Junseok Kim (Andrew)

Hello 👋! I am a Ph.D student at Seoul National University, advised by Prof. Kyomin Jung. My research centers on LLM Reliability and LLM Agents. I seek to understand how language models work and why they behave as they do, and use these insights to improve their reasoning, decisions, and interactions with users.

Email  /  Google Scholar  /  Github  /  Linkedin

News

Research

My research spans two connected directions: 1) Reliability: understanding and improving model behavior. I investigate how internal mechanisms shape reasoning, confidence, and failure, and use these insights to develop inference-time strategies for reliable decisions under uncertainty. 2) LLM Agents: extending reliability to memory and interaction. I study how agents read and use the given context and interact with users and tools over multiple turns and long horizons, while adapting to user-specific needs and constraints such as personalization and privacy. Across both directions, understanding model behavior guides my use of interpretability, inference-time control, reinforcement learning and post-training, and memory and context management.

We propose Personalized Privacy Control in LLMs via Attention Head Intervention, a method that allows users to control the privacy of their interactions with large language models by intervening in the attention mechanisms.

We propose Tool Selection Bias Amplifies in Multi-turn User–Agent Interactions, a study that examines how tool selection bias can amplify in multi-turn interactions between users and agents.

We propose Reliability-Aware Adaptive Self-Consistency (ReASC), an adaptive self-consistency framework that incorporates a response-level confidence as a reliability signal to guide how evidence is accumulated at inference time.

Persona Switch is a training-free decoding method that improves reasoning by step-wise switching between zero-shot and role-play prompting based on token-level confidence at decoding time.

We analyze the impact of role-play prompts in zero-shot reasoning tasks and show that they can be detrimental to performance in some cases depending on the designed persona.

We introduces "Decomposition of Experts" (DoE), a framework that accelerates inference by dynamically identifying and activating only task-specific neurons within a language model to reduce computational costs without sacrificing accuracy.

Last updated in October 2026. This page is based on Jon Barron's website template.