Human-AI Engineering undergraduate at Sangmyung University
Trustworthy AI · Model interpretability · Language model evaluation
I’m interested in how to evaluate the explanations models give and use that evidence to guide human review.
- Faithfulness. Do model explanations reflect the evidence that drove a prediction?
- Fine-tuning. Which errors decrease after fine-tuning, and which persist?
- Human review. What evidence helps a reviewer decide when to step in?
| Project | Approach |
|---|---|
| HEX Review Demo | A review prototype for comparing news headlines with article text and recording the evidence behind a decision. Includes synthetic examples. |
| PM10 Forecasting | Jupyter experiments that examine errors during rapid changes in PM10 and compare features, prediction aggregation, and models. |
모델이 제시한 설명이 실제 판단 근거를 반영하는지, 파인튜닝 후 어떤 오류가 줄고 어떤 오류가 남는지에 관심이 있습니다. 사람이 모델의 출력을 검토할 때 어떤 근거가 개입 여부를 판단하는 데 도움이 되는지도 살펴보고 싶습니다.