Published in Under Review, 2025
Taero Kim, Hoyoon Byun, Youngjun Choi, Sungrae Park, Kyungwoo Song
We propose MIDUS, a depth up-scaling method for large language models that infuses memory mechanisms into newly inserted layers. MIDUS enables efficient model depth expansion while preserving pretrained knowledge and improving parameter utilization.
Published in Efficient Reasoning Workshop at Neural Information Processing Systems (NeurIPS), 2025
Jung Lee*, Taero Kim*, Nikhil Verma (* equal contribution)
We propose Multi-linguistic LoRA Merging (MLM), a method that merges language-specific LoRA adapters to efficiently transfer multilingual capability into a single model without full retraining, enabling scalable multilingual large language model deployment.
Published in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025
Taero Kim, Subeen Park, SungJun Lim, Yonghan Jung, Krikamol Muandet, Kyungwoo Song
We identify that existing invariant learning methods are overly conservative—they discard predictive spurious features to ensure invariance. We propose Sufficient Invariant Learning (SIL), which leverages flatness-aware optimization to learn features that are both invariant and maximally predictive across environments, improving out-of-distribution generalization.
Published in CVPR 2025 Workshop on Domain Generalization: Evolution, Breakthroughs and Future Horizon, 2025
Taero Kim, Seonggyun Lee, Joonseong Kang, Youngjun Choi, Wonsang Yun, Nicole Hee-Yeon Kim, Ziyu Chen, Lexing Xie, Kyungwoo Song
We introduce IMC, a benchmark that exposes fundamental failure modes of existing domain generalization methods under multiple causally-intertwined spurious correlations—a realistic but underexplored setting that current benchmarks fail to capture.
Published in Computers in Biology and Medicine, 2024
Taero Kim, Hyeonjeong Lee, Minseong Kim, Kwang-Yong Kim, Kyu Hyung Kim, Kyungwoo Song
We propose GloGen, a global PPG prompt generation framework for few-shot transfer learning in blood pressure estimation. By generating subject-adaptive prompts from a small number of labeled samples, GloGen enables rapid personalization of blood pressure models with minimal calibration data.
Published in Biomedical Signal Processing and Control, 2025
Sungjun Lim*, Taero Kim*, Hyunjung Lee*, Yewon Kim, Minhoi Park, Minseung Kim, Kyuhyung Kim, Kyungwoo Song (* equal contribution)
We apply distributionally robust optimization to PPG-based blood pressure estimation to address data imbalance and distribution shift across subjects. Our approach improves estimation robustness and generalization to underrepresented physiological conditions.
