Preprint 2026
Learning How to Remember: Memory Modality Selection for GUI Agents
Selecting the right memory representation for each step of a GUI agent’s interaction.
AI agents · Memory · Multimodal reasoning
Taeyun Roh
노태윤Building agents that remember,
reason, and act.
I am an AI Research Intern at Samsung Research America in Mountain View, and a researcher at Korea University’s DMIS Lab, advised by Prof. Jaewoo Kang.
My research focuses on memory systems for AI agents and visually grounded reasoning. I study how agents can organize experience, recall the right evidence, and learn to make better decisions.
Recently
All updatesSelected work
Research
ACL Findings 2026
CLAG: Adaptive Memory Organization via Agent-Driven Clustering for Small Language Model Agents
Self-organizing memory that helps small language model agents retrieve and reuse experience.
Preprint 2026
BUZZY: Contrastive Scoring to Mitigate Text-Induced Bias in Multimodal Multiple-Choice QA
Grounding answers in visual evidence through training-free contrastive scoring.
EMNLP Main 2026 · Accepted
Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning
Tailoring reinforcement learning to the steps where multimodal reasoning breaks down.
Research directions
From experience to action.
Memory
How should an agent organize experience and decide what to remember?
SwitchMem / CLAGVisual reasoning
How can an agent ground its decisions in what it actually sees?
BUZZYLearning to act
How can post-training improve planning, tool use, and reliable reasoning?
Expert-level planning