LVRPO: Language-Visual Alignment with GRPO for Multimodal Understanding and Generation
LVRPO:基于GRPO的语言-视觉对齐用于多模态理解和生成
Shentong Mo, Sukmin Yun
机构
*
Department of Machine Learning, CMU, USA(卡内基梅隆大学机器学习系,美国)
;
Department of Machine Learning, MBZUAI, UAE(穆罕默德·本·扎耶德人工智能大学机器学习系,阿联酋)
;
Department of Artificial Intelligence, Hanyang University ERICA, South Korea(汉阳大学ERICA校区人工智能系,韩国)
机构
*
Beijing Wenge Technology Co., Ltd.(北京文歌科技有限公司)
;
Hebei Provincial Hospital of Traditional Chinese Medicine(河北省中医院)
;
Institute of Automation, Chinese Academy of Sciences(中国科学院自动化研究所)
;
Tianjin University(天津大学)
机构
*
Lyles School of Civil and Construction Engineering, Purdue University(普渡大学莱尔斯土木与建筑工程学院)
;
Department of Civil and Environmental Engineering, University of Wisconsin-Madison(威斯康星大学麦迪逊分校土木与环境工程系)
;
Google(谷歌)
Comments20 pages, 10 figures, 3 tables. Training-free harmful-prompt detector via angular deviation in LLM residual streams. Evaluated on six Qwen variants (base / instruct / abliterated). Achieves AUROC over 0.937 (harmful-vs-normative) and 1.000 (harmful-vs-benign-aggressive) with no harmful training data
机构
*
Guangdong Provincial Key Laboratory of Space-Aerial Networking and Intelligent Sensing, Harbin Institute of Technology, Shenzhen(广东省空天网络与智能感知重点实验室,哈尔滨工业大学(深圳))
;
Peng Cheng Laboratory (PCL), Shenzhen(鹏城实验室)
;
Information Systems Technology and Design, Singapore University of Technology and Design(新加坡科技设计大学信息系统技术与设计系)
;
School of Computer Science and Engineering, Central South University, Changsha(中南大学计算机科学与工程学院)