EXAONE 4.5 技术报告
计算与语言
2026-04-13 v1
摘要
本技术报告介绍了EXAONE 4.5，这是LG AI Research首个发布的开源权重视觉语言模型。EXAONE 4.5通过将专用视觉编码器集成到现有的EXAONE 4.0框架中，实现了对视觉和文本两种模态的原生多模态预训练。该模型在大规模数据上进行训练，特别强调与LG战略应用领域相匹配的文件中心语料。这种针对性的数据设计使模型在文档理解及相关任务上获得显著性能提升，同时也在广泛的语言能力方面实现整体性改进。EXAONE 4.5将上下文长度扩展至256K token，促进长上下文推理和企业级应用场景。对比评估显示，EXAONE 4.5在通用基准测试中表现出竞争力，同时在文档理解和韩语语境推理方面超越了规模相似的SOTA模型。作为LG持续致力于实际工业部署的一部分，EXAONE 4.5旨在能够持续扩展其他领域和应用场景，以推动AI为更美好的生活而进步。
引用
@article{arxiv.2604.08644,
title = {EXAONE 4.5 Technical Report},
author = {Eunbi Choi and Kibong Choi and Sehyun Chun and Seokhee Hong and Junwon Hwang and Hyojin Jeon and Ahra Jo and Hyunjik Jo and Yeonsik Jo and Joonkee Kim and Seonghwan Kim and Soyeon Kim and Sunkyoung Kim and Yireun Kim and Yongil Kim and Changhun Lee and Haeju Lee and Jinsik Lee and Kyungmin Lee and Sangha Park and Kwangrok Ryoo and Minju Seo and Sejong Yang and Heuiyeen Yeen and Hwan Chang and Stanley Jungkyu Choi and Yejin Choi and Kyubeen Han and Joonwon Jang and Kijeong Jeon and Geunyeong Jeong and Gerrard Jeongwon Jo and Jiyeon Jung and Daeseong Kim and Dohoon Kim and Dohyun Kim and Hyunseo Kim and Minu Kim and Myoungshin Kim and Youchul Kim and Byungoh Ko and Christopher Lee and Edward Hwayoung Lee and Honglak Lee and Jiyoung Lee and Sangeun Lee and Seungwon Lim and Woohyung Lim and Jueun Mun and Jaewoo Park and Jimin Park and Jinho Park and Yongmin Park and Wooseok Seo and Yongwoo Song and Sihyuk Yi and Kyungjae Yoo and Sangyeon Yoon},
journal= {arXiv preprint arXiv:2604.08644},
year = {2026}
}