Professor Choe Jun-suk’s Research Team of the Department of Computer Science and Engineering Has Paper Accepted at ICLR 2026, a Premier International Conference in Artificial Intelligence
Professor Choe Jun-suk’s research team of the Department of Computer Science and Engineering (Lee Min-young, PhD student; Park Ye-ji, Integrated Master’s–PhD student; Hwang Dong-jun, PhD student; Kim Ye-jin, Master’s student) has had a paper accepted to the International Conference on Learning Representations (ICLR) 2026. The paper analyzes the role of separator tokens in Large Vision-Language Models (LVLMs) within multi-image input environments and proposes an effective technique to enhance performance through this analysis. ICLR is a prestigious international conference in the field of artificial intelligence and machine learning, scheduled to be held in Rio de Janeiro, Brazil, from April 23 to 27.
LVLMs demonstrate excellent performance on single-image tasks. However, when processing multiple images simultaneously, they suffer from cross-image information leakage, where information from different images becomes mixed, leading to a significant degradation in inference performance. While existing models use delimiter tokens to separate images, the research team's analysis revealed that these tokens do not effectively prevent information leakage between images.
To address this issue, the research team proposed a simple yet effective technique: scaling the hidden state of delimiter tokens. This method enhances intra-image interactions while suppressing unnecessary interactions between different images. This enables the model to more clearly differentiate image-specific information and perform accurate multi-image inference.
This study is significant as it reexamines the role of delimiter tokens, which have been relatively overlooked in LVLMs, and presents a practical solution that reliably improves multi-input inference performance without requiring modifications to the model architecture or additional training. It is expected to serve as a core technology for enhancing the reliability and accuracy of multi-image and multi-document AI systems in the future.
▶Title: Enhancing Multi-Image Understanding through Delimiter Token Scaling
▶Authors: Lee Min-young (First author, Sogang University), Park Ye-ji (Second author, Sogang University), Hwang Dong-jun (Third author, Sogang University), Kim Ye-jin (Fourth author, Sogang University/KAIST), Oh Sung-jun (Fifth author, KAIST/University of Tübingen), Choe Jun-suk (Corresponding author, Sogang University)
▶Paper Link: https://arxiv.org/abs/2602.01984
▶Code link: https://github.com/MYMY-young/DelimScaling
[SEO 키워드]
Professor Choe Jun-suk, ICLR 2026, Large Vision-Language Models
[Summary]
A research team led by Professor Choe Jun-suk from the Department of Computer Science and Engineering had a paper accepted at ICLR 2026, a premier AI conference. The study introduces a novel technique to enhance Large Vision-Language Models by scaling delimiter tokens, effectively preventing information leakage in multi-image environments and improving inference accuracy without additional training.
LVLMs demonstrate excellent performance on single-image tasks. However, when processing multiple images simultaneously, they suffer from cross-image information leakage, where information from different images becomes mixed, leading to a significant degradation in inference performance. While existing models use delimiter tokens to separate images, the research team's analysis revealed that these tokens do not effectively prevent information leakage between images.
To address this issue, the research team proposed a simple yet effective technique: scaling the hidden state of delimiter tokens. This method enhances intra-image interactions while suppressing unnecessary interactions between different images. This enables the model to more clearly differentiate image-specific information and perform accurate multi-image inference.
This study is significant as it reexamines the role of delimiter tokens, which have been relatively overlooked in LVLMs, and presents a practical solution that reliably improves multi-input inference performance without requiring modifications to the model architecture or additional training. It is expected to serve as a core technology for enhancing the reliability and accuracy of multi-image and multi-document AI systems in the future.
▶Title: Enhancing Multi-Image Understanding through Delimiter Token Scaling
▶Authors: Lee Min-young (First author, Sogang University), Park Ye-ji (Second author, Sogang University), Hwang Dong-jun (Third author, Sogang University), Kim Ye-jin (Fourth author, Sogang University/KAIST), Oh Sung-jun (Fifth author, KAIST/University of Tübingen), Choe Jun-suk (Corresponding author, Sogang University)
▶Paper Link: https://arxiv.org/abs/2602.01984
▶Code link: https://github.com/MYMY-young/DelimScaling
[SEO 키워드]
Professor Choe Jun-suk, ICLR 2026, Large Vision-Language Models
[Summary]
A research team led by Professor Choe Jun-suk from the Department of Computer Science and Engineering had a paper accepted at ICLR 2026, a premier AI conference. The study introduces a novel technique to enhance Large Vision-Language Models by scaling delimiter tokens, effectively preventing information leakage in multi-image environments and improving inference accuracy without additional training.