Eight papers by a research team led by Professor Jungwoo Lee of the Department of Electrical and Computer Engineering at Seoul National University College of Engineering have been accepted to NeurIPS 2026, a leading international conference in artificial intelligence and machine learning, and EMNLP 2026, a major international conference in natural language processing.
The team presented methods for reducing the impact of environmental noise when robots learn actions from video, as well as technology for generating more diverse synthetic training data for large language models (LLMs). The studies aim to improve AI performance and efficiency across a wide range of fields, including physical AI, reinforcement learning, the training, inference, and code-generation capabilities of large language models, and quantum AI.
Of the accepted papers, five are scheduled to be presented at NeurIPS 2026, to be held in Sydney, Australia, from December 6 to 12, while the remaining three will be presented at EMNLP 2026, to be held in Budapest, Hungary, from October 24 to 29. The research involved members of Professor Lee’s Cognitive Machine Learning Lab and researchers from AI platform company HodooAI, among others.
Among the studies accepted to NeurIPS 2026, “Why Latent Actions Fail, and How to Prevent It” addresses a method for reducing the influence of environmental factors unrelated to robot actions in video-based robot action learning.
When a robot learns actions from video, it can learn not only the movements of objects but also changes in the background and movements in the surrounding environment. When such action-irrelevant information becomes mixed into the learning process, it becomes more difficult for the robot to accurately identify the actions it is actually supposed to learn. The researchers theoretically analyzed this problem and proposed a learning method that reduces the influence of unnecessary noise in the latent space, where information from video is represented in compressed form. By applying an additional loss function that guides learning in the desired direction, the researchers improved learning performance by 50–70% compared with existing methods.
The findings are expected to help robots learn action-relevant information more reliably across diverse environments. In the future, the approach could also be applied to research aimed at advancing the intelligence of humanoid robots that must operate in complex environments, including domestic service robots and caregiving robots.
In the field of reinforcement learning, “MINT: Meeting-time INdicators for Truncation in Multi-Step Off-Policy RL,” with Seungyub Han as first author, was also accepted. Reinforcement learning is a technology that enables AI to learn better actions through trial and error. The researchers proposed a method for reducing variability in the target reward function during a learning process that uses multi-step experience, improving on the theoretical limitations of existing learning methods.
In the field of large language models, “Mechanistic Circuit Identification for Controllable Data Generation” is a study that increases the diversity of synthetic training data by analyzing the internal mechanisms of AI models.
As securing high-quality training data becomes increasingly important, synthetic data technologies—in which AI generates new training data itself—are attracting growing attention. However, if generated data becomes concentrated around similar content and expressions, its usefulness for training can be limited.
The researchers analyzed the characteristics of synthetic data from three perspectives: learnability, difficulty, and alignment with the learning objective. They then identified the internal information-processing pathways of the model associated with each characteristic and controlled data generation by strengthening the desired characteristics. As a result, the semantic diversity of the generated data increased by more than 10% compared with existing methods. The technology is expected to help secure training data for large language models and improve the training efficiency and performance of domestic large language models and AI foundation models.
Other research accepted to NeurIPS 2026 proposed a method for efficiently determining training configurations for large language models by using training information from smaller models. In the field of quantum AI, the researchers proposed a method for optimizing quantum circuits that compress quantum states and, for the first time, provided a theoretical basis for the recoverability of compressed quantum states.
The three papers accepted to EMNLP 2026 focus on improving the reasoning and code-generation performance of large language models. The researchers studied methods for refining generated answers and using pause tokens during the reasoning process, as well as a method for improving code-generation accuracy by incorporating function-level code execution results into training.
The research team plans to apply the related patented technologies to HodooAI’s AI platform, where they will be used as core technologies for physical AI and natural language processing-based large language model systems.
Professor Lee, whose recent research has focused on physical AI and quantum AI, previously founded HodooAI, a natural language processing and physical AI company, together with five former students from his laboratory. The accepted papers were co-authored by researchers Seungyub Han, Jung Min Lee, Nakyung Lee, Jimin Seo, Hyunho Cha, Minhae Oh, Suhwan Kim, Jaehyeon Kim, and Idris Nechnech. They are currently focusing on physical AI and large language model research and plan either to pursue follow-up research in academia or conduct research at global corporate research laboratories. Other collaborators included researchers Taehyun Cho, Dohyeong Kim, Kyungjae Lee, Li Zhao, Sangwoo Hong, Chae-Yeun Park, Gyubok Lee, Yeonsik Jo, Kiwoong Yoo, Yeongoon Kim, Jin Woo Koo, Minsik Seol, Giho Lee, and Sehwan Kim.
Seungyub Han, first author of the MINT paper, said, “I believe this work will play an important role in bringing SNU Engineering’s research capabilities in reinforcement learning theory to a world-class level.” He added, “I expect that it will help improve the performance and safety of physical AI and large language models in the future.” Han plans to focus on research applying reinforcement learning technologies to physical AI.
Professor Jungwoo Lee, who supervised the research, said, “I am delighted that a total of eight papers have been accepted to NeurIPS 2026 and EMNLP 2026, two of the world’s leading AI conferences.” He added, “We will continue to develop innovative technologies and contribute to strengthening SNU Engineering’s AI research capabilities and raising the technological level of Korean AI startups.”
The research was supported by projects funded by the National Research Foundation of Korea and the Institute for Information and communications Technology Promotion.
□ Introduction to the SNU College of Engineering
Seoul National University (SNU) founded in 1946 is the first national university in South Korea. The College of Engineering at SNU has worked tirelessly to achieve its goal of ‘fostering leaders for global industry and society.’ In 12 departments, 323 internationally recognized full-time professors lead the development of cutting-edge technology in South Korea and serving as a driving force for international development.
Computational simulation/modeling
Not applicable
Professor Jungwoo Lee is a co-founder of HodooAI, whose researchers participated in the studies. The research team plans to apply related patented technologies to HodooAI's AI platform. No additional author-provided conflict-of-interest statement is available.