👋 Welcome
I'm Taewhoo Lee, a PhD student at Korea University under the guidance of Professor Jaewoo Kang. I build large language models with strong context processing capabilities, grounded by concrete analysis of their inner mechanisms and thorough performance evaluation. My recent work centers on training self-improving search agents, interpreting how LLMs reason over implicit information, distilling domain-specific knowledge via fine-tuning, and evaluating how LLMs utilize given context.
I'm also focused on applying my research at production scale. I am currently an Applied Scientist Intern on the Alexa AI team at Amazon, studying efficient training methods that bridge the gap between what reasoning LLMs internally represent and how they act in agentic search scenarios. Previously, I was a research intern on the post-training team at Upstage, equipping Solar Open 2 and Solar Pro 4 with web search capabilities, and at AIGEN Sciences, exploring effective strategies for training biomedical tool-use agents.
🔬 Research Focus
My work centers on training and aligning large language models for effective context processing, grounded by concrete analysis of their inner mechanisms and thorough performance evaluation.
- Pre/Post-Training: reinforcement learning for developing self-aware search agents (SARL), long-horizon RL on a 250B-scale LLM (Solar Open 2), encoder pre-training for both classification and retrieval (UltraBERT), training frameworks for context compression (CompAct) and domain-specific knowledge distillation (Meerkat, OLAPH), outlier-safe pre-training for robust quantization (OSP)
- Evaluation: metric and benchmark for long-context tasks with high information coverage (ETHIC), tool-use tasks with multi-round, multi-party dialogues (DICE-BENCH)
- Analysis: interpreting the inner mechanisms of how LLMs process implicit information, under a challenging task of analogical reasoning (The Curious Case of Analogies)
📰 Latest News
New Preprint: SARL
This paper introduces Self-Assessment Reinforcement Learning (SARL), a framework that cultivates search agents with high search-state awareness, opening up opportunities for agents that improve themselves by recognizing their own knowledge gaps.
New Preprint: UltraBERT
This paper revisits core design choices in encoder pre-training and presents UltraBERT, the first encoder to achieve state-of-the-art performance on both classification and retrieval tasks. Check out the blog and model (10k+ downloads) for more details!
Career Update: Amazon
I joined Amazon as an Applied Scientist Intern on the Alexa AI team, working on agentic reinforcement learning. Specifically, my research will focus on bridging the gap between what reasoning LLMs internally represent and how they act in agentic search scenarios.
Career Update: Upstage
I joined Upstage's post-training team as a research intern, with a mission to equip Solar Open 2 and Solar Pro 4 with web search capabilities. Check out the technical report for more details!
1 Paper Accepted @ AAAI'26
This paper explores analogical reasoning in LLMs, revealing three distinct findings that help understand how models process implicit information.
2 Papers Accepted @ ACL'25
This paper proposes a pre-training strategy that prevents activation outliers from hindering 4-bit quantization.
This paper introduces a metric and benchmark for evaluating tool-use capabilities of LLMs under multi-round, multi-party dialogue situations.
Paper Accepted @ npj Digital Medicine'25
This paper presents Meerkat-7B, a medical language model that learns from real medical textbooks and synthetic datasets to outperform similar-scale models across medical exams, challenges, and expert evaluations.
Career Update: AIGEN Sciences
I joined AIGEN Sciences as a research intern, working on developing post-training strategies to train efficient and effective biomedical tool-use agents for real-world deployment.
1 Paper Accepted @ NAACL'25
This paper presents Information Coverage (IC), a metric that quantifies the proportion of input context necessary to answer a given query, and ETHIC, a unique long-context benchmark that requires high IC unlike other benchmarks.
1 Paper Accepted @ EMNLP'24
This paper presents CompAct, a post-training framework that employs an active strategy to condense extensive documents without losing key information.
New Preprint: OLAPH
This paper presents MedLFQA, a biomedical long-form QA dataset, and OLAPH, a post-training framework that leads to significant improvements in biomedical factuality.
Career Update: Korea University
I joined DMIS Lab (Dep. of Computer Science and Engineering, Korea University) as a M.S/Ph.D integrated student, advised by Professor Jaewoo Kang.
📝 Selected Publications
🤝 Professional Service
Primary Reviewer
- Acknowledged as a Silver Reviewer for ICML 2026!
- NeurIPS 2026 (ongoing), ICML 2026
Secondary Reviewer
- AAAI 2026, ACL 2025, NAACL 2025, EMNLP 2024