Analyzing the architecture of a domestically developed NPU Improving computational efficiency through software optimization Achieving a throughput of 821.1 tokens per second ▲ Graduate students from the Department of Electrical and Computer Engineering at Seoul National University took second place in the NPU Model Optimization Competition at the IEEE/ACM International Symposium on Microarchitecture (MICRO 2026), held in Athens, Greece, on November 1 (local time) A research team from Seoul National University took second place in an international competition focused on optimizing artificial intelligence (AI) models. According to Seoul National University on November 2, Jae-min Yoon and Jangwoon Park, master’s students at the Real-Time Operating Systems Laboratory (RTOSLab) in the Department of Electrical and Computer Engineering, placed second in the NPU Model Optimization Competition held at the IEEE/ACM International Symposium on Microarchitecture (MICRO 2026). The competition challenged participants to accelerate Google’s multimodal AI model “Gemma 4 12B” on “Renegade (RNGD),” a neural processing unit (NPU) developed by AI semiconductor company FuriosaAI. In the first round, more than 100 teams worked on optimizing three core operations of the AI model. The 64 teams that advanced to the second round improved the model’s overall performance in actual service conditions. The SNU RTOSLab team designed a new inference engine that processes requests from multiple users simultaneously. The team minimized waiting time by preparing the next task while the NPU was executing the current one and improved processing efficiency by reducing computational precision while maintaining accuracy. Through these methods, the team improved the inference engine’s performance more than fivefold, ultimately achieving a throughput of 821.1 tokens per second. Professor Seongsoo Hong of the Department of Electrical and Computer Engineering at Seoul National University emphasized, “The service performance of AI semiconductors depends heavily on software, not just the chips themselves. This achievement is meaningful because the team directly analyzed the hardware architecture of a domestically developed AI semiconductor and improved its performance by optimizing the software accordingly.” Source: https://m.sedaily.com/article/20109132 Translated by: JungHwan Kim, English Editor of the Department of Electrical and Computer Engineering, onzdavid@snu.ac.kr...

▲ (From left) Professor Yongtaek Hong, Professor Jonghyun Choi, and Professor Se Young Chun of the Department of Electrical and Computer Engineering at Seoul National University. Two research projects, one by the joint research team of Professors Yongtaek Hong and Jonghyun Choi and the other by Professor Se Young Chun’s research team in the Department of Electrical and Computer Engineering at Seoul National University, have been selected for Samsung’s Research Funding for Future Technology Program for the second half of 2026. The two teams will address key challenges in robotic sensing and artificial intelligence (AI) memory, respectively. By studying robots that independently adjust their actions based on tactile information and AI that selectively removes unnecessary memories, they plan to develop next-generation AI technologies that respond flexibly to real-world environments. In particular, the joint project led by Professors Hong and Choi was the only project selected in the program’s next-generation robotics category. [Robots That Adjust Their Grip When Objects Slip and Let Go When They Are Hot: Developing Robots with Tactile Physical AI] The joint research team led by Professors Hong and Choi will study next-generation robotic AI that enables robots to touch and feel objects and independently adjust their actions, much like humans. Even without looking, people adjust their grip when a cup slips from their hand and reflexively withdraw their hand upon touching a hot object. They assess situations and change their actions based on sensory information obtained through their skin. Robots, on the other hand, have made significant advances in recognizing their surroundings and planning tasks using cameras and AI. However, when grasping objects or assembling components, the contact area between the hand and the object is often hidden from the camera. This makes it difficult to accurately detect slight slipping, abnormal forces, or dangerous temperature changes. These “sensory blind spots” can lead to task failures or component damage when assembling small parts or handling unfamiliar objects. They remain a major limitation to robots performing tasks with human-like precision in diverse environments. To address these challenges, the team will develop high-resolution electronic skin that deforms flexibly like human skin while precisely detecting pressure, temperature, and heat flow. The team will also study AI technology that combines tactile information from the electronic skin with camera images and language information, enabling robots to decide how to act independently. The key is to combine Professor Hong’s flexible and stretchable sensor technology with Professor Choi’s robotic AI technology. The team plans to develop technology that allows robots to adjust their grip when they detect an object slipping, independently release hot components to prevent damage, and modify their posture and movements when unexpected resistance occurs during assembly. The team will also test whether robots can understand contact conditions hidden from view and respond appropriately, even when handling objects they have not previously been trained on or working in unpredictable manufacturing environments. The goal is to establish a foundation for core intelligence technologies for both industrial robots performing precision manufacturing and assembly tasks and humanoid robots handling diverse objects safely and delicately. [Should AI Forget to Become Smarter? Developing AI with “Selective Forgetting”] Professor Chun’s research team will study AI capable of “selective forgetting,” identifying and removing unnecessary memories. Rather than remembering everything, the human brain lets unimportant details fade while selecting and retaining memories that will be useful in the future. Research in memory science has suggested that selective forgetting reduces interference from outdated information and supports generalization by preserving the essential elements of experiences. In contrast, AI research has largely focused on retaining as much learned information as possible without forgetting it. However, AI lacks the ability to independently decide what to forget. As information accumulates, important clues can become buried, or existing capabilities can deteriorate while the AI learns new information. The team will investigate the hypothesis that selectively removing unnecessary memories can improve AI performance. It plans to develop technology that enables AI to assess the future usefulness of its memories, retain those that are needed, and remove those that are unnecessary. The technology will be progressively evaluated in multimodal AI agents that understand both images and language and sustain extended conversations and tasks. Through this research, the team aims to lay the groundwork for AI assistants that maintain context throughout lengthy tasks and on-device AI that operates with less memory and power. Media Coverage External Link: Edaily: https://www.edaily.co.kr/News/Read?newsId=03670326645612264&mediaCodeNo=257 Seoul Economic Daily: https://www.sedaily.com/article/20101652?ref=naver The Robot News: https://www.irobotnews.com/news/articleView.html?idxno=48946 VERITAS: https://www.veritas-a.com/news/articleView.html?idxno=631374 Source: https://ece.snu.ac.kr/ece/news?md=v&bbsidx=58181 Translated by: JungHwan Kim, English Editor of the Department of Electrical and Computer Engineering, onzdavid@snu.ac.kr...

Late Chairman Jeong-sik Kim Selected as a Recipient of the 36th Proud SNU Alumni Award ▲ The late Jeong-sik Kim, former Chairman of Daeduck Electronics Co., Ltd. and former Chairman of the Haedong Science and Culture Foundation. In October 2026, the late Chairman Jeong-sik Kim, who graduated from the Department of Communication Engineering at Seoul National University’s College of Engineering in 1956, was selected as a recipient of the 36th Proud SNU Alumni Award. In February 2019, Chairman Kim donated his personal funds toward the construction and operation of the Haedong Advanced Engineering Center to foster interdisciplinary talent and advance artificial intelligence (AI), a key field in the Fourth Industrial Revolution. His donations to his alma mater, including the Haedong Academic Building Construction Fund, the Jeong-sik Kim Development Fund, and the Daeduck Electronics Academic Fund, totaled KRW 65.7 billion, the largest individual donation in Seoul National University’s history. In particular, his full support for the establishment of the Haedong Advanced Engineering Center, an AI research and education platform at Seoul National University, laid the foundation for interdisciplinary research and the development of future talent across engineering fields, including robotics, semiconductors, energy, and biotechnology. These contributions have been highly recognized for advancing national science and technology and strengthening industrial competitiveness. Seoul National University’s College of Engineering deeply honors Chairman Kim’s lifelong dedication to his alma mater and society and his commitment to giving. Carrying forward his noble spirit, the college will further strengthen its efforts to nurture creative engineering talent who will lead the future. Source: https://ece.snu.ac.kr/ece/news?md=v&bbsidx=58179 Translated by: JungHwan Kim, English Editor of the Department of Electrical and Computer Engineering, onzdavid@snu.ac.kr...

▲ Professor Kyomin Jung and researchers from the Department of Electrical and Computer Engineering at Seoul National University. (Top row, from left) Professor Kyomin Jung, Nakyeong Yang, Minbeom Kim, Yerin Hwang, and Yunah Jang. (Bottom row, from left) Seongho Joo, Minha Jhang, Kyeongman Park, Junseok Kim, and Dohyung Kim. Professor Kyomin Jung’s research team in the Department of Electrical and Computer Engineering at Seoul National University has had a total of 44 papers accepted at leading international artificial intelligence (AI) conferences over two years: 21 in 2025 and 23 in 2026. The Machine Intelligence Lab (MILAB), led by Professor Jung, has conducted extensive research on reasoning and training efficiency in large language models (LLMs), consistency evaluation of AI-generated images, and AI agent safety. In 2026, the team had nine papers accepted at EMNLP and five at ACL, two leading international conferences in natural language processing. It also had one paper accepted at NeurIPS, three at ICML, and one at ICLR, major international conferences in machine learning. With two additional papers accepted at EACL and one each at AAAI and ECCV, the team achieved a total of 23 accepted papers this year. In 2025, the team had 21 papers accepted: nine at EMNLP, six at NAACL, two at ACL, and one each at NeurIPS, ICLR, CIKM, and COLING. This marks two consecutive years with more than 20 accepted papers annually. Among the representative studies, “Evaluating Visual Narrative Coherence in Story Visualization via Diversified Storylines,” by Minha Jhang and colleagues, was selected for an oral presentation at ACL 2026. The study proposed a method for evaluating visual narrative coherence across sequences of AI-generated images. It comprehensively assesses not only whether individual images match their descriptions, but also whether characters and backgrounds remain consistent across scenes. In experiments comparing AI evaluations with human judgments, the method achieved approximately 95% of the agreement level observed between human evaluators. The technology is expected to help identify inconsistencies between scenes in AI-generated picture books and story content, reducing the burden of manual review for creators. “Beyond Normalization: Rethinking the Partition Function as a Difficulty Scheduler for RLVR,” by Dohyung Kim and colleagues, was accepted at ICML 2026. The study proposed “PACED-RL,” a reinforcement learning technique that efficiently improves AI’s mathematical reasoning and code generation capabilities. It uses the partition function, traditionally used mainly to normalize computed values, to estimate the difficulty of individual problems and concentrate computational resources on problems that are beneficial for learning. In experiments, accuracy on a mathematical reasoning evaluation increased from 23.9% with the existing method to 28.7%. In a code generation evaluation, the technique reduced the training time needed to reach the existing method’s peak performance by approximately half. By reducing the time and cost of AI training, the technology is expected to ease the burden of AI development for university laboratories and small and medium-sized enterprises. The team also achieved results in improving AI agent safety. “LiSA: Lifelong Safety Adaptation via Conservative Policy Induction,” by Minbeom Kim and colleagues, was accepted at NeurIPS 2026. The technology continuously improves AI agents’ safety judgments based on errors reported by users. It derives rules for privacy protection and task execution from cases of incorrect judgment, then selects and applies rules supported by sufficient evidence. This allows AI agents to respond to organization-specific security requirements and emerging risks without retraining the AI model. In experiments, LiSA distinguished safe actions from unsafe ones more accurately than existing methods. It also maintained better performance than the comparison methods even when 20% of user feedback contained incorrect information. The technology is expected to help prevent personal information leaks from enterprise AI assistants and improve the safety of workflow automation. ▲ How LiSA improves AI agent safety. The technology derives general safety policies and context-specific rules from user-reported errors, verifies the reliability of the general policies, and uses them to make safety judgments. The team is also improving the reliability and practical applications of AI through research on personalized guidance in long-term conversations, privacy-preserving language generation, and the evaluation of AI’s ability to follow instructions. The team has also achieved results in training researchers. Dr. Yerin Hwang, who earned her Ph.D. in August 2026, was appointed as a professor in the Department of Computer Science and Artificial Intelligence at Dongguk University in September of the same year. Dr. Minbeom Kim, who completed his doctoral studies during the same period, received the Best Ph.D. Dissertation Award from Seoul National University’s Interdisciplinary Program in Artificial Intelligence and is currently working at the SK hynix AI Research Center. Professor Jung, who supervised the research, stated, “At a time when language-processing-based AI research, including LLMs, is becoming increasingly important, it is encouraging that we have been able to lead research by actively presenting papers at the world’s leading conferences over several years. Through creative and challenging research, we will continue contributing to the AI research capabilities of SNU’s College of Engineering and Korea’s academic and industrial communities.” The research introduced in this article was supported by the National Research Foundation of Korea (NRF) and the Institute of Information & Communications Technology Planning & Evaluation (IITP), with funding from the Ministry of Science and ICT. [References] Laboratory website: https://milab.snu.ac.kr/ Representative Paper 1: Evaluating Visual Narrative Coherence in Story Visualization via Diversified Storylines (ACL 2026, Oral) DOI: https://doi.org/10.18653/v1/2026.acl-long.1578 Representative Paper 2: Beyond Normalization: Rethinking the Partition Function as a Difficulty Scheduler for RLVR (ICML 2026) DOI (arXiv): https://doi.org/10.48550/arXiv.2602.12642 Representative Paper 3: LiSA: Lifelong Safety Adaptation via Conservative Policy Induction (NeurIPS 2026) DOI (arXiv): https://doi.org/10.48550/arXiv.2605.14454 [Contact] Professor Kyomin Jung, Department of Electrical and Computer Engineering, Seoul National University / +82-2-880-1937 / kjung@snu.ac.kr Source: https://ece.snu.ac.kr/ece/news?md=v&bbsidx=58178 Translated by: JungHwan Kim, English Editor of the Department of Electrical and Computer Engineering, onzdavid@snu.ac.kr...

The teams of researchers—Sarang Lee, Seohee Jung, Se.ohyeon Kwak, and Seungjae Lee; Jongwha Kim, Yuhwan Shin, Hojun Jeon, and Minho Jeon; and Myeongho Han, Jaeho Kim, Hyunjun Song, and Sijin Lee—from the Integrated Circuits and Systems Laboratory (ICSL), led by Professor Jaehyouk Choi, received the Minister of Trade, Industry and Energy Award and two Corporate Special Awards from MathWorks Korea and SK Hynix, respectively, at the 27th Korea Semiconductor Design Competition. Minister of Trade, Industry and Energy Award: A W-Band Fully Symmetric Quadrature Local Oscillator Generator with a Fast Quadrature Calibration Technique Corporate Special Award (MathWorks Korea): A Low-Power Direct-Frequency-Tracking Digital CDR Supporting Per-Lane Data-Rate Variation and Deep-SSC Tolerance in AI-Infrastructure Multi-Lane Interconnects Corporate Special Award (SK Hynix): Next-Generation HBM Interface Clock Stabilization Technology Using a Strobe-Triggered Digital Regulator and Local Replica Regulators Source: https://ece.snu.ac.kr/ece/news?md=v&bbsidx=58142 Translated by: JungHwan Kim, English Editor of the Department of Electrical and Computer Engineering, onzdavid@snu.ac.kr...

Professor Jung-Ik Ha of the Department of Electrical and Computer Engineering at Seoul National University received the 36th Science and Technology Excellent Paper Award on September 30. Established by the Korean Federation of Science and Technology Societies (KOFST), the Science and Technology Excellent Paper Award is presented to scientists and engineers who publish outstanding papers in domestic academic journals through creative research. Each year, papers nominated by academic societies are selected through an in-depth evaluation of their creativity and academic contributions, and the award is presented to the principal investigator of each selected paper. In the award-winning paper, Professor Ha proposed a new control method for unbalanced power systems. Unlike conventional approaches that use phase voltages, the proposed method uses line-to-line voltages to improve power-control performance during transient conditions. Source: https://ece.snu.ac.kr/ece/news?md=v&bbsidx=58140 Translated by: JungHwan Kim, English Editor of the Department of Electrical and Computer Engineering, onzdavid@snu.ac.kr...

▲ Participants at the launch ceremony for the “K-AI Network Competitiveness Enhancement” project, held on October 1st at the conference room of the New Media Communications Joint Research Institute at Seoul National University. The National Information Society Agency (NIA) and the Ministry of Science and ICT (MSIT) held a launch ceremony for the “K-AI Network Competitiveness Enhancement” project at Seoul National University on October 1st, initiating next-generation AI network demonstrations. For the project’s core initiative, the “Campus AI Network Demonstration,” KRW 2.2 billion in government funding will be invested in 2026. NewGens will serve as the lead organization, with Seoul National University, KAIST, Yonsei University, and HFR participating in the project through December 2027. The consortium will establish AI-RAN demonstration environments based on domestically developed equipment and solutions across the campuses of Seoul National University, KAIST, and Yonsei University. The three universities will be connected through the Korea Advanced Research Network (KOREN) to verify RAN performance and interoperability, as well as demonstrate services in the fields of physical AI and robotics. Seoul National University will demonstrate an integrated campus guidance service using AI hologram technology, while KAIST will demonstrate an on-campus transportation and delivery service based on autonomous driving robots. Yonsei University will develop an indoor autonomous patrol and escort guidance service using multimodal quadruped robots. Through the project, NIA plans to identify models for optimizing RAN performance using AI and demonstrating physical AI services, thereby supporting the wider adoption of domestic AI network technologies in industry and laying the groundwork for overseas expansion. At the launch ceremony, NIA and the National Robotics Research Institute at Seoul National University declared a vision for collaboration between AI-RAN and physical AI. Participants also discussed the potential for linking AI-RAN demonstration environments with physical AI research, including robotics. NIA President Hyeong-Cheol Kim stated, “By connecting the research capabilities of universities with the technological capabilities of industry, we will strengthen the competitiveness of domestic AI-RAN technology and support the expansion of demonstration results accumulated on university campuses into industries and overseas markets.” Source: https://www.joongang.co.kr/article/25466687 Translated by: JungHwan Kim, English Editor of the Department of Electrical and Computer Engineering, onzdavid@snu.ac.kr...

Professor Emeritus Jae Hong Lee of the Department of Electrical and Computer Engineering at Seoul National University received the 2026 Stuart F. Meyer Memorial Award from the IEEE Vehicular Technology Society in September. The award was presented at the IEEE 104th Vehicular Technology Conference: VTC2026-Fall, held in Boston, USA, in recognition of his outstanding and exemplary contributions to the advancement of wireless technology and science, as well as his service to the development of the Society. Source: https://ece.snu.ac.kr/ece/news?md=v&bbsidx=58117 Translated by: JungHwan Kim, English Editor of the Department of Electrical and Computer Engineering, onzdavid@snu.ac.kr...