Ayoung Lee

Enhancing Multilingual Reasoning in Language Models
- Developing a new dataset that focuses on cultural adaptation, addressing the limitations of direct translation.
- Exploring different rollout methods and loss functions in RL to improve multilingual reasoning.
- Generating synthetic training datasets to improve multilingual reasoning.
Ongoing project

LiveOIBench: Can Large Language Models Outperform Human Contestants in Informatics Olympiads?
- Implemented self-debugging baselines.
Kaijian Zou, Aaron Xiong, Yunxiang Zhang, Xinliang Frederick Zhang, Yueqi Ren, Jirong Yang, Ayoung Lee, Shitanshu Bhushan, Lu Wang
Submitted to ICLR 2026, Initial avg. score in top ≈25%

Logit Arithmetic Elicits Long Reasoning Capabilities Without Training
- Ran supervised fine-tuning experiments, with and without LoRA, for distillation baselines.
Yunxiang Zhang, Muhammad Khalifa, Lechen Zhang, Xin Liu, Ayoung Lee, Xinliang Frederick Zhang, Farima Fatahi Bayat, Lu Wang
COLM 2025, The 1st Workshop on Test-time Scaling and Reasoning Models, Submitted to ARR 2025 Oct [paper]

CLASH: Evaluating Language Models on Judging High-Stakes Dilemmas from Multiple Perspectives
- Assessed the ability of language models to respond to dilemmas based on character descriptions as perspectives.
- Analyzed reasoning chains and identified new failure modes in the value-understanding domain.
- Converted a non-verifiable dilemma task into a verifiable format.
Ayoung Lee, Ryan Sungmo Kwon, Peter Railton, Lu Wang
Preprint, Submitted to ICLR 2026, Initial avg. score in top ≈20% [paper]

On Consistency Training for Language-Based Image Editing Interface
- Generated training dataset using stable diffusion and YOLOv7 to enforce greater consistency in object-level image edits.
Youngwon Lee*, Ayoung Lee*, Yeonjoon Jung, Seung-won Hwang (* denotes equal contribution)
IJCNLP-AACL 2023, Second Workshop on Natural Language Interfaces (Oral) [paper] [code]