2026
TaRA: Training-Aware Low-Rank Adaptation Initialization
Taehyeon Kim, Eunhyeok Park
The Conference on Empirical Methods in Natural Language Processing (EMNLP) 2026 Accepted
TaRA is a training-aware LoRA initialization that matches combined low-rank adapter gradients to full-rank gradients, improving optimization and consistently outperforming prior SVD-based initializations with minimal overhead.
TaRA: Training-Aware Low-Rank Adaptation Initialization
Taehyeon Kim, Eunhyeok Park
The Conference on Empirical Methods in Natural Language Processing (EMNLP) 2026 Accepted
TaRA is a training-aware LoRA initialization that matches combined low-rank adapter gradients to full-rank gradients, improving optimization and consistently outperforming prior SVD-based initializations with minimal overhead.
MOB-KV: Mixture-of-Basis for Key-Value Cache Compression
Jinseop Yeom*, Taehyeon Kim*, Eunhyeok Park (* equal contribution)
The Conference on Empirical Methods in Natural Language Processing (EMNLP) 2026 Accepted
MOB-KV makes adaptive low-rank KV cache compression practical by replacing costly online SVD with offline key bases, combining low-rank key compression and value quantization for strong memory-accuracy-latency trade-offs.
MOB-KV: Mixture-of-Basis for Key-Value Cache Compression
Jinseop Yeom*, Taehyeon Kim*, Eunhyeok Park (* equal contribution)
The Conference on Empirical Methods in Natural Language Processing (EMNLP) 2026 Accepted
MOB-KV makes adaptive low-rank KV cache compression practical by replacing costly online SVD with offline key bases, combining low-rank key compression and value quantization for strong memory-accuracy-latency trade-offs.
2024
Experiments and analysis of large language model's performance on code clone detection for obfuscated code
Taehyeon Kim*, Yujin Ahn*, Goun Pyeon*, Sungho Lee (* equal contribution)
Korea Software Congress 2024 Accepted
Evaluated how well large language models detect code clones in obfuscated source code using POJ104 augmented with representative obfuscation techniques.
Experiments and analysis of large language model's performance on code clone detection for obfuscated code
Taehyeon Kim*, Yujin Ahn*, Goun Pyeon*, Sungho Lee (* equal contribution)
Korea Software Congress 2024 Accepted
Evaluated how well large language models detect code clones in obfuscated source code using POJ104 augmented with representative obfuscation techniques.
2023
Landmark-guided clustering for predicting time-based public bicycle demand in daejeon city
Chaeeun Noh*, Sujin Noh*, Taehyeon Kim*, Eunbin Kim*, Hyunsu Mun, Youngseok Lee (* equal contribution)
Korea Software Congress 2024 Accepted
Landmark-based station clustering, which incorporates spatial characteristics, improved public bike usage prediction on Daejeon Tashu data, achieving the best regression performance
Landmark-guided clustering for predicting time-based public bicycle demand in daejeon city
Chaeeun Noh*, Sujin Noh*, Taehyeon Kim*, Eunbin Kim*, Hyunsu Mun, Youngseok Lee (* equal contribution)
Korea Software Congress 2024 Accepted
Landmark-based station clustering, which incorporates spatial characteristics, improved public bike usage prediction on Daejeon Tashu data, achieving the best regression performance