The paper "Learning Decomposition for Source-free Universal Domain Adaptation" (LEAD) addresses the challenge of Universal Domain Adaptation (UniDA) in the absence of source data, a scenario known as Source-free Universal Domain Adaptation (SF-UniDA). The primary goal is to identify target-private data, which are not present in the source domain, without relying on manual thresholding or iterative clustering, which can be time-consuming and unstable. LEAD proposes a novel approach called Learning Decomposition (LEAD), which decomposes features into source-known and source-unknown components. This decomposition helps in identifying target-private data by leveraging the orthogonality of the feature spaces. The method then builds instance-level decision boundaries using a metric that considers distances to target prototypes and source anchors. This approach is more robust and efficient compared to existing methods, as demonstrated through extensive experiments on various datasets and scenarios. Key contributions of LEAD include: 1. **Feature Decomposition**: LEAD decomposes features into source-known and source-unknown components, facilitating the identification of target-private data. 2. **Instance-Level Decision Boundaries**: It establishes instance-level decision boundaries using a metric that accounts for distances to target prototypes and source anchors. 3. **Efficiency and Robustness**: LEAD is more efficient and robust compared to existing methods, especially in handling covariate shifts and label shifts. Experiments on datasets such as Office-31, OfficeHome, VisDA, and DomainNet show that LEAD outperforms or matches the performance of existing methods, particularly in challenging scenarios like Open-Partial Set Domain Adaptation (OPDA). The code for LEAD is available at <https://github.com/ispcc-lab/LEAD>.The paper "Learning Decomposition for Source-free Universal Domain Adaptation" (LEAD) addresses the challenge of Universal Domain Adaptation (UniDA) in the absence of source data, a scenario known as Source-free Universal Domain Adaptation (SF-UniDA). The primary goal is to identify target-private data, which are not present in the source domain, without relying on manual thresholding or iterative clustering, which can be time-consuming and unstable. LEAD proposes a novel approach called Learning Decomposition (LEAD), which decomposes features into source-known and source-unknown components. This decomposition helps in identifying target-private data by leveraging the orthogonality of the feature spaces. The method then builds instance-level decision boundaries using a metric that considers distances to target prototypes and source anchors. This approach is more robust and efficient compared to existing methods, as demonstrated through extensive experiments on various datasets and scenarios. Key contributions of LEAD include: 1. **Feature Decomposition**: LEAD decomposes features into source-known and source-unknown components, facilitating the identification of target-private data. 2. **Instance-Level Decision Boundaries**: It establishes instance-level decision boundaries using a metric that accounts for distances to target prototypes and source anchors. 3. **Efficiency and Robustness**: LEAD is more efficient and robust compared to existing methods, especially in handling covariate shifts and label shifts. Experiments on datasets such as Office-31, OfficeHome, VisDA, and DomainNet show that LEAD outperforms or matches the performance of existing methods, particularly in challenging scenarios like Open-Partial Set Domain Adaptation (OPDA). The code for LEAD is available at <https://github.com/ispcc-lab/LEAD>.