🚀 About Me

📫 E-mail : [email protected]

📥 Github : 🖱Github

💻 Linkedin : 📎 linkedin

📞 Phone : +82-10-4909-4946

💡 Strength:

• Comprehensive experience in NLP and CV research, with hands-on work in LLMs, VLMs, alignment training, RLHF, GRPO, and large-scale dataset construction.
• Proficient in implementing and customizing libraries such as Transformers, Accelerate, vLLM, and Django, and quick at translating theoretical understanding into practical code.
• Strong communicator and active collaborator, fostering teamwork and leadership in project-based settings.

🔬 Publications

I0T: Embedding Standardization Method Toward Zero Modality Gap

• Eunki Kim*, Na Min An*, James Thorne, Hyunjung Shim
• ACL ‘25 Main(Oral), Outstanding paper
• TLDR: We attribute and mitigate explicit factors in the image and text embeddings that lead to the modality gap.
• arxiv: https://arxiv.org/abs/2412.14384

Learning to Insert [PAUSE] Tokens for Better Reasoning

• Eunki Kim*, Sangryul Kim*, James Thorne
• ACL ‘25 Findings
• TLDR: We introduce a novel approach termed Dynamic Inserting Token Training (DIT), which identifies positions within sequences where model confidence, determined by token log likelihood, is lowest, and strategically inserts [PAUSE] tokens to bolster the model's predictive capabilities for subsequent tokens.
• arxiv: https://arxiv.org/abs/2506.03616

Sightation Counts: Leveraging Sighted User Feedback in Building a BLV-aligned Dataset of Diagram Descriptions

• Wan Ju Kang, Eunki Kim, Na min An, Sangryul kim, Haemin Choi, Ki Hoon Kwak, James Thorne
• ACL ‘25 Main
• TLDR: We introduce Sightation, a 137k-sample dataset of diagram descriptions for BLV users, created through sighted assessments of model-generated content. The dataset supports training in completion, preference, retrieval, QA, and reasoning, enhancing BLV education and model fine-tuning.
• arxiv: https://arxiv.org/abs/2503.13369

On the Robustness of Reward Models for Language Model Alignment

• Jiwoo Hong, Noah Lee, Eunki Kim, Guijin Son, Woojin Chung, Aman Gupta, Shao Tang, James Thorne