Professor Choi Jun-seok's research team (Lee Min-young, doctoral; Park Ye-ji, combined master's–doctoral; Hwang Dong-jun, doctoral; Kim Ye-jin, master's) has had a paper accepted at the distinguished international conference International Conference on Learning Representations (ICLR) 2026, analysing the role of delimiter tokens in large vision-language models (LVLMs) under multi-image input and proposing an effective technique that improves performance on that basis. ICLR is a globally prestigious international conference in artificial intelligence and machine learning, and will be held in Rio de Janeiro, Brazil, from 23 to 27 April.

LVLMs perform well on single-image tasks. When several images are input at once, however, inference performance degrades markedly because of cross-image information leakage, in which information from different images is mixed. Existing models use delimiter tokens to separate images, but the team's analysis confirmed that these tokens do not in fact block information leakage between images effectively.

The team therefore proposed a simple but effective technique that scales the hidden state of the delimiter tokens. The method strengthens intra-image interaction between tokens while suppressing unnecessary interaction between different images, allowing the model to distinguish per-image information more clearly and perform accurate multi-image inference.

The research is significant in that it re-examines the relatively neglected role of delimiter tokens in LVLMs and presents a practical solution that reliably improves multi-input inference performance without changing the model architecture or additional training. It is expected to be used as a core technique for raising the reliability and accuracy of future multi-image and multi-document AI systems.

References:

▶ Paper link: https://arxiv.org/abs/2602.01984

▶ Code link: https://github.com/MYMY-young/DelimScaling