Abstract: To address the demand for maintaining the segmentation accuracy while enhancing the real-time performance of large-scale point cloud scenes in point cloud semantic segmentation, a point cloud semantic segmentation algorithm based on adaptive spatial mixing is proposed. Firstly, the point features and domain features are fused in the embedding layer to enhance the original point data and capture local geometric information. Subsequently, the embedded point features are iteratively updated through the spatial and channel mixing layers. In the spatial mixing stage, the point cloud is projected using a sparse matrix, an axis-aware feature mixing module is introduced to capture cross-dimensional context information, and adaptive fusion of local features and query features is achieved. The fused features are then connected with the residual of the embedded features and input into the channel mixing stage. In the channel mixing stage, a low-parameter low-rank approximation Multi-Layer Perceptron (MLP) module is designed to replace the traditional MLP, for efficient processing of the features. After multiple alternating mixings, the updated features are classified using convolution. Experimental results on the SemanticKITTI and SemanticPOSS datasets show that the proposed method achieves a good balance between efficiency and accuracy, with the mean intersection over union (mIoU) scores on the validation set reaching 64.0% and 52.2%, respectively.