Junik Bae

I am an incoming M.S. student in Computer Science at Stanford University, starting September 2026. My studies are supported by the Korean government's Study Abroad Scholarship Program (NIIED). Since my undergraduate studies in Computer Science and Engineering at Seoul National University, I have worked primarily with Prof. Youngwoon Lee at Yonsei RLLAB. I am currently visiting Prof. Yuke Zhu's Robot Perception and Learning (RPL) Lab at UT Austin.

My long-term research goal is to develop exploration-driven, creative agents that can achieve superhuman performance in open-ended environments. In particular, I am interested in:

General unsupervised exploration algorithms (e.g., TLDR)
Knowledge-guided, sample-efficient exploration (e.g., PokeAgent)
LLM-based reasoning and agents (e.g., PokeAgent, ESREAL)

In the past, I served in Korean Air Force, specializing in deep learning model development. After that, I joined Naver Cloud Speech team as a research intern. During my undergraduate studies, I was involved in an undergraduate research program at Language and Data Intelligence Lab, focusing on Natural Language Understanding for code refinement. Also, I worked as a lab intern at SNU Vision and Learning Lab, studying Retrieval augmented generation in NLP.

Email / CV / Github

Research and Projects

Below are selected publications and projects. See my CV for a complete list.

	PokeAgent Challenge: 1st Place in RPG Speedrunning Track NeurIPS, 2025 Award: $1,500 + $700 (GCP credits), competition report co-authorship Developed an autonomous Pokemon Emerald speedrun agent for long-horizon decision-making. Built a knowledge-guided RL pipeline using Voyager-style executable skill generation combined with reinforcement learning.
	TLDR: Unsupervised Goal-Conditioned RL via Temporal Distance-Aware Representations Junik Bae, Kwanyoung Park, Youngwoon Lee CoRL, 2024 TLDR is a novel unsupervised GCRL algorithm that leverages Temporal Distance-aware Representations to improve both goal-directed exploration and goal-conditioned policy learning. TLDR significantly outperforms previous unsupervised GCRL methods in achieving a wide variety of states.
	Exploiting Semantic Reconstruction to Mitigate Hallucinations in Vision-Language Models Minchan Kim, Minyoung Kim, Junik Bae, Suhwan Choi, Sungkyung Kim, Buru Chang ECCV*, 2024 We introduce ESREAL, a novel unsupervised learning framework designed to reduce hallucinations in vision-language models by accurately identifying and penalizing hallucinated tokens. ESREAL results in significant reductions in hallucinations in models like LLaVA, InstructBLIP, and mPLUG-Owl2, without needing image-text pairs.
	2022 Korean AI Competition 2nd Place, Awarded by Korean Minister of Science and Technology: 10,000,000 KRW. Developed a model targeting the tasks of Korean Free-chat, Command-chat and Dialect-chat recognition using a variant of wav2vec model. Implemented Transformer LM beam-search decoder for the inference, enabling it to consider long-range contextual information effectively.
	2022 Military AI Competition 1st Place, Awarded by Korean Minister of Science and Technology: 20,000,000 KRW Developed a model for Change Segmentation and Image Denoising tasks, utilizing a variant of U-Net model and several engineering techniques for the tasks.
	NaturalSpeech Implementation (400+ stars) Implemented Microsoft’s NaturalSpeech: End-to-End Text to Speech Synthesis with Human-Level Quality, which was a SOTA for the LJ Speech Dataset.

Talks

Understanding RLHF in Detail

Hosted a presentation titled "Understanding RLHF in Detail" at Deepest (SNU Deep Learning Research Club), providing an in-depth explanation of Proximal Policy Optimization (PPO) with preferences datasets and Direct Policy Optimization (DPO).

A Path Towards Autonomous AI

Hosted a presentation titled "A Path Towards Autonomous AI" at Deepest (SNU Deep Learning Research Club), which explains motivations for Joint-Embedding Predictive Architecture proposed by Yann LeCun, and some recent works utilizing this architecture.

Website source code credit to Dr. Jon Barron.