%61%63%63%75%72%61%74%65
-
AI Gets a Private Tutor, Learning Human Preferences More Accurately
< Professor Junmo Kim and Ph.D. candidate Minchan Kwon, School of Electrical Engineering >
No matter how much data they learn, why do Artificial Intelligence (AI) models often miss the mark on human intent? Conventional "comparison learning," designed to help AI understand human preferences, has frequently led to confusion rather than clarity. A KAIST research team has now presented a new learning solution that allows AI to accurately learn human preferences even with limited data by assigning it a "private tutor."
On December17th, a research team led by Professor Junmo Kim of KAIST School of Electrical Engineering announced the development of "TVKD" (Teacher Value-based Knowledge Distillation), a reinforcement learning framework that significantly improves data efficiency and learning stability while effectively reflecting human preferences.
Existing AI training methods typically rely on collecting massive amounts of "preference comparison" data—simple structures like "A is better than B." However, this approach requires vast datasets and often causes the AI to become confused in ambiguous situations where the distinction is unclear.
To solve this problem, the research team proposed a method in which a ‘Teacher model’ that has first deeply understood human preferences delivers only the core information to a ‘Student model.’ This can be compared to a private tutor who organizes and teaches complex content, and the research team named this ‘Preference Distillation.’
The biggest feature of this technology is that instead of simply imitating ‘good or bad,’ it is designed so that the teacher model learns a ‘Value Function’ that numerically judges how valuable each situation is, and then delivers this to the student model. Through this, the AI can learn by making comprehensive judgments about ‘why this choice is better’ rather than fragmentary comparisons, even in ambiguous situations.
< Conceptual diagram of TVKD: After teaching the human preference dataset to the teacher model, learning proceeds by delivering the teacher's information and the dataset to the student model >
The core of this technology is twofold. First, by reflecting value judgments that consider the entire context into the student model, learning that understands the overall flow rather than fragmentary answers has become possible. Second, a technique was introduced to adjust learning importance according to the reliability of preference data. Clear data is significantly reflected in learning, while the influence of ambiguous or noisy data is reduced, allowing the AI to learn stably even in realistic environments.
As a result of the research team applying this technology to various AI models and conducting experiments, it showed more accurate and stable performance than methods previously known to have the best performance. In particular, it recorded achievements that stably outperformed existing top technologies in major evaluation indices such as MT-Bench and AlpacaEval.
Professor Junmo Kim said, “In reality, human preference data is not always sufficient or perfect,” and added, “This technology will allow AI to learn consistently even under such constraints, so it will be highly practical in various fields.”
< Performance comparison results for each task of MT-Bench. It can be confirmed that the proposed TVKD framework records generally higher scores than existing methods. >
< Visualization results of the Shaping term. The top tokens (converted into words) judged as important by the teacher model within the response are displayed in red, intuitively showing which tokens have a greater influence during the value-based alignment process. >
Ph.D. candidate Minchan Kwon from our university’s School of Electrical Engineering participated as the first author, and the research results were accepted at ‘NeurIPS 2025’, the most prestigious international conference in the field of artificial intelligence. The research was presented at a poster session on December 3, 2025 (US Pacific Time).
※ Paper Title: Preference Distillation via Value based Reinforcement Learning, DOI: https://doi.org/10.48550/arXiv.2509.16965
Meanwhile, this research was carried out with support from the Information & Communications Technology Planning & Evaluation (IITP) funded by the government (Ministry of Science and ICT) in 2024 (No. RS-2024-00439020, Development of Sustainable Real-time Multimodal Interactive Generative AI, SW Star Lab).
2025.12.17 View 4077 -
Accurate Real time ECG Measurement While Comfortably Lying Down at Home
< (From left) Professor Chul Kim of the Department of Bio and Brain Engineering, Ph.D. candidate Minjae Kim, researcher Premravee Teeravichayangoon >
KAIST's research team has developed a technology that can measure electrocardiogram (ECG) and heart rate variability (HRV) in real time by simply lying on a bed with clothes on, without having to go to the hospital. This technology is expected to evolve into a daily heart health monitoring platform in conjunction with remote healthcare, and further expand into various bio-healthcare fields such as sleep and stress analysis, contributing to personalized prevention and early diagnosis for patients.
KAIST announced on the 19th that Professor Chul Kim's research team from the Department of Bio and Brain Engineering has developed an 'in-bed cardiac monitoring on-device system'.
The research team manufactured a flexible substrate sensor that integrates the electronic circuit and electrodes into one to increase precision, and implemented an integrated system that can perform signal-noise separation, heart beat signal (R-peak) detection, and heart rate variability analysis in real time through on-device signal processing.
Existing ECG measurement had the inconvenience of visiting a hospital, taking off clothes, and attaching wet electrodes to the skin. Because of this, long-term monitoring was difficult, and it was not easy for the elderly or patients with chronic diseases to use it daily. Non-contact methods also had a technical limitation of being vulnerable to external noise.
To solve these problems, the research team applied a circuit that blocks external noise (active shielding) and a circuit that stably captures minute current changes in the human body (right-leg drive circuit). In addition, they implemented a mathematical transformation technique (wavelet transform) that extracts only the important parts from the heart beat signal and a calculation method (peak detection algorithm) that accurately identifies the moment of the heart's electrical beat (R-peak) as on-device signal processing techniques, allowing for precise real-time analysis of the signal.
As a result, users can obtain stable and accurate ECG signals even when lying on their backs with clothes on.
< Figure. Overall structural diagram of the developed non-contact in-bed cardiac monitoring on-device system, schematic diagram of the R-peak detection algorithm, real-time ECG and HRV measurement screen >
This research presents new possibilities for managing chronic cardiovascular diseases and supporting the health of the elderly, as it can be easily used not only in hospitals but also at home.
Professor Chul Kim said, "This system, which can extract signals in real time even in a noisy environment, can be used to easily check heart health in daily life," and added, "In the future, it will become the foundation of sleep health management by adding the measurement of various bio-signals."
This paper, in which Ph.D. candidate Minjae Kim and researcher Premravee Teeravichayangoon from the Department of Bio and Brain Engineering participated as co-first authors, was published online in the international journal 'Biosensors and Bioelectronics' on August 9, 2025.
※ Paper title: A homecare in-bed hardware system for precise real-time ECG and HRV monitoring with layered clothing. DOI: https://doi.org/10.1016/j.bios.2025.117838
※ Author information: Minjae Kim (KAIST Department of Bio and Brain Engineering, First Author), Premravee Teeravichayangoon (KAIST Department of Bio and Brain Engineering, First Author), Chul Kim (KAIST Department of Bio and Brain Engineering, Corresponding Author)
Meanwhile, this research was carried out with the support of the National Research Foundation of Korea's Basic Research Lab and Bio-medical Technology Development Project, and the KAIST-Ceragem Future Healthcare Research Center.
2025.09.19 View 4084 -
Bioengineers develop a new strategy for accurate prediction of cellular metabolic fluxes
A team of pioneering South Korean scientists has developed a new strategy for accurately predicting cellular metabolic fluxes under various genotypic and environmental conditions. This groundbreaking research is published in the journal Proceedings of the National Academy of Sciences of the USA (PNAS) on August 2, 2010.
To understand cellular metabolism and predict its metabolic capability at systems-level, systems biological analysis by modeling and simulation of metabolic network plays an important role. The team from the Korea Advanced Institute of Science and Technology (KAIST), led by Distinguished Professor Sang Yup Lee, focused their research on the development of a new strategy for more accurate prediction of cellular metabolism.
“For strain improvement, biologists have made every effort to understand the global picture of biological systems and investigate the changes of all metabolic fluxes of the system under changing genotypic and environmental conditions,” said Lee. The accumulation of omics data, including genome, transcriptome, proteome, metabolome, and fluxome, provides an opportunity to understand the cellular physiology and metabolic characteristics at systems-level. With the availability of the fully annotated genome sequence, the genome-scale in silico (means “performed on computer or via computer simulation.”) metabolic models for a number of organisms have been successfully developed to improve our understanding on these biological systems. With these advances, the development of new simulation methods to analyze and integrate systematically large amounts of biological data and predict cellular metabolic capability for systems biological analysis is important.
Information used to reconstruct the genome-scale in silico cell is not yet complete, which can make the simulation results different from the physiological performances of the real cell. Thus, additional information and procedures, such as providing additional constraints (constraint: a term to exclude incorrect metabolic fluxes by restricting the solution space of in silico cell) to the model, are often incorporated to improve the accuracy of the in silico cell.
By employing information generated from the genome sequence and annotation, the KAIST team developed a new set of constraints, called Grouping Reaction (GR) constraints, to accurately predict metabolic fluxes. Based on the genomic information, functionally related reactions were organized into different groups. These groups were considered for the generation of GR constraints, as condition- and objective function- independent constraints. Since the method developed in this study does not require complex information but only the genome sequence and annotation, this strategy can be applied to any organism with a completely annotated genome sequence.
“As we become increasingly concerned with environmental problems and the limits of fossil resources, bio-based production of chemicals from renewable biomass has been receiving great attention. Systems biological analysis by modeling and simulation of biological systems, to understand cellular metabolism and identify the targets for the strain improvement, has provided a new paradigm for developing successful bioprocesses,” concluded Lee. This new strategy for predicting cellular metabolism is expected to contribute to more accurate determination of cellular metabolic characteristics, and consequently to the development of metabolic engineering strategies for the efficient production of important industrial products and identification of new drug targets in pathogens.”
2010.08.05 View 24533