Six Papers Accepted at NeurIPS 2026, the World’s Most Prestigious AI Conference / Research Team Led by Professors Lee Hyungi and Kim Jangho (Department of Artificial Intelligence)
- 26.11.17 / 홍유민
A research team led by Professors Lee Hyungi and Kim Jangho from the School of Artificial Intelligence at Kookmin University (President Jeong Seung Ryul) has achieved the distinction of having a total of six papers accepted at “NeurIPS 2026 (Conference on Neural Information Processing Systems),” the world’s most prestigious international conference in the field of artificial intelligence. Notably, Kookmin University AI and Software Engineering Department master’s and doctoral students Hwang Sol-hee, Jeong Jin-woo, and Lee Hye-eun served as first authors on their respective papers and led the research. They demonstrated the research capabilities of Kookmin University’s graduate students by achieving outstanding results in the fields of inference efficiency optimization for large-scale vision-language models, low-bit quantization of Mamba-based super-resolution models, and personalization of large-scale language models.
NeurIPS, along with ICML and ICLR, is a leading international conference in the fields of artificial intelligence and machine learning, serving as an academic platform where researchers from universities and global companies around the world present the latest AI technologies and research findings. Notably, NeurIPS is recognized as one of the outstanding international conferences under the BK21+ program, and this achievement serves as a testament to the research competitiveness of Kookmin University’s Department of Artificial Intelligence.
The accepted papers cover a wide range of core AI research areas, including improving the inference efficiency of large-scale vision-language models (VLMs), low-bit quantization, data-efficient meta-learning, personalization of large language models (LLMs), accelerating generative models, and improving the training efficiency of vocabulary-expanded language models.
First, “Q-ViK: Question-Guided Visual KV Cache Eviction for Large Vision-Language Models” proposes a technique that predicts and selectively retains visual information critical to a question during the inference process of large-scale vision-language models. The study investigated methods to reduce memory usage and inference costs while preserving model performance to the greatest extent possible.
“Hardware-Friendly Token-Group Activation Quantization for Low-Bit Mamba Super-Resolution” is a study on low-bit quantization techniques for Mamba-based super-resolution models. It focused on enhancing the model’s computational efficiency and practical hardware applicability through hardware-friendly token-group-level activation quantization.
“Improving Neural Processes in the Low-Data Regime via Context-Subset Training and Self-Distillation” is a study on improving the performance of Neural Process models in environments with limited training data. By utilizing context-subset training and self-distillation, the study proposed a method for effective learning and prediction even with limited observational data.
“PriSM: Prior-guided Shared-basis Mixture Personalization for LLMs under Sparse User Histories” proposes a technique for effectively personalizing large language models even when user-specific data is insufficient. The key lies in utilizing a shared learning basis and prior information to construct personalized models with only a small amount of user history.
“ManiFusion: Unlocking High-Throughput Generation via Superposition in Manifold Space,” a joint study with Professor Kwak No-jun’s research team at Seoul National University, proposes a technique to increase the throughput of generative models. The study investigated a method to enhance generation efficiency while maintaining quality by processing multiple samples simultaneously through superposition in manifold space.
Finally, “SAVE: Sparsity-Aware Influence Estimation for Vocabulary-Expanded LLMs,” a joint study with Professor Lee Ju-ho’s research team at KAIST, proposed a technique to improve the training performance and transferability of large-scale language models with expanded vocabularies. By selecting effective additional training data through a sparsity-aware influence estimation technique, the study simultaneously enhanced performance on specific tasks and the ability to transfer knowledge to related tasks.
These achievements are significant in that they span a wide range of research areas, from lightweighting and optimizing large-scale AI models to addressing data scarcity, personalization, and accelerating generative models. In particular, they present practical research directions that go beyond simply improving the performance of individual models to enable the effective utilization of AI technology even under constraints of limited computational resources and data environments.
Through the acceptance of six papers at NeurIPS 2026, the research team led by Professors Lee Hyungi and Kim Jangho demonstrated its research competitiveness in various fields, including the optimization of AI model training and inference, personalization, and the advancement of generative models. In particular, the team achieved outstanding results through collaborative research with major domestic institutions such as Seoul National University and KAIST, and plans to continue contributing to the advancement of AI technology through active academic exchange and joint research.
By conducting a wide range of research, including predictive technologies utilizing intelligence-based surrogate models, the team has demonstrated its research competitiveness in the field of computational materials science.

▲ Professors Kim Jangho and Lee Hyungi
|
This content is translated from Korean to English using the AI translation service DeepL and may contain translation errors such as jargon/pronouns. If you find any, please send your feedback to kookminpr@kookmin.ac.kr so we can correct them.
|
|
Six Papers Accepted at NeurIPS 2026, the World’s Most Prestigious AI Conference / Research Team Led by Professors Lee Hyungi and Kim Jangho (Department of Artificial Intelligence) |
||||
|---|---|---|---|---|
|
2026-11-17
166
A research team led by Professors Lee Hyungi and Kim Jangho from the School of Artificial Intelligence at Kookmin University (President Jeong Seung Ryul) has achieved the distinction of having a total of six papers accepted at “NeurIPS 2026 (Conference on Neural Information Processing Systems),” the world’s most prestigious international conference in the field of artificial intelligence. Notably, Kookmin University AI and Software Engineering Department master’s and doctoral students Hwang Sol-hee, Jeong Jin-woo, and Lee Hye-eun served as first authors on their respective papers and led the research. They demonstrated the research capabilities of Kookmin University’s graduate students by achieving outstanding results in the fields of inference efficiency optimization for large-scale vision-language models, low-bit quantization of Mamba-based super-resolution models, and personalization of large-scale language models.
▲ Professors Kim Jangho and Lee Hyungi
|
||||






