Research

My research sits at the intersection of multi-agent AI systems, human cognition, and collaborative interaction. I investigate how artificial agents can maintain distinct personalities during extended dialogue and how group dynamics shape collective intelligence.

Current Research

Multi-Agent LLM Dialogue Systems

Current

Core question: "Does turn-taking behavior match the agent's persona?" I investigate why AI agents with distinct personalities gradually converge to similar responses during extended conversations, and how to achieve genuine collective intelligence rather than simple information aggregation.

Key Contributions

  • Agent Complex: A dual-process cognitive architecture featuring Social Brain (fast decision-making) and Higher Cognition (deliberate response generation) modules, inspired by Dual Process Theory
  • Prefix-tuned personality vectors: 17-dimensional personality conditioning that provides meaningful resistance to homogenization
  • Motivation-based turn-taking: Eight-heuristic system (Relevance, Information Gap, Balance, Dynamics, Coherence, Expected Impact, Originality, Urgency) achieving 82% user preference over round-robin systems
  • Discovery of the Deliberation Paradox: Multi-agent discussions achieve perfect information pooling but produce lower-quality conclusions than single agents with direct information access

Theoretical Foundations

Social Brain HypothesisDual Process Theory (Kahneman)Hidden Profile ParadigmBeneficial Inconsistency

Collaborative Interaction Analysis

Current

Contributing to an audio-to-text dataset construction pipeline for analyzing group collaboration recordings at CSTL Lab. Enhancing speaker diarization through parallelization and speaker embedding techniques.

Key Contributions

  • Building robust audio-to-text pipeline for collaborative interaction recordings
  • Improving speaker diarization accuracy through parallelization strategies
  • Developing speaker embedding techniques for multi-speaker environments

Previous Research

LLM & RAG System Optimization

Previous

Samsung Electronics, Aug 2024 -- Feb 2025

Developed and optimized enterprise-scale RAG systems at Samsung Electronics, focusing on query decomposition, relevance checking, and evaluation frameworks for Korean-language applications.

Key Contributions

  • Prompt-Induced Reasoning based Query Decomposition improving search quality by 28%
  • Query-Document Relevance Check achieving 98.1% score stability
  • RAG evaluation system applied to 3,000+ real-world evaluations
  • Korean-optimized LLM prompt engineering

Computer Vision

Previous

Explored generative models, anomaly detection, and CNN optimization across multiple research projects.

Key Contributions

  • StyleCLIP-based localized image editing in GAN latent space
  • Audio anomaly detection for industrial machinery (DCASE Challenge)
  • CNN architecture optimization (Stanford CS231n based research)