LEAD: Learning Decomposition for Source-free Universal Domain Adaptation This paper proposes a novel framework called LEAD for Source-free Universal Domain Adaptation (SF-UniDA), which aims to transfer knowledge between domains without access to source data. LEAD addresses the challenge of identifying target-private unknown data by decomposing features into source-known and -unknown components. The framework leverages orthogonal decomposition to construct two orthogonal feature spaces, source-known and -unknown, and then uses instance-level decision boundaries to adaptively identify target-private data. LEAD is evaluated across various scenarios, including Partial Domain Adaptation (PDA), Open-set Domain Adaptation (OSDA), and Open-partial-set Domain Adaptation (OPDA). The results show that LEAD outperforms existing methods in terms of performance and efficiency, particularly in the OPDA scenario on the VisDA dataset, where it achieves a 3.5% improvement in H-score compared to GLC. LEAD is also complementary to most existing SF-UniDA methods and can be seamlessly integrated into them to further boost performance. The code for LEAD is available at https://github.com/ispc-lab/LEAD.LEAD: Learning Decomposition for Source-free Universal Domain Adaptation This paper proposes a novel framework called LEAD for Source-free Universal Domain Adaptation (SF-UniDA), which aims to transfer knowledge between domains without access to source data. LEAD addresses the challenge of identifying target-private unknown data by decomposing features into source-known and -unknown components. The framework leverages orthogonal decomposition to construct two orthogonal feature spaces, source-known and -unknown, and then uses instance-level decision boundaries to adaptively identify target-private data. LEAD is evaluated across various scenarios, including Partial Domain Adaptation (PDA), Open-set Domain Adaptation (OSDA), and Open-partial-set Domain Adaptation (OPDA). The results show that LEAD outperforms existing methods in terms of performance and efficiency, particularly in the OPDA scenario on the VisDA dataset, where it achieves a 3.5% improvement in H-score compared to GLC. LEAD is also complementary to most existing SF-UniDA methods and can be seamlessly integrated into them to further boost performance. The code for LEAD is available at https://github.com/ispc-lab/LEAD.