Reiss Koh

Woosung (Reiss) Koh

Graduate Student

About Me

👋 Hi! I am interested in expanding what is possible for humanity. To this end, I work on improving models and systems along the performance and cost axes. My work has spanned efficient long-context inference, scaling pre-training, post-training, reasoning, agents, feedback/evaluation, and inference-time scaling. These days, I work on (i) continual post-training, (ii) self-distillation/OPD, and (iii) VideoLLMs, while remaining broadly interested in high-impact problems. I am bilingual in English and Korean. If you share similar interests, let’s connect.

Interests

Reasoning and Agents
Efficient Learning and Models
Multi-modality

Education

Graduate Student
KAIST AI
Bachelor's Degree
Yonsei University
Diploma Programme
International Baccalaureate

Publications

[P2] Namgyu Ho*, Huzama Ahmad*, Woosung Koh*, Se-Young Yun, Tal Schuster, Cicero Nogueira Dos Santos. Large Language Models Can Control Their Own Attention Span. Pre-print

Long-context Efficient Inference Sparse Attention

[P1] Woosung Koh*, Jeyoung Jeon*, Youngjin Song, Yujin Cheon, Soowon Oh, Jaehyeong Choi, Se-Young Yun. mSFT: Addressing Dataset Mixtures Overfiting Heterogeneously in Multi-task SFT. Pre-print

PDF CODE Post-training Multi-task Generalization

[C5] Woosung Koh*, Sungjun Han*, Segyu Lee, Se-Young Yun, Jay Shin. Generative Visual Code Mobile World Models. ICML 2026

PDF CODE PROJECT World Model Mobile GUI Code Generation VLM Post-training

[C4] Woosung Koh, Juyoung Suk, Sungjun Han, Se-Young Yun, Jay Shin. Predicting LLM Reasoning Performance with Small Proxy Model. ICLR 2026

PDF DATASET Pre-training Scaling Reasoning Efficiency

[C3] Woosung Koh, Wonbeen Oh, Jaein Jang, MinHyung Lee, Hyeongjin Kim, Ah Yeon Kim, Joonkee Kim, Junghyun Lee, Taehyeon Kim, Se-Young Yun. AdaSTaR: Adaptive Data Sampling for Training Self-Taught Reasoners. NeurIPS 2025

PDF CODE Post-training Reasoning Data Sampling Training Efficiency

[C2] Woosung Koh*, Jang Han Yoon*, MinHyung Lee, Youngjin Song, Jaegwan Cho, Jaehyun Kang, Taehyeon Kim, Se-Young Yun, Youngjae Yu, Bongshin Lee. C$^2$: Scalable Auto-Feedback for LLM-based Chart Generation. NAACL 2025 Main Long (Oral)

PDF CODE PROJECT VIDEO Chart Generation Code Generation VLM-as-a-Judge Inference-time Scaling

[C1] Woosung Koh, Wonbeen Oh, Siyeol Kim, Suhin Shin, Hyeongjin Kim, Jaein Jang, Junghyun Lee, Se-Young Yun. FlickerFusion: Intra-trajectory Domain Generalizing Multi-Agent RL. ICLR 2025

PDF CODE PROJECT Multi-Agent RL Domain Generalization

[W2] Insu Choi*, Woosung Koh*, Gimin Kang, Yuntae Jang, Woo Chang Kim. Encoding Temporal Statistical-space Priors via Augmented Representation. IJCAI 2024 STRL Workshop (Oral)

PDF Spatio-temporal Prediction Financial Markets

[W1] Woosung Koh*, Insu Choi*, Yuntae Jang*, Gimin Kang, Woo Chang Kim. Curriculum Learning and Imitation Learning for Model-free Control on Financial Time-series. AAAI 2024 AI4TS Workshop (Oral)

PDF Reinforcement Learning Financial Markets

[J1] Insu Choi*, Woosung Koh*, Bonwoo Koo*, Woo Chang Kim. Network-based exploratory data analysis and explainable three-stage deep clustering for financial customer profiling. Engineering Applications of Artificial Intelligence (SCIE, Q1)

PDF Explainable AI Personalized AI