Prev: Google Research (NYC), Qualcomm AI, Dynamo AI (YCW22) LG AI Research Scientist @Superintelligence LabEducation
PhD - KAIST AI BS - KAIST Mathematical Science[Minor-Intellectual Property]
I am a Research Scientist at LG AI Research, focusing on thinking and reasoning strategies for LLMs, including scalable test-time inference and model behavior optimization. I received my Ph.D. in Optimization & Statistical Inference (OSI) Lab @ KAIST, advised by Prof. Se-Young Yun. I worked as a PhD Intern @ Google Research, Qualcomm AI, and DynamoFL (YCW22). My expected graduation date is Feb. 2025.
Contact: potter32 [at] kaist.ac.kr, kimtaehyeon610 [at] gmail.com (permanent)
Contact: potter32 [at] kaist.ac.kr, kimtaehyeon610 [at] gmail.com (permanent) Short Bio
Taehyeon Kim is a Research Scientist at LG AI Research, where he focuses on thinking and reasoning strategies for large language models (LLMs), including scalable test-time inference and model behavior optimization. He received his Ph.D. in AI from the Korea Advanced Institute of Science and Technology (KAIST), advised by Prof. Se-Young Yun in the OSI Lab.
During his Ph.D., he gained diverse research and engineering experience through collaborations with Google Research (2023), Dynamo AI (2023), the Korean National Institute of Meteorological Sciences (2022), and Qualcomm AI (2021). His work has been recognized with spotlight and oral presentations at top venues (e.g., ICLR, ICML Workshops), as well as several NeurIPS competition awards and leadership roles.
Taehyeon’s research has centered around efficient and effective LLM inference, with contributions to speculative decoding, instructive decoding, and collaborative inference frameworks. His broader work spans real-world challenges including distributed optimization for heterogeneous data, semi-supervised federated object detection, automated hyperparameter search, instruction-following alignment, and domain-specific forecasting for weather prediction in South Korea.
Looking ahead, he is particularly interested in collaborative decoding among multiple LLMs via prompt optimization, client-centric test-time adaptation incorporating user preferences, and fast, scalable reasoning under compute and latency constraints. His research integrates both theoretical insights in matrix analysis and practical algorithmic design, enabling robust performance across diverse application domains.
(Click to Open) → Featured Publications (1st Authored)
CV (Updated: Dec, 2024)
️ News
Apr. 2025
Feb. 2025
Successfully received my PhD!Oct. 2024
1 Accepted @ NeurIPS2024W: Speculative Decoding with multiple drafters 1 Accepted @ TMLR 2024: Federated Learning with Noisy Labels Google Conference Scholarship - NeurIPS 2024Sep. 2024
Successfully passed my PhD Proposal! 1 Accepted @ EMNLP2024 Main - Specialized Speculative Decoding! 2 Accepted @ NeurIPS 2024 - Speculative Decoding and Block Transformer!May. 2024
Attending ICLR 2024 @ Vienna, Aus Jan. 2024
1 Accepted @ ICLR2024 (Spotlight): Instruction Following on Large Language Model Working at/with
Table
Search
Publications & Technical Reports
Table
List
Default view
Search
Preprints (Under Review)
Leadership Awards Activities
List view
Search
Research Projects
Table
List view
Search
Invited Talks
Table
Search
Services & Others
Gallery view
Search