01Structure
Give language models a sense of discourse.
Discourse-level structure as both a training signal and an analysis lens: structural alignment for coherent long-form generation, and discourse motifs that tell human writing from machine text.
Zae Myung Kim
Linguistic structure · metacognitive self-improvement · human-AI collaboration
I teach language models to write with structure, then to judge and recursively improve their own reasoning.
I am a Ph.D. candidate in the Minnesota NLP group at the University of Minnesota, advised by Dongyeop Kang. My research develops a meta-scaffolding paradigm for large language models: discourse-level structure as a training signal for coherent long-form generation, metacognitive feedback inside reinforcement learning to stabilize optimization, and test-time metacognition that lets agents recursively improve themselves.
Before the Ph.D. I worked on machine translation at NAVER LABS Europe and Papago, and I have interned at Amazon AWS, Amazon AGI, Salesforce AI Research, and Grammarly.
Current research focus
02Metacognition
Let models judge, and improve, themselves.
Metacognitive feedback inside the training loop: evolving reward models, evaluators audited for cognitive bias, and recursive self-improvement whose layers take on emergent roles.
03Collaboration
Build AI that writes and works alongside people.
Interactive revision systems, sensemaking scaffolds for researchers, and benchmarks that probe how agents navigate real tools and tasks.
Selected publications
A fixed meta-operation applied recursively to its own outputs lets LLM agents build deeper self-improving layers, beating prior self-improvers on eight benchmarks.
A benchmark of fork-merge tool-use puzzles over Wikipedia shows agents rarely fail at calling tools but often fail at navigating to the right pages.
Token-level RL rewards derived from human discourse structure make LLMs write longer, more coherent essays and summaries than standard or RLHF-tuned models.
A meta-reward model rewrites the reward model's prompt during RL training, curbing reward hacking and removing most manual reward-prompt engineering.
Hierarchical discourse motifs reveal that human writing is structurally more varied than LLM text, and these features improve machine-text detectors even out of distribution.
An end-to-end revision system first detects which spans to edit and why, then revises them, beating baselines across several text-editing tasks.
R3 is a human-in-the-loop system where a model proposes edits, writers accept or reject them, and the loop repeats until the text converges.
The most important attention heads in a many-to-one multilingual NMT model are largely shared across language pairs, and a third can be pruned safely.