older adults repeatedly found in mass media.
This study is significant because it quantitatively identified subtle biases embedded in generative AI by combining social science theory with computational analysis techniques. The findings show that generative AI tends to portray older adults as a “warm but less competent” group, a pattern similar to typical stereotypes of older adults repeatedly found in mass media.
“Bias in AI is not merely a technological issue, but a social one,” said Professor Moon Choi. “To build inclusive artificial intelligence, people from diverse generations must participate in the development process.”
The study was conducted with Ph.D. student Wan Hong of the Graduate School of Science and Technology Policy as the first author. The findings were published in the February 2026 special issue of The Gerontologist, a leading international journal in the field of gerontology with an impact factor of 5.7.
※ Paper title: “An Exploratory Semantic Analysis of Age-Related Stereotypes in OpenAI’s GPT-4o Model”
※ DOI: https://doi.org/10.1093/geront/gnaf291
This research was supported by the National Research Foundation of Korea through the Mid-Career Research Program for Convergence between Science and Technology and the Humanities and Social Sciences.
※ Research team homepage: https://aging.kaist.ac.kr
2026.07.01 View 1042 -
KAIST Begins Developing the World’s First Brain-to-Robot Technology, Moving Robots by Thought and Sending Sensation Back to the Brain
KAIST researchers have begun developing a next-generation brain-robot interface platform that uses human brain signals to control an exoskeleton in real time and sends the tactile and force information sensed by the robot back to the brain.
KAIST, led by President Kwang-Hyung Lee, announced on the 25th that research teams led by Professors Kyoungchul Kong and Jung Kim of its Department of Mechanical Engineering, together with Angel Robotics Co., Ltd., have launched the world’s first bidirectional “Brain-to-Robot” system as a flagship initiative of the Korea Medical Device Development Fund (KMDF). The project runs from April 2026 to December 2032.
Professor Kyoungchul Kong is a world-renowned wearable-robotics researcher who founded Angel Robotics, a developer of walking-assist exoskeletons, and led his team to back-to-back gold medals at Cybathlon, the international competition for assistive technologies for people with disabilities. Professor Jung Kim is a globally recognized researcher who received the Scientist and Engineer of the Month Award for his work on robotic skin. Together, the two teams have formed a consortium to develop a Brain-to-Robot platform that merges neural interfaces with exoskeleton robotics.
Brain interface technologies that let users move a cursor or operate a smartphone with brain signals have already reached the stage of human clinical trials, and U.S. companies such as Neuralink and Synchron are accelerating their development. Existing approaches, however, have struggled to link actual movement and sensory feedback at the same time. They have also concentrated largely on advancing signal decoding itself, without clearly defining the target of control, namely what the brain signals actually drive and what kind of sensory information is returned.
Brain-to-Robot is designed to overcome these limitations head-on. It sets the exoskeleton itself as the control target: brain signals read the user’s movement intentions to drive the robot, and at the same time the robot’s sensory readings are delivered back to the brain. These readings include ground reaction force (the force the floor exerts on the foot), joint torque (rotational force at the joints), and tactile information. The aim is a complete bidirectional interface.
According to the research team, no fully bidirectional Brain-to-Robot system that combines exoskeleton control with sensory feedback has yet been reported anywhere in the world, and the project is expected to mark a turning point in brain interface technology.
Within this system, the KAIST teams are responsible for the core technologies. Professor Kong’s team will develop wearable-robot control and AI-based interpretation of movement intention, and will design a somatosensory interface, a system for transmitting bodily sensory information, that delivers the robot’s sensory data accurately to the Brain Chip, the semiconductor that processes brain signals.
Professor Kim’s team will develop robotic skin that senses in place of impaired sensation for people with disabilities, along with AI-based interpretation of somatosensory information.
The teams will also develop AI-based encoding and decoding algorithms that turn brain signals into robot commands and send the robot’s sensory information back to the brain. A key challenge is processing hundreds of channels of cortical signals, the neural signals generated in the cerebral cortex, while stably maintaining an ultra-low-latency closed loop, a control cycle in which signals are exchanged continuously in real time.
Commercialization of the flagship project will be led by Angel Robotics (KOSDAQ: 455900), the company founded by Professor Kong. The team plans to pursue full-cycle commercialization, from regulatory approval by the Ministry of Food and Drug Safety through to real-world deployment.
“If this technology succeeds, it will open a new rehabilitation paradigm in which people with quadriplegia can move beyond the hospital to walk on their own, pick up objects, and even feel sensation at their fingertips in everyday life,” Professor Kong said.
The research team stressed that, because this is an unprecedented and highly complex convergence technology never attempted at home or abroad, long-term safety, clinical validation, and a regulatory approval framework must advance in parallel with the technology itself. To reach the global market, they added, safety and efficacy testing, the accumulation of clinical evidence, a risk-management system, protection of brain-signal data, cybersecurity, and ethical review must all be addressed in an integrated way.
Meanwhile, KAIST is conducting a wide range of fundamental research in the field of brain interfaces. A research team led by Professor Hyung-Soon Park of the Department of Mechanical Engineering is studying wearable rehabilitation robot technologies based on neural intention-recognition interfaces, which identify users’ movement intentions from brain signals, for the effective treatment of neurological disorders. A research team led by Professor Sungho Cho of the School of Computing is developing AI-based brain-signal interpretation technologies.
A research team led by Professor Jihoon Lee of the Department of Brain and Cognitive Sciences is conducting next-generation brain–machine interface research focused on ultra-low-power bio/neural interface circuits, which connect and process biological and neural signals with low power consumption; wireless neural signal measurement technologies, which measure neural signals without wires; and on-device AI-based closed-loop neuromodulation technologies, which use cyclical control structures to exchange signals in real time.
In addition, a research team led by Professor Hyunjoo Lee of the School of Electrical Engineering is conducting research on high-resolution neural signal measurement and precision brain stimulation based on ultra-miniaturized multimodal neural electrodes, which can simultaneously measure and stimulate multiple types of neural signals. A research team led by Professor Minkyu Je of the Department of AI Semiconductor Systems is studying AI-based semiconductor integrated circuits and system technologies for next-generation neural interfaces. A research team led by Professor Jae-Woong Jeong of the School of Electrical Engineering is conducting research on high-precision brain-signal measurement, which precisely measures neural signals generated in the brain, and neuroengineering based on neural stimulation.
“This Brain-to-Robot flagship project is a world-class, highly challenging convergence research initiative led by the teams of Professors Kyoungchul Kong and Jung Kim,” said KAIST President Kwang-Hyung Lee. “KAIST has a wide range of researchers studying fundamental technologies in brain interfaces, AI, semiconductors, and robotics, and based on this foundation, we will lead innovation in next-generation Brain-to-Robot technologies.”
2026.06.25 View 2939