Research

I work on human-centered NLP, multi-agent systems, and the computational analysis of group interaction. On the human side, I build corpora of real teams at work. On the model side, I ask whether cognitive diversity can be created in LLMs, and whether it survives long conversations.

01Publications

[C.1]

TIDES: A Longitudinal Bilingual Dataset for Modeling Multi-Party Social Dynamics

Heechan Lee*, Jeonggyu Kang*, Junho Myung, Jaywoong Jeong, Juho Kim, Joseph Seering

COLM 2026Conference on Language Modeling (COLM)

*Equal contribution

12 teams · 104 meetings · 76K+ utterances

02Current Work

2025 — now · CSTL

TIDES & Computational Analysis of Group Interaction

Co-leading TIDES, a longitudinal bilingual corpus for modeling multi-party social dynamics, annotated with speaker roles, developmental stages, and interaction-process categories.

  • 12 teams · 104 meetings · 76K+ utterances
  • Accepted to COLM 2026 (co-first author)

2025 — now · CSTL

Manufacturing Cognitive Diversity via Internal Model Controls

Can internal model controls manufacture useful cognitive diversity within a single base model? I'm building a systematic falsification record to find out.

  • Token-matched baselines and effect-size analysis

2025 — now · CSTL

Diversity Persistence in Multi-Agent Conversation

Does diversity survive extended multi-agent conversation, or do agents inevitably converge?

  • Multi-turn debate among stance-seeded agents

03Previous Work

Aug 2024 — Feb 2025 · Samsung Electronics

LLM & RAG System Optimization

Built LLM-enhanced RAG systems at Samsung Electronics' Device Solution Division during a 7-month internship.

  • Prompt-Induced Reasoning based Query Decomposition: +28% search quality, -43% zero-hit queries
  • Query-Document Relevance Check (DeepSeek-R1-Distill-Qwen-32B): 98.1% score stability
  • RAG evaluation system applied to 3,000+ real-world evaluations
  • Korean-optimized LLM prompt engineering

2024 — 2025 · KAIST

Computer Vision

Course and research projects in generative models, anomaly detection, and CNN optimization.

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