%53%63%68%6f%6f%6c%20%6f%66%20%43%6f%6d%70%75%74%69%6e%67
-
KAIST Develops Vector Database Technology for Reliable Retrieval of Up-to-Date Information
When new information is continually added and old information is removed, an AI system may fail to find the correct material even when that material exists. KAIST researchers have developed a vector database technology that preserves the search paths leading to the needed information as data changes. This allows AI to find up-to-date information more accurately and quickly.
KAIST (President Choongsik Bae) announced on October 5 that a research team led by Professor Min-Soo Kim from the School of Computing has developed CONDA (Connectivity-Aware Dynamic Index). CONDA is a dynamic vector database indexing technology that maintains high search accuracy even as data is continuously added and deleted.
Generative AI models do not inherently have access to information that emerges after training. Retrieval-augmented generation (RAG) has been widely adopted to address this limitation. RAG searches sources such as recent news or internal company documents and uses the results in its answers.
For RAG to work properly, the system needs a search technology that can quickly and accurately find the information relevant to a question among vast amounts of data. Many AI services today convert the meaning of documents or images into vectors, which are numerical representations that computers can process. They then link pieces of information with similar meanings and follow these links to find the material they need.
The challenge is that data in real companies and on the internet changes constantly. New documents, news, and product information keep arriving, while older information is modified or deleted. As these changes repeat, the links built between data points can gradually break, a problem the researchers call "connectivity collapse." The needed information is still in the database, but the AI can no longer reach it, and search accuracy drops. In other words, the information itself has not been lost. The search path that leads to it has been broken.
CONDA considers both distances between data points and whether the search paths needed to reach the information remain connected. When new data is added or existing data is deleted, CONDA preserves the links that are important for search. This prevents specific pieces of information from becoming isolated in the search graph.
In experiments where data was constantly added and deleted, CONDA improved search accuracy (recall) by up to 24.5% compared with state-of-the-art methods. It also increased data update throughput by up to 1.90 times.
The team also ran searches and data updates at the same time for six hours on a large-scale dataset containing 100 million vectors. Throughout the test, CONDA recorded the lowest search latency and the highest search accuracy among the compared methods. This result suggests that the technology can be applied to large-scale AI services.
The technology can be used in knowledge search systems where information such as internal company regulations and work documents changes frequently. It can also be used in a range of AI services where data changes continuously, including news search, product search, and recommendation systems. In particular, it is expected to serve as a foundational technology for RAG, in which generative AI searches external information and uses it in its answers. CONDA allows this up-to-date information to be retrieved reliably over extended periods.
The research results are also being applied to a commercial product. CONDA is being integrated into AkasicDB, the database product of GraphAI, an AI data infrastructure company founded by Professor Kim. It is scheduled for commercial release in the fourth quarter of this year.
"The key value of retrieval-augmented generation (RAG) is that it allows large language models (LLMs) to find and use the latest information they were not trained on at the moment it is needed," said Professor Kim. He added that the study is significant because it presents a data infrastructure that lets AI use up-to-date knowledge accurately over long periods, even in real-world environments where data changes constantly.
Darae Lee, who completed a master's degree at the KAIST School of Computing, is the first author of the study, and Professor Kim is the corresponding author. The findings were presented on September 2 at VLDB 2026 (the International Conference on Very Large Data Bases), a leading international conference in the database field.
Paper title: CONDA: A Connectivity-Aware Dynamic Index for Approximate Nearest Neighbor Search over Evolving Data
DOI: 10.14778/3836663.3836694
This research was supported by the Mid-Career Researcher Program of 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).
2026.10.06 View 764 -
KAIST Develops Smartphone AI That Uses Past Solutions to Tackle Similar Problems
A new AI technology has been developed that retains and reuses knowledge, much like writing down a solution in a notebook and applying it to similar problems rather than asking an expert for help each time. Researchers at KAIST have developed a way for small AI models running on smartphones to store and reuse knowledge from a large server model. The approach reduced server calls by an average of 55.61% compared with a non-cumulative approach while maintaining high accuracy, suggesting that mobile AI could make faster decisions with less server support as it encounters similar problems.
KAIST (President Choongsik Bae) announced on September 21 that a research team led by Professor Jae-Gil Lee from the School of Computing has developed CURE (Cumulative Knowledge Reuse), a technology that enables a small AI model on a device to work efficiently with a large model on a server.
Smartphones have limited processing power and memory, so they typically use small, lightweight AI models. These models can handle simple tasks quickly but may be less accurate when identifying complex or unfamiliar images.
Sending every input to a powerful server model can improve accuracy, but transferring data and waiting for a response takes time. It also adds to network traffic and the server’s computational workload, increasing delays for users and operating costs for service providers.
Researchers have explored a collaborative approach in which the on-device model assesses each input first and sends only difficult cases to the server. Without a way to retain the server’s knowledge, however, the device uses each answer once and may need the same help when a similar input appears. This is much like a student who fails to write down the solution to a difficult problem and has to ask for help again when faced with another of the same kind. The team set out to turn these one-time answers into knowledge the device could continue to use.
CURE first checks whether the on-device model can handle an input reliably. If it cannot, the system consults knowledge previously obtained from the server and stored on the device. It contacts the server only if that knowledge is also insufficient. The process follows three steps, moving from solving a problem independently to consulting previous lessons and, when necessary, asking an expert.
For example, an on-device model that cannot identify a car model in a photo may turn to the server for help. CURE uses the server’s prediction and the image’s features to update the device’s knowledge store. When the device later receives a photo of a similar car model, it can use that knowledge to identify it locally.
CURE does not keep a collection of the original photos processed by the server. Instead, it stores a summary of their shared features and differences. This is like noting a car’s distinguishing features, such as its body shape or headlight design, rather than memorizing the entire photo. The stored knowledge can therefore help the system recognize not only images it has already seen but also similar images it encounters for the first time.
The team tested CURE using vision-language models, which connect images with text to understand visual information. These models link what they see to language, much as people do.
The researchers used MobileCLIP2 on the device and EVA-CLIP, which has 18 billion parameters, on the server. Parameters are numerical values within a model that encode information learned during training. They help the model distinguish objects and identify relationships between them.
In tests on a range of image classification datasets, CURE achieved accuracy close to that of an approach that sends every input to the large server model. It also made an average of 55.61 percent fewer server calls than a device-server collaboration baseline that does not retain knowledge from previous server responses.
In end-to-end tests that included communication time, CURE ran up to 2.80 times as fast as the non-cumulative device-server collaboration baseline and up to 3.67 times as fast as the approach that sends every input to the server. Making fewer server calls reduced the time spent transferring data and waiting for responses.
CURE requires no retraining of either the on-device model or the server model. It leaves both models unchanged and adds a separate store for knowledge obtained from the server, making it adaptable to a range of AI models and services.
The technology could reduce repeated data transfers and server computation image recognition. This could mean shorter waits for users and lower operating costs for service providers.
Potential applications also include robots and wearables that need to recognize their surroundings with limited computing resources. Robots operating over slow or unreliable networks, for example, could use previously acquired knowledge to make more decisions locally without waiting for a server response.
The benefits in practice will depend on the device’s processing power and storage capacity, network conditions, and the characteristics of the input data. Further testing across different devices and environments is therefore needed.
“CURE allows a small on-device AI model to remember and reuse knowledge it has already obtained, rather than repeatedly asking the server the same question,” said Professor Jae-Gil Lee. “By maintaining high accuracy while reducing communication demands and response times, we expect it to help smartphones, robots, and wearables use powerful AI models more efficiently.”
Dr. Youngjun Lee, a postdoctoral researcher from the KAIST Institute of Information Electronics, was the study’s first author, and Professor Jae-Gil Lee from the School of Computing was the corresponding author. Co-authors were Doyoung Kim from Amazon, Junhyeok Kang from LG AI Research, and Professor Hwanjun Song from the KAIST Department of Industrial and Systems Engineering. The findings were presented on September 10 at the European Conference on Computer Vision (ECCV 2026), a leading international computer vision conference held in Malmö, Sweden, from September 8 to 12.
Paper title “CURE: Cumulative Knowledge Reuse for Efficient Device-Server Hybrid Inference in Vision-Language Models”,
DOI: https://doi.org/10.1007/978-3-032-37627-5_21
Author information — Dr. Youngjun Lee (postdoctoral researcher, KAIST Institute of Information Electronics, first author), Doyoung Kim (KAIST graduate, now at Amazon, co-author), Junhyeok Kang (KAIST graduate, now at LG AI Research, co-author), Professor Hwanjun Song (KAIST Department of Industrial and Systems Engineering, co-author), Professor Jae-Gil Lee (KAIST School of Computing, corresponding author)
This work was supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2020-II200862, DB4DL: High-Usability and Performance In-Memory Distributed DBMS for Deep Learning, 50% and No. RS-2022-II220157, Robust, Fair, Extensible Data-Centric Continual Learning, 40%), the InnoCORE pro-gram of the Ministry of Science and ICT (AI Meta-Scientist, N10260110, 10%), and Samsung Electronics Co., Ltd.(IO251216-14634-01).
2026.09.21 View 2385 -
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 2781 -
KAIST Develops Smartphone-Based Technology to Detect Hidden Cameras
A smartphone can now be transformed into a “hidden-camera detector.” KAIST researchers have developed an AI technology that can detect hidden cameras using only a smartphone and a low-cost LED device. This new security technology enables users to protect their privacy more easily and is expected to help prevent illegal filming in everyday spaces such as hotels and short-term rentals.
KAIST (President Choongsik Bae) announced on August 30 that a research team led by Professor Jun Han of the School of Computing, in collaboration with the National University of Singapore and Singapore Management University, has developed “SweepLED,” a technology that detects hidden cameras by attaching an LED case to a smartphone.
As hidden cameras are increasingly being installed in everyday spaces such as hotels, short-term rentals, and restrooms, the need is growing for detection technology that everyday users can easily use. However, existing portable detectors require users to visually identify bright reflective spots, which can lead to false positives by mistaking reflections from metal, glass, or glossy plastic surfaces for camera lenses.
SweepLED works by keeping the smartphone camera fixed while changing only the direction of the LED illumination, then analyzing the patterns of reflected light that appear on object surfaces. Reflections from ordinary glossy objects tend to move or disappear depending on the direction of the light. In contrast, camera lenses show distinctive deformation patterns in their reflections due to their internal lens, aperture, and sensor structures.
The research team uses deep learning-based analysis to distinguish these differences in temporal reflection patterns. While conventional detection methods rely on the user’s eyes to simply look for “bright spots,” SweepLED is different in that it analyzes both the movement and shape changes of reflections across multiple lighting angles.
This enables more reliable detection of hidden camera lenses inside various everyday objects commonly found in lodging spaces, such as chargers, clocks, remote controls, and everyday objects.
The research team evaluated SweepLED on 30 objects that may be found in real-world environments and found that it achieved approximately 94% detection accuracy. It also took less than five seconds to inspect a single object.
In addition, the core components of the LED case attached to the smartphone cost less than USD 7, or about KRW 10,000, demonstrating the potential for this technology to be developed into an affordable detection tool that general users can easily access.
Professor Jun Han said, “Hidden cameras pose a serious threat to personal safety and privacy in everyday spaces,” adding, “This research is meaningful in that it combines low-cost smartphone-based hardware with AI analysis to present the possibility of a practical detection technology that even non-experts can use.”
This paper, with KAIST doctoral student Jonghyuk Yun as first author, was presented on June 20 at ACM MobiSys 2026, one of the leading international conferences in the field of mobile computing.
Paper title: Hide-and-Sweep: Detecting Concealed Cameras via LED Illumination Sweeps
https://doi.org/10.1145/3812835.3814866
Author information: Jonghyuk Yun (first author), Jaeyoung Moon, Yunseo Park, Sean Rui Xiang Tan, Byunghyun Kim, Rajesh Krishna Balan, and Professor Jun Han (corresponding author)
This research was supported by the STEAM Global Convergence Research Support Program and the Mid-Career Researcher Program of the Ministry of Science and ICT and the National Research Foundation of Korea.
2026.08.31 View 10185 -
KAIST Develops AI to Detect ‘Foreign-Linked Opinion Manipulation’ in 110 Million News Comments
During election seasons or major national issues, online news comment sections often become heated spaces of conflict across gender, generation, and political lines. For years, there have been persistent concerns that behind some of these conflicts may lie “foreign winds,” or interventions by foreign actors seeking to manipulate public opinion and deepen social divisions. A KAIST research team has now developed a technology that uses big data on two decades worth of news comments and artificial intelligence (AI) to precisely detect traces of such hidden influence operations.
KAIST (President Choongsik Bae) announced on the 12th of August that a joint research team led by Professor Wonjae Lee of the Graduate School of Culture Technology, Professor Meeyoung Cha of the School of Computing (Scientific director at Max Planck Institute for Security and Privacy) and Professor Alice Oh of the School of Computing, in collaboration with Professor Thorsten Holz of the Max Planck Institute, has developed an explainable AI technology that automatically detects patterns suspected of foreign-linked influence operations in online news comments and provides specific evidence for its judgments.
The organized and repeated posting of comments or content by certain actors to shape public opinion in a desired direction is known as an “online influence operation.” Existing AI-based detection technologies have had a key limitation: even when they classify certain accounts as belonging to a “blacklist,” they often fail to provide clear evidence explaining why those accounts should be considered influence-operation accounts.
To overcome this limitation, the research team used 70 foreign-linked accounts previously identified by the Institute for National Security Strategy as starting points. They then tracked groups of accounts connected to them or repeatedly commenting on the same news articles, ultimately collecting and analyzing a large-scale dataset of 110 million comments posted on Naver News over a 20-year period from 2006 to 2025.
In particular, the AI developed by the research team examines accounts through a careful three-step process. First, it checks whether there are clues suggesting that the author may be linked to a foreign source. Second, it examines whether the comment contains emotionally polarizing expressions, such as moral condemnation or blind praise. Third, it identifies which country or target the emotional framing is directed toward.
The model does not stop at simply labeling an account as suspicious. It also highlights the specific phrases in the comments that served as the basis for its judgment. The team further combined this with multidimensional behavioral-pattern analysis, including account activity frequency, account lifespan, and activity links with other suspected accounts. As a result, among approximately 4 million Naver News users, the model ultimately identified 23,998 accounts exhibiting patterns consistent with suspected public-opinion manipulation.
The analysis also revealed the more subtle strategy of these suspected accounts. Their main target was not the victory of a particular political camp, but rather the maximization of division and confrontation within Korean society.
Among the top 10 targets that drew the highest public engagement, measured through likes and other reactions, seven were prominent domestic political figures. Notably, the attacks were not concentrated on a single party or ideology. Former and current presidents, presidential candidates, and political parties from both progressive and conservative camps were targeted across the spectrum. According to the research team, this suggests a sophisticated strategy aimed not so much at supporting a particular group, but at inflaming domestic political conflict and increasing social distrust and polarization.
This study is significant because it provides data-based evidence for influence-operation activity that had previously been discussed largely in terms of suspicion, while also offering a potential defense mechanism for protecting healthy online public discourse. In the future, portal platforms and related organizations could use this technology during elections or national crises to monitor the influx of suspicious accounts in real time and prioritize the review of coordinated attacks against domestic political figures. However, the research team emphasized that the AI should not be used to block accounts indiscriminately, but rather as an explainable content-moderation tool that supports the judgment of expert reviewers.
Professor Wonjae Lee said, “By analyzing 20 years of data, we found that suspected accounts tended to use messages criticizing Korea and domestic political figures rather than directly praising foreign countries, and that these messages gained higher visibility,” adding, “This research can provide empirical criteria for when and which messages platforms and monitoring organizations should prioritize for review, especially during socially sensitive periods such as elections.”
Professor Alice Oh said, “This is a meaningful achievement in which AI precisely identified not only the surface meaning of words in massive comment datasets, but also subtle emotional patterns and organized behavioral signals intended to provoke conflict,” adding, “It can become a powerful defense system against online influence operations, which are becoming increasingly sophisticated.”
Professor Meeyoung Cha said, “This study goes beyond simple blacklist-account analysis and represents the outcome of actionable data science that addresses real-world problems and drives practical change,” adding, “In an online environment where social conflict is deepening, we hope this technology will serve as a practical tool for protecting the transparency and trustworthiness of the digital public sphere.”
This research was led by KAIST Ph.D. candidate Jaehong Kim and master’s student Hyeonseung Kim as co-first authors. The paper is scheduled to be presented at the USENIX Security Symposium 2026, one of the most prestigious conferences in the field of computer security.
Paper title: Cross-National Information Attacks: A Two-Decade Analysis of Troll Behavior in Korea,
DOI: 10.48550/arXiv.2606.22785
This research was supported by the Hyundai Motor Chung Mong-Koo Foundation, the Institute of Information & Communications Technology Planning & Evaluation, and the National Research Foundation of Korea, funded by the Ministry of Science and ICT.
2026.08.12 View 4514 -
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 2256 -
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 3984 -
KAIST Develops AI That Learns to Theorize the World from Observation, Inspired by How Children Learn
A KAIST research team has developed a next-generation world model, an internal model an AI builds to understand and predict the world, that learns executable theories from observation alone.
KAIST (President Choongsik Bae) announced on the 15th of July that a team led by Professor Sungjin Ahn from the School of Computing has proposed a new learning paradigm called Learning-to-Theorize (L2T), which trains AI to theorize how the world works using only observed information. The team also built the Neural Theorizer (NEO), a neural network-based model that implements this paradigm.
The research was selected for an oral presentation at the 43rd International Conference on Machine Learning (ICML 2026), held in Seoul from July 6 to 11, and was presented on July 9. This places it among the top 0.7 percent (168 papers) of the 23,918 total submissions. The paper was also selected for the Best Paper Award at the Compositional Learning Workshop.
World models are a foundational technology across robot control, autonomous driving, generative AI, and autonomous agents — AI systems capable of judging and acting on their own. Until now, world models have mainly focused on predicting what happens next. Even when a model predicts the next scene accurately, it doesn't necessarily understand why that change occurred — that is, the underlying principle governing the world.
The team found a solution in how humans learn. Long before children acquire language, they build their own internal theories of how the world works. Applying this view from developmental cognitive science to AI, the researchers built a system that understands the principles behind the world, rather than one that simply predicts the future.
The team's proposed L2T framework provides no predetermined answers or rules. Given only a "before" and "after" observation, the AI discovers on its own which rule produced the change. While conventional AI models focus on "guessing what comes next," this approach is built to understand "why the change happened."
To implement this, the team also developed Neural Theorizier, NEO. The model discovers reusable primitives hidden within observed transformations and composes them into executable programs. These learned primitives can then be systematically recombined to explain new situations.
For example, NEO independently learns primitives corresponding to basic operations such as rotation, movement to the left or downward, and coloring. Even when presented with a combination it has never encountered during training, such as “move down, then color, then rotate”, it can recombine the primitives it has already learned to explain and solve the new situation.
Conventional AI, by contrast, tends to memorize entangled patterns as a single unit, so its performance drops sharply when faced with an unfamiliar combination. Through a range of experiments, the team demonstrated that NEO outperforms existing approaches in compositional generalization, the ability to combine learned basic rules to solve problems never seen before.
"It points to a new direction beyond prediction-centric world models, what we call a 'World Theory Model.' We expect this to develop into a core technology across fields including intelligent robots, autonomous agents, and AI that supports scientific discovery." said Professor Sungjin Ahn.
Master's students Doojin Baek and Gyubin Lee from the School of Computing served as co-first authors on the study.
Paper title: Learning to Theorize the World from Observation.
DOI: https://doi.org/10.48550/arXiv.2605.03413
Authors: Doojin Baek*, Gyubin Lee*, Junyeob Baek, Hosung Lee, Sungjin Ahn (*Co-first authors)
The research was supported by the National Research Foundation of Korea (NRF).
2026.07.27 View 3356 -
KAIST Develops Robot Learning Technology Capable of Precisely Imitating Even “Rough” Demonstrations
Robots with increasingly precise dexterity are becoming essential in everyday life and industrial settings, from assembling tiny smartphone components to assisting doctors in surgery. However, teaching robots delicate human movements has traditionally required collecting vast amounts of data at extremely fine time intervals, resulting in significant costs and time burdens. KAIST researchers have developed a robot artificial intelligence technology that can perform sophisticated tasks by autonomously adjusting precision according to the situation, even when trained only on coarsely (sparsely) sampled demonstrations.
KAIST, led by President Kwang Hyung Lee, announced on the 24th that a research team led by Professor Daehyung Park of the School of Computing has developed DiSPo, a multi-granularity manipulation model that generates fine-grained robot motions tailored to a user’s desired level of precision, even from rough human demonstrations.
Existing robot learning methods, such as Behavior Transformer and Diffusion Policy, are limited by their dependence on the time intervals of the data used during training. As a result, learning precision manipulation tasks such as screw fastening or component insertion has required collecting large volumes of high-frequency data at very short time intervals. This has significantly increased data collection costs and slowed down the inference speed of robot AI models.
To overcome these limitations, the research team combined Mamba, a state-space model capable of predicting time intervals, with a diffusion model that enables rich action representation. The team also introduced a new Step-scale factor mechanism, which allows users to directly control the time intervals used by the robot.
As a result, even when trained on only low-frequency (coarse) demonstration data, the robot can generate high-precision motions during inference without additional training by autonomously subdividing actions through a discretization process.
DiSPo achieved up to an 81% higher task success rate compared to state-of-the-art models in simulation environments. In real-world experiments using a collaborative robot, DiSPo stably performed challenging tasks such as passing a clamp through a narrow gap with only a 2.5 mm radial clearance and accurately pressing a small shutter button on a smartphone. This performance was up to four times higher than that of existing AI models.
The technology is expected to make a significant contribution to automation in a wide range of everyday and industrial service fields that require high precision, including precision component assembly, cable connection, medical surgery, and precision machining.
“This study demonstrates that robots can learn precise motions from coarse demonstrations and autonomously adjust their level of precision according to the task situation,” said Professor Daehyung Park. “Moving forward, this technology is expected to dramatically reduce data collection costs while serving as a general-purpose robot learning technology for various industrial fields, including precision assembly and medical applications.”
The study was led by Nayoung Oh, a master’s student at the KAIST Graduate School of AI, as the first author, and was presented on June 1 at the 2026 IEEE International Conference on Robotics and Automation, or ICRA 2026, one of the world’s most prestigious robotics conferences, held in Vienna, Austria.
Paper Title: DiSPo: Diffusion-SSM based Policy Learning for Coarse-to-Fine Action Discretization
DOI: https://doi.org/10.48550/arXiv.2409.14719
2026.06.24 View 3261 -
KAIST Develops Next-Generation Database Technology That Reduces AI Hallucinations and Improves Accuracy by 78%
One of the greatest weaknesses of AI agents that read and understand vast amounts of enterprise data is “hallucination” — the generation of plausible-sounding but factually incorrect information. KAIST researchers have developed a next-generation database technology capable of understanding documents, data, and relationships among entities all at once. The technology improves AI response accuracy by up to 78% and processing speed by up to 20 times, addressing a key challenge in the commercialization of enterprise AI.
KAIST, led by President Kwang Hyung Lee, announced on the 19th that a research team led by Professor Min-Soo Kim of the School of Computing, in collaboration with faculty startup GraphAI Co., Ltd., has developed “AkasicDB,” a next-generation database technology that integrates the functions of vector databases, graph databases, and relational databases into a single database management system (DBMS). Based on this technology, the team also developed a new Retrieval-Augmented Generation (RAG) method called “Omni RAG.”
AkasicDB is designed to integrate and execute the functions of vector databases, which convert the meaning of documents or images into numerical vectors to search for similar information; graph databases, which store and analyze relationships among entities such as people, companies, and products; and relational databases, which systematically manage data in table form. Omni RAG, developed on this foundation, improves the accuracy of generative AI responses by simultaneously utilizing semantic information from documents, relationships among entities, and structured data.
AI agents have recently been spreading rapidly based on RAG technology, which searches vast collections of corporate documents and expert knowledge and generates responses based on the retrieved information. However, real-world enterprise data is distributed across various forms, including documents, tables, and relationships among entities, making it difficult for AI to comprehensively understand and use the data. As a result, AI may generate factually incorrect responses without sufficient grounding, creating hallucination issues that have been regarded as a major obstacle to the broader adoption of enterprise AI.
Conventional RAG typically works by converting user queries and documents into vectors, retrieving semantically similar documents, and providing them to a large language model (LLM), an AI model trained on massive datasets to generate human-like language. While this approach is effective for searching unstructured documents, it has limitations when handling complex queries that must also consider relationships among entities in documents or structured conditions such as specific periods, types, or ranges.
For example, a query such as, “Find clauses related to Company A among contracts signed last year, and explain how those clauses are connected to product supply issues,” requires vector search to identify document meaning, graph search to explore relationships among entities, and relational queries to filter by date and type. In existing systems, this required building multiple types of databases separately and combining the results at the application layer, leading to management complexity and response delays.
To solve this problem, the research team proposed Omni RAG, which integrates vector similarity search, graph traversal, and relational filtering within a single query and execution plan. Omni RAG identifies more accurate evidence by simultaneously utilizing semantic information from documents, relationship information from knowledge graphs, and structural conditions from tabular data, significantly reducing AI hallucinations.
AkasicDB, developed to support this method, adopts a new architecture that integrates graph databases, vector databases, and relational databases into a single engine. Users can express complex RAG queries that combine vector search, graph traversal, and relational filtering as a single SQL/GQL* query, and AkasicDB optimizes and processes the query as one unified execution plan.
SQL/GQL, or Structured Query Language/Graph Query Language, refers to command languages used to search or modify information stored in databases. SQL is the traditional language used to handle tabular data, while GQL is a language dedicated to graph data and is used to analyze connections among entities such as people, companies, and products.
Through this integrated architecture, AkasicDB minimizes unnecessary intermediate result generation and data movement, greatly reducing the number of tokens used by LLMs and shortening response latency. In experiments, complex search queries that took up to 21.3 seconds in existing systems were processed in under one second, achieving a performance improvement of more than 20 times. Omni RAG also improved response accuracy by up to 78% compared with conventional RAG. These results demonstrate its potential to substantially mitigate hallucination, one of the core challenges for enterprise AI agents.
Professor Min-Soo Kim said, “For AI agents to accurately understand and utilize the vast amounts of data held by enterprises, data infrastructure capable of processing vector, graph, and relational data in an integrated manner within a single system is essential. AkasicDB is a next-generation database technology for the era of AI agents, and we expect it to be used as core data infrastructure in fields requiring high reliability, including defense, manufacturing, finance, law, science, and technology.”
KAIST School of Computing Ph.D. student Geonho Lee participated in this research as the first author. The research results were presented as a demo paper on June 2 at ACM SIGMOD 2026, one of the world’s most prestigious international conferences in the field of databases, where they drew strong interest from global companies and researchers.
※ Paper title: AkasicDB: Demonstrating Omni RAG with a Unified Vector-Graph-Relational DBMS
DOI: https://doi.org/10.1145/3788853.3801609
※ Author information: Geonho Lee, KAIST, first author; Jeongho Park and Donghyoung Han, GraphAI Co., Ltd., co-authors; Professor Min-Soo Kim, KAIST, corresponding author
※ Demonstration video: https://www.youtube.com/watch?v=KD6MznZ61P4
2026.06.19 View 4645 -
Professor Insik Shin Becomes First Korean to Win the RTSS Most Influential Paper Award
< KAIST Professor Insik Shin >
KAIST announced on December 21st that Professor Insik Shin from the School of Computing has received the Influential Paper Award 2025 at the IEEE Real-Time Systems Symposium (RTSS), the world's most prestigious international conference in the field of real-time systems.
This honor is a "Test of Time Award," presented to papers that have exerted a sustained and significant influence on both academia and industry for more than 10 years after publication. This marks the first time a Korean researcher has received this prestigious award. The ceremony took place at IEEE RTSS 2025 in Boston, USA, on December 4th (local time).
Professor Shin’s award-winning research is the "Periodic Resource Model," co published in 2003 with Professor Insup Lee of the University of Pennsylvania. Rather than trying to verify a complex machine or system all at once, this study developed a method to verify individual components—much like LEGO blocks—to ensure each meets its designated timing requirements. It mathematically guarantees that when these components are assembled, the entire system will operate safely.
Paper Title: Periodic Resource Model for Compositional Real-Time Guarantees
DOI: 10.1109/REAL.2003.1253249
Thanks to this research, it has become possible to design real-time systems that cannot tolerate even a moment of delay—such as autonomous vehicles, aircraft, and industrial robots—with greater precision and safety. This breakthrough overcame the limitations of existing methods that required analyzing an entire system at once, which had become nearly impossible as the complexity of modern real-time systems increased rapidly.
Professor Shin presented a method to divide a system into small modules, verify that each module satisfies its time constraints, and mathematically prove that the safety of the entire system is guaranteed upon integration. This work is credited with establishing the foundation for modern compositional real-time scheduling theory.
At the time of its initial publication in 2003, this paper won the 'Best Paper Award' at RTSS—another first for a Korean researcher. Now, 20 years later, its academic and industrial value has been officially recognized once again. This is because the theory has transcended academic boundaries to become a core analytical tool in various safety-critical industries, including autonomous driving, aerospace control, and industrial automation.
The IEEE Technical Committee stated, "This model has established itself as a core language for modern real-time system design and has guided the direction of research and industry for the past 20 years." The paper is currently featured in textbooks at major universities in the United States and Europe, serving as a standard theory in the field.
"As a scholar, this is the award I have wanted most in my life," said Professor Shin. "I am honored to have it recognized that research from 20 years ago has truly had a major impact on the world. This was made possible by the many researchers and companies who applied this theory to actual systems."
Meanwhile, Professor Shin is expanding his research beyond real-time systems into the field of Artificial Intelligence (AI). He founded the faculty-led startup Fluiz and developed FluidGPT, a mobile AI agent technology that allows users to execute smartphone apps via voice commands. This technology recently won the AI Champion Competition hosted by the Ministry of Science and ICT. Experts evaluate Professor Shin as achieving rare success by bridging basic theory and applied technology, effectively linking research to industry.
2025.12.22 View 4353 -
AI Technology World No. 1 in Finding the Exact Moment in a Video: Where is the First Place?
< (From left) Professor Joon Hyuk Noh (Assistant Professor, Department of Artificial Intelligence, Ewha Womans University), Seojin Hwan, Yoonki Cho (Ph.D. Candidate), Professor Sung-Eui Yoon (School of Computing, KAIST) >
When faced with a complex question like 'What object disappeared while the camera was pointing elsewhere?', a common problem is that AI often relies on language patterns to guess a 'plausible answer,' instead of actually observing the real situation in the video. To overcome this limitation, our university's research team developed a technology that enables the AI to autonomously identify the 'exact critical moment (Trigger moment)' within the video, and the team’s excellence was proven by winning an international AI competition with this technology. The university announced on the 28th that the research team led by Professor Sung-Eui Yoon from the School of Computing, in collaboration with Professor Joon Hyuk Noh's team from Ewha Womans University, took 1st place in the Grounded Video Question Answering track of the Perception Test Challenge held at ICCV 2025, a world-renowned computer vision conference. The Perception Test Challenge held at ICCV 2025 was organized by Google DeepMind with a total prize pool of 50,000 Euros (approximately 83 million KRW). It assesses the cognitive and reasoning abilities of multimodal AI, which must comprehensively understand various data, including video, audio, and text. Crucially, the core evaluation factor is the ability to make judgments based on actual video evidence, moving beyond language-centric bias. Unlike conventional methods that analyze the entire video indiscriminately, our university's research team developed a new technology that instructs the AI to first locate the core scene (Trigger moment) essential for finding the correct answer. Simply put, this technology is designed to make the AI autonomously determine: “This scene is decisive for answering this question!” The research team calls this framework CORTEX (Chain-of-Reasoning for Trigger Moment Extraction). The research team's system consists of a three-stage structure where three models performing different functions operate sequentially. First, the Reasoning AI (Gemini 2.5 Pro) reasons about which moment is required to answer the question and finds candidate Trigger moments. Next, the Object Location Finding Model (Grounding Model, Molmo-7B) accurately identifies the exact location (coordinates) of people, cars, and objects on the screen during the selected moment. Finally, the Tracking Model (SAM2) precisely tracks the movement of objects in the time frame before and after the selected scene, using that scene as a reference, thereby reducing errors. In short, the 'method of accurately pinpointing a key scene and tracking the evidence for the answer centered on that scene' significantly reduced problems like initial misjudgment or occlusion in the video. In the Grounded Video Question Answering (Grounded VideoQA) track, which saw 23 participating teams, the KAIST team SGVR Lab (Scalable Graphics, Vision & Robotics Lab) recorded 0.4968 points in the HOTA (Higher Order Tracking Accuracy) metric, overwhelmingly surpassing the 2nd place score of 0.4304 from Columbia University, USA, to secure 1st place. This achievement is nearly double the previous year's winning score of 0.2704 points. This technology has wide-ranging applications in real-life settings. Autonomous driving vehicles can accurately identify moments of potential accident risk, robots can understand the surrounding environment smarter, security and surveillance systems can rapidly locate critical scenes, and media analysis can precisely track the actions of people or objects in chronological order. This is a core technology that enables AI to judge based on "actual evidence in the video." The ability to accurately pinpoint how objects behave over time in a video is expected to greatly expand the application of AI in real-world scenarios in the future.
< Pipeline image of the grounding framework for video question answering proposed by the research team >
This research was presented on October 19th at ICCV 2025, the 3rd Perception Test Challenge conference. The achievement was supported by the Ministry of Science and ICT's Basic Research Program (Mid-Career Researcher), the SW Star Lab Project's 'Development of Perception, Action, and Interaction Algorithms for Open-World Robot Services,' and the AGI Project's 'Reality Construction and Bi-directional Capability Approach based on Cognitive Agents for Embodied AGI' tasks."
2025.12.01 View 4347