<p>Manual annotation of large-scale point cloud datasets for tasks such as 3D object classification, segmentation, and detection is often laborious due to the irregular structure of point clouds. Self-supervised learning, which operates without human labeling, presents a promising approach to address this challenge. Following the success of contrastive learning in image and video understanding, recent works have investigated self-supervised contrastive learning for point cloud analysis. However, existing contrastive learning approaches face a critical bottleneck caused by the incorporation of false negatives in the InfoNCE loss. This leads to inefficient learning when negative sample pairs are insufficient to differentiate all actual object classes, forcing models to over-separate representations of the same class. To address this limitation, we propose PointAT, a lightweight point cloud contrastive learning framework. Our approach employs a Focused Sampling (FS) module for geometric keypoint identification, enabling discriminative object representation learning. Both coarse-grained point clouds from the FS module and original fine-grained point clouds are processed through parallel streams in a Local Residual Attention module. This dual-stream architecture facilitates comprehensive feature extraction and fusion across multiple scales. Finally, we introduce an Adaptive Truncated Loss (AT) to align multi-scale representations while adaptively managing negative sample separation in the embedding space. We provide theoretical analysis validating our proposed loss function. Experimental results demonstrate that PointAT outperforms previous unsupervised methods across diverse downstream tasks, including 3D object classification on ModelNet40 and ScanObjectNN datasets, and part segmentation on ShapeNet. Comprehensive ablation studies further validate the effectiveness of our approach.</p>