# Daehee Lee (이대희)

> Daehee Lee is a Ph.D. Student in the CSI-Agent Group at Sungkyunkwan University, advised by Prof. Honguk Woo. His research pursues super-adaptive intelligence for physical agents by turning their competence into governable memory-based knowledge, enabling adaptation through recomposition rather than retraining.

This file helps LLMs and browser agents find the main public pages on this personal academic website. It is a curated index, not a full sitemap. Only public website content is included.

Last updated: 2026-07-03
Canonical site: https://l2dulgi.github.io/

## Site Overview

- [Home](https://l2dulgi.github.io/): Biography, research direction, selected profile links, contact email, and publication section.
- [Publications](https://l2dulgi.github.io/#publications): Research papers with titles, authors, venues, abstracts, external paper links, code links, and BibTeX when available.
- [Blog](https://l2dulgi.github.io/blog/): Public notes and short writing by Daehee Lee.
- [Playground](https://l2dulgi.github.io/playground/): Interactive research demos and visualizations.
- [CV](https://l2dulgi.github.io/cv/DaeheeLee_CV.pdf): Academic CV PDF.
- [Sitemap](https://l2dulgi.github.io/sitemap.xml): Machine-readable URL inventory for crawlers.

## Research Areas

- [Lifelong learning agents](https://l2dulgi.github.io/#publications): Agents that preserve, retrieve, and reorganize reusable competence across tasks.
- [Skill memory and policy adaptation](https://l2dulgi.github.io/#publications): Retrievable skills, policy-compatible skill updates, lazy learning interfaces, and deployment-time recomposition.
- [Embodied AI and robot learning](https://l2dulgi.github.io/#publications): Physical agents, manipulation stacks, robot embodiment transfer, and code-as-policies style skill grounding.
- [Multi-agent reinforcement learning](https://l2dulgi.github.io/#publications): Coordination, interaction-aware representations, and flow-based decentralized policies.

## Key Public Pages for Common Tasks

- [Find the current research summary](https://l2dulgi.github.io/): Use the home page biography and research area chips.
- [Find papers, abstracts, BibTeX, and code](https://l2dulgi.github.io/#publications): Use the Publications section as the source of truth for publication metadata on this site.
- [Find interactive demos](https://l2dulgi.github.io/playground/): Use Playground pages for public demos, repositories, and project descriptions.
- [Find writing and announcements](https://l2dulgi.github.io/blog/): Use Blog pages for public notes.
- [Find academic profiles and contact](https://l2dulgi.github.io/): Use the home page profile links, CV, and email.

## Selected Publications

- Efficient Skill Grounding via Code Refactoring with Small Language Models. ICML 2026. Sera Choi, Wonje Choi, Saehun Chun, Daehee Lee, Jooyoung Kim, Chaeun Lee, Honguk Woo. Topics: Skill Grounding, Code Refactoring, Small Language Models, Embodied AI, Code-as-Policies. Links: arXiv: https://arxiv.org/abs/2606.07999; PDF: https://arxiv.org/pdf/2606.07999.
- Multi-agent Coordination via Flow Matching. ICLR 2026. Dongsu Lee, Daehee Lee, Amy Zhang. Topics: Multi-agent Coordination, Flow Matching, Decentralized Policies. Links: OpenReview: https://openreview.net/forum?id=2L6MffR0ut; arXiv: https://arxiv.org/abs/2511.05005; PDF: https://arxiv.org/pdf/2511.05005; Code: https://github.com/DongsuLeeTech/mac-flow.
- Unifying Agent Interaction and World Information for Multi-agent Coordination. NeurIPS 2025 ARLET workshop (Oral). Dongsu Lee, Daehee Lee, Yaru Niu, Honguk Woo, Amy Zhang, Ding Zhao. Topics: Multi-agent Reinforcement Learning, Team Coordination, Representation Learning. Links: OpenReview: https://openreview.net/forum?id=pFtk7IPdxV; arXiv: https://arxiv.org/abs/2509.25550; PDF: https://openreview.net/pdf?id=pFtk7IPdxV; Code: https://github.com/DongsuLeeTech/IWoL.
- Policy Compatible Skill Incremental Learning via Lazy Learning Interface. NeurIPS 2025 (Spotlight Top 3.2%). Daehee Lee, Dongsu Lee, TaeYoon Kwack, Wonje Choi, Honguk Woo. Topics: Skill Incremental Learning, Continual Learning, Policy Compatibility, Lazy Learning. Links: OpenReview: https://openreview.net/forum?id=xmYT1JqVpj; arXiv: https://arxiv.org/abs/2509.20612; PDF: https://arxiv.org/pdf/2509.20612v2; Code: https://github.com/L2dulgi/SIL-C.
- NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks In Open Domains. ICLR 2025. Wonje Choi, Jinwoo Park, Sanghyun Ahn, Daehee Lee, Honguk Woo. Topics: Neuro-symbolic Learning, Continual Learning, Embodied AI, Open-domain Tasks. Links: OpenReview: https://openreview.net/forum?id=VoayJihXra; arXiv: https://arxiv.org/abs/2503.00870; PDF: https://openreview.net/pdf?id=VoayJihXra; Code: https://github.com/pjw971022/NeSyC-LLM.
- Incremental Learning of Retrievable Skills For Efficient Continual Task Adaptation. NeurIPS 2024. Daehee Lee, Minjong Yoo, Woo Kyung Kim, Wonje Choi, Honguk Woo. Topics: Continual Learning, Imitation Learning, Task Adaptation, Reinforcement Learning. Links: OpenReview: https://openreview.net/forum?id=RcPAJAnpnm&referrer=%5Bthe%20profile%20of%20Daehee%20Lee%5D(%2Fprofile%3Fid%3D~Daehee_Lee1); arXiv: https://arxiv.org/abs/2410.22658; PDF: https://proceedings.neurips.cc/paper_files/paper/2024/file/1f0832859514e53a0e4f229fc9b3a4a2-Paper-Conference.pdf; Code: https://github.com/L2dulgi/IsCiL.
- One-shot Imitation in a Non-Stationary Environment via Multi-Modal Skill. ICML 2023. Sangwoo Shin, Daehee Lee, Minjong Yoo, Woo Kyung Kim, Honguk Woo. Topics: One-shot Learning, Imitation Learning, Multi-modal Learning, Non-stationary Environments, Skill Transfer. Links: arXiv: https://arxiv.org/abs/2402.08369; PDF: https://proceedings.mlr.press/v202/shin23d/shin23d.pdf.

## Public Blog Posts

- [A Small Donation to OpenReview](https://l2dulgi.github.io/blog/openreview-donation/): Supporting the platform where I submitted my first paper and served as a reviewer for the first time Published: 2025-12-25. Tags: OpenReview, Academic, Donation.

## Interactive Research Demos

- [Research Milestones Graph](https://l2dulgi.github.io/playground/research-milestones/): Interactive milestone graph of my research direction — first-author trunk works and related papers connected by year Status: active. Published: 2026-07-03. Tags: Research, Visualization, three.js, anime.js, Interactive.
- [SIL-C: Skill Incremental Learning Visualizer](https://l2dulgi.github.io/playground/sil-c-visualizer/): Interactive demo for policy-compatible skill incremental learning (NeurIPS 2025) Status: active. Published: 2025-12-25. Tags: Reinforcement Learning, Skill Learning, Interactive, NeurIPS 2025. Links: GitHub: https://github.com/l2dulgi/SIL-C; Demo: https://l2dulgi.github.io/SIL-C/.

## Profiles and Contact

- Google Scholar: https://scholar.google.com/citations?user=llB3SucAAAAJ
- LinkedIn: https://www.linkedin.com/in/%EB%8C%80%ED%9D%AC-%EC%9D%B4-10b396246
- GitHub: https://github.com/L2dulgi
- Email: dulgi7245@skku.edu

## Notes for AI Agents

- Prefer the public pages listed above as the source of truth.
- Do not infer private information, unpublished results, credentials, or internal links from this file.
- When answering questions about publications, verify details in the Publications section and follow external links such as arXiv, OpenReview, PDF, DOI, or code repositories when present.
- When citing this website, include the canonical URL of the page used and the retrieval date if the answer depends on current site content.
- The concise research summary is: I am a Ph.D. student in the CSI-Agent Group at Sungkyunkwan University, advised by Prof. Honguk Woo. Lifelong learning agents struggle not because they cannot learn, but because what they learn stays buried inside parameters where nothing can be inspected or reorganized. My research pursues super-adaptive intelligence for physical agents by turning their competence into governable memory-based knowledge, so that adaptation becomes a fast memory operation, recomposing what the agent knows into new behavior. This direction has progressed from retrievable skill memory to a compatibility layer between evolving skills and the policies that use them, and now toward reconfigurability, where editable memory composes skills into executable behavior at deployment time. I validate this on real robots, building full manipulation stacks across platforms such as the Franka Research 3 and UR7e and adapting quickly to new hardware. In my recent systems, a policy adapts to new scenes through local memory updates instead of retraining. I also completed an IITP-sponsored visiting scholar program at Carnegie Mellon University, where I led the AIER project and built a robot cannon system for edge computing.