Hyunin Lee

Ph.D at UC Berkeley

Contact: hyunin(at)berkeley(dot)edu

Research Interests: Reinforcment Learning, Recursive Self-Improvement

About

I am Hyunin Lee, a final-year Ph.D. candidate at UC Berkeley. My research background is in reinforcement learning, and I am currently investigating how reinforcement learning can enable recursive self-improvement (RSI) in AI systems.

I am fortunate to work with Somayeh Sojoudi. During my Ph.D., I have also had the opportunity to conduct research at Meta (ads ranking team), OpenAI (paperbenchmark project), and Sakana AI, working on topics spanning large language models, AI agents, and reinforcement learning. Prior to joining Berkeley, I received my B.S. degree from Seoul National University where I worked on research projects in neuroscience and machine learning with Yong-Lae Park.

Besides my research, I am interested in AI education 📚👩🏻‍💻. I co-founded and currently advise OUTTA.

Blog

• An Orthogonal Alignment Phenomenon in Cross-Attention

News

• [2026.06] A new sakanaAI model; Fugu is out! Try the model: technical report, blog.
• [2026.05] Joining SakanaAI as research intern (manager: Yujin Tang). I will be in Japan!
• [2026.01] My research will be further funded by Meta (RL for multimodal LLM).
• [2025.10] New paper on Multi-Modal Recommendation (work done during Meta internship): Orthogonal Alignment and its blog post: An Orthogonal Alignment Phenomenon in Cross-Attention
• [2025.06] Extension of Black swan paper: Antifragile will be presented in ICML2025.
• [2025.05] Joining Meta this summer as a research scientist intern (Ranking team).
• [2025.04] Black Swan Hypothesis 🦢 as oral presentation at ICLR25 workshop: Financial AI
• [2025.04] Invited to AI safety panel at Appen
• [2025.03] Started research associate with OpenAI
• [2025.02] New paper on proposing the necessity of antifragility in AI safety : Antifragile
• [2024.02] Serving as a reviewer for ICML, RLC, and Neurips 2025.
• [2025.01] New paper proposing a novel AI safety perspective, Black Swan Hypothesis 🦢, in ICLR2025 (& workshop: Financial AI 25): A Black Swan Hypothesis: The Role of Human Irrationality in AI Safety
• [2024.09] Serving as a reviewer for ICLR, AISTATS 2025.
• [2024.07] Attending RLC.
• [2024.07] Serving on the program chair committee for AAAI 2025.
  
• [2024.05] New paper on Safe reinforcement Learning: Policy-based Primal-Dual Methods for Concave CMDP with Variance Reduction
• [2024.05] New paper on optimal early stopping of non-stationary Reinforcement Learning in ICML 2024 (Oral, top 1%): Pausing Policy Learning in Non-stationary Reinforcement Learning. [Talk/codes]
• [2024.04] Received a Berkeley research fellowship.
• [2024.03] Served as a reviewer for ICML 2024.
• [2024.02] New paper on proposing new policy gradient estimator in reinforcement Learning in IEEE TAC: Beyond Exact Gradients: Convergence of Stochastic Soft-Max Policy Gradient Methods with Entropy Regularization.
• [2023.10] Received a NeurIPS scholar award
• [2023.08] OUTTA has launched a new spinoff team, AI PLAYGROUND, dedicated to creating intuitive AI educational materials for elementary education! I lead the team.
• [2023.05] New paper on proposing Tempo in Non-stationary Reinforcement Learning in NeurIPS 2023: Tempo Adaption in Non-stationary Reinforcement Learning. [slides] [codes]
• [2023.01] New paper on Causal Machine Learning in CDC 2023: Initial State Interventions for Deconfounded Imitation Learning.
• [2022.08] I started Ph.D. at UC Berkeley.
• [2022.05] Received the Kwanjeong Education Foundation Scholarship.
• [2022.03] New paper on Neuroscience and Machine Learning in IEEE TNSRE: Explainable Deep Learning Model for EMG-Based Finger Angle Estimation using Attention. [slides] [videos] [codes]
• [2022] I have co-founded 🚀OUTTA🚀 with Haeun (MIT), Chankyo (UMich)

Education

• Ph.D., University of California, Berkeley, 2022.08 - current
• B.S., Seoul National University, 2015.03 - 2022.02
  (Served military service in Korean Combat Traning Center, 2017.07 - 2019.03)