in progress
Prior-Knowledge-Driven Generalization on ARC-AGI-3
Applying pretrained prior knowledge to unseen ARC environments via BAMDP and Meta-RL. Active competition participant.
Woochang Sim심우창
Ph.D. Student, AI Convergence
Data Science Lab, Gwangju Institute of Science and Technology (GIST)
Advisor: Prof. Sundong Kim
About
I am a Ph.D. student in AI Convergence at the Gwangju Institute of Science and Technology (GIST), advised by Prof. Sundong Kim in the Data Science Lab. My research aims to bridge the gap between the capabilities of large language models and human-level abstract reasoning, anchored by the ARC-AGI benchmark.
Recently, I have been studying how models can leverage prior knowledge acquired during pretraining to generalize to genuinely unseen tasks. I test these ideas on ARC-AGI-3, where I am an active competition participant, and I am currently exploring methods grounded in Bayes-Adaptive MDPs (BAMDP) and Meta-Reinforcement Learning.
Education
- Gwangju Institute of Science and Technology (GIST)M.S. in Artificial IntelligenceFeb 2023 – Feb 2025
- Chosun UniversityB.S. in Computer Science2016 – 2023
Research Interests
- Abstract reasoning & LLM reasoning (ARC-AGI)
- Agentic AI and Meta-Reinforcement Learning
- Analogical & transfer learning — leveraging prior knowledge for unseen tasks
- Prior-knowledge-driven generalization on ARC-AGI-3 via BAMDP / Meta-RLCurrent focus
Featured Research
under review
GIFARC: Synthetic ARC-style Dataset from GIF-derived Analogies
Building ARC-style tasks from human analogies extracted from GIFs.
ACM TIST 2025
Reasoning Abilities of Large Language Models: In-Depth Analysis on the Abstraction and Reasoning Corpus
In-depth analysis of LLM reasoning on the Abstraction and Reasoning Corpus (ARC) — measuring where the models succeed and where they break down across inferential, abstraction, and recombination capabilities.
Publications
Journal· 2
S Lee*, W Sim*, D Shin*, W Seo, J Park, S Lee, S Hwang, S Kim, S Kim(* equal contribution)
Reasoning Abilities of Large Language Models: In-Depth Analysis on the Abstraction and Reasoning Corpus
ACM Transactions on Intelligent Systems and Technology (TIST), 2025
reasoningLLMevaluationARCW Sim, H Jin, S Kim, S Kim
The Possibility of Prompt Engineering for ARC Problem Solving
정보과학회 컴퓨팅의 실제 논문지 30(2):63–69, 2024
reasoningprompt engineeringARCdomestic journal
Conference· 2
W Seo, W Sim, S Kim
Augmenting Few-Shot Demonstrations with Large Language Model
한국정보과학회, 2023
LLMdata augmentationfew-shotdomestic conferenceG Gu, W Sim, J Im, S Kim, S Kim
Using Contrastive Learning for Abstraction and Reasoning Task
한국정보과학회 학술발표논문집:828–830, 2023
contrastive learningreasoningARCdomestic conference
Workshop· 2
S Lee, W Sim, D Shin, S Kim, S Kim
Reasoning Abilities of Large Language Models through the Lens of Abstraction and Reasoning
The First Workshop on System-2 Reasoning at Scale, NeurIPS, 2024
reasoningLLMevaluationARCNeurIPSJ Lee*, W Sim*, S Kim, S Kim(* equal contribution)
Enhancing Analogical Reasoning in the Abstraction and Reasoning Corpus via Model-Based Reinforcement Learning
Workshop on the Interactions between Analogical Reasoning and Machine Learning (IARML @ IJCAI), 2024
analogical reasoningreinforcement learningARCIJCAI
News
- 2026-07Co-organizing the TAIC 2026 workshop at GIST (Pre-ICML event).