Woochang Sim심우창

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. I study how language-model agents can reason, adapt, and make effective use of computation.

My current work follows two directions: diagnostic tools for automated improvement of agent harnesses, and learned internal signals for self-monitoring and computation control. In the harness work, I use ARC-AGI-3 to study how solver token probabilities and the recorded origins of model context can help locate and assess harness edits. In the model-level work, I investigate whether learned readouts from unused output coordinates can support judgment and self-monitoring while preserving language-generation ability, with adaptive compute allocation as a further research goal.

My previous work includes controlled evaluation of LLM reasoning and synthetic benchmark construction. For our ACM TIST study, I implemented reasoning experiments and human baselines; for GIFARC, I developed a pipeline that turns GIF-derived analogies into executable ARC-style tasks and evaluated them through in-context learning and supervised fine-tuning experiments.

My broader interests remain adaptive agents under uncertainty, probabilistic task and environment modeling, Bayesian and meta-reinforcement learning, and generalization to unseen tasks and environments. I am also interested in how pretrained language-model knowledge and world models can support interactive adaptation.