Taehyeon Kim

Taehyeon Kim (김태현)

Senior Member of Technical Staff

Agent reasoning · Reinforcement learning · Embodied AI

Holiday Robotics

Education

PhD — KAIST AI

BS — KAIST Mathematical Science

Minor: Intellectual Property

Bio

I am a Senior Member of Technical Staff at Holiday Robotics. I study how agents reason, choose actions, and improve from feedback, drawing on my research in efficient LLM inference, adaptive reasoning, and reinforcement learning.

At Holiday Robotics, I develop agent harnesses with planning, spatial perception, and act–verify loops, alongside harness learning with dense feedback. I also work on text-and-motion-conditioned RL policies, robot data flywheels, and embodied-agent benchmarks. Previously, I led agentic search and LLM alignment at LG AI Research, combining planning, tool use, retrieval, and post-training in a system for expert-domain tasks.

My published work spans adaptive thinking, instruction-guided decoding, speculative and parallel decoding, and knowledge distillation. I received my Ph.D. in AI from KAIST, advised by Prof. Se-Young Yun, and interned at Google Research, Dynamo AI, and Qualcomm AI.

I love coffee, running, and long walks at Seoul Forest, and I enjoy the Seongsu neighborhood culture. 커피를 좋아하고, 러닝과 서울숲 산책을 즐기며 성수동 문화를 좋아합니다.

Contact: kimtaehyeon610 [at] gmail [dot] com

Research Focus

My research connects efficient reasoning, agent execution, and learning from feedback, with a current focus on embodied agents. 효율적 추론과 에이전트의 행동, 피드백을 통한 학습을 연결하며, 현재는 이를 실제 환경에서 작업을 수행하는 embodied agent로 확장하고 있습니다.

Publications

† corresponding authors; * equal contribution.

Work Experience

Holiday Robotics

Senior Member of Technical Staff

–Present

Founding member, Embodied AI Team.
Developing harness planning, spatial perception, and act–verify loops, alongside meta-harness-style learning with dense feedback. Built an algorithm to automate text-and-motion-conditioned RL action-policy construction, using behavior-cloning regularization.
Data flywheel work includes grasping and video progress labeling for real-world and simulated robot data.
Building embodied-agent benchmarks (in progress).

LG AI Research

Superintelligence Lab

–

Tech Lead (Jan.–Jun. 2026)

Squad Lead (Sep.–Dec. 2025)

Research Scientist (Apr.–Aug. 2025)

Led 10+ researchers (6+ PhDs) on agentic search and LLM alignment. Built planning and tool-use workflows, retrieval systems, and RL/post-training methods.
Our expert-document search system achieved higher accuracy than Perplexity, Codex, and Anthropic systems in internal evaluations on this task.
Worked with Moontae Lee.

Google Research

Ph.D. Research Intern

–

Efficient LLM inference.
Worked with Adrian Benton and Ananda Theertha Suresh.

Dynamo AI

Ph.D. Research Intern

–

Privacy-preserving federated learning; worked with Eric Lin.

Qualcomm AI

Ph.D. Research Intern

–

Worked with Heesoo Myeong on efficient models for autonomous driving.

Honors

  • Technical Reviewer, ICML 2026 Gold · Top 25%
  • Google Conference Scholarship Dec. 2024 For NeurIPS 2024 multi-token prediction work.
  • Winner, Qualcomm AI Fellowship Oct. 2022 FINE Samples for Learning with Noisy Labels (NeurIPS '21).
  • NeurIPS Weather4Cast 4th · 2022 Project Lead · High-resolution precipitation forecasting from weather sequences.
  • NeurIPS Black-Box Optimization 8th · 2020 Project Lead · Bayesian optimization for AutoML and hyperparameter search.
  • NeurIPS MicroNet (CIFAR-100) 2nd & 3rd · 2019 Project Lead · Compact image classifiers for the CIFAR-100 track.

Leadership

  • Doctoral Student Representative, KAIST AI Mar. 2021-Feb. 2022
  • Lab Master, OSI Lab Mar. 2020-Feb. 2021
  • Vice Captain, KAIST ELKA Oct. 2014-Nov. 2015