Ongoing
Cost-Aware Tool Use in LLM Agents
Agents that call external tools have to weigh what a task needs against what each option costs. We study how dependable that trade-off is, and what pulls it off course.
LUMI Lab · Dongguk University
LUMI Lab
Language Understanding and Machine Intelligence
We research natural language processing, large language models, and trustworthy AI—including LLM evaluation, reliable agents, multimodal systems, and AI safety—under Prof. Yerin Hwang at Dongguk University.
We are growing. Graduate students, undergraduate research interns, and research collaborators are welcome to get in touch.
Researching reliability where it matters most
Our work sits at the intersection of natural language processing, trustworthy AI, agentic systems, and multimodal evaluation. Strong benchmark scores say little about how a model behaves once instructions conflict, evaluations carry bias, or a first attempt needs fixing. We investigate where AI systems fail under that kind of realistic pressure, and build the benchmarks, analyses, and methods that make their behavior more reliable.
Ongoing Projects
What we are working on
Capable systems still fall short in real use. Right now we focus on three of those gaps: weighing cost against capability, keeping priorities straight under conflict, and recovering from a result that missed.
Ongoing
Instruction Hierarchy Under Accumulated Conflict
A model is expected to respect which source of instructions outranks another. We study how well that ordering holds up once conflicting instructions start to pile up.
Ongoing
Feedback and Correction in Image Editing
When an edit misses what the user actually wanted, can a model tell that something went wrong and put it right? We work on evaluating multimodal systems on recovery, not only on first attempts.
Research Areas
From evaluation to reliable action
We study intelligence not only through what models know, but through how they judge, follow priorities, interact with tools, and respond when their first attempt is wrong.
Trustworthy Evaluation
We study bias, robustness, uncertainty, and faithfulness when language and vision–language models are used as evaluators.
Reliable Agents
We examine instruction hierarchy, tool use, cost-aware decisions, and safety when agents face conflicting or misleading signals.
Feedback & Multimodal Systems
We evaluate whether models can diagnose failures, express actionable feedback, and carry out precise corrections without causing new errors.
Data-Centric & Korean NLP
We develop evaluation and training data, metrics, and analyses that reflect language-specific and real-world model behavior.
Publications
Selected Work by the PI
Selected research that informs the questions we are pursuing at LUMI.
How Instruction Hierarchy Breaks Under Pairwise Multi-Conflict Pressure
Fooling the LVLM Judges: Visual Biases in LVLM-Based Evaluation
LLMs can be easily Confused by Instructional Distractions
News
Latest from the lab
Three papers accepted to EMNLP 2026
Two to the main conference and one to Findings.
Join LUMI
Come work on these questions with us.
We welcome students and collaborators who want to study trustworthy AI through careful experiments and evaluations they can stand behind.