The field of 3D point cloud registration is rapidly advancing, with a focus on developing more robust and efficient methods for aligning multiple point clouds. Recent research has explored the use of attention mechanisms, graph neural networks, and adaptive fusion techniques to improve the accuracy and speed of registration algorithms. One notable direction is the use of learning-based approaches, which can learn to extract distinctive features from point clouds and achieve state-of-the-art results on benchmark datasets. Another area of research is the development of real-time registration methods, which can be deployed on resource-constrained devices such as embedded GPUs. Noteworthy papers in this area include: LAHNet, which introduces a local attention mechanism for point cloud registration, achieving significant results on real-world indoor and outdoor benchmarks. Register Any Point, which casts registration as conditional generation and achieves state-of-the-art results on pairwise and multi-view registration benchmarks. TALO, which proposes a higher-DOF and long-term alignment framework for online 3D reconstruction, consistently yielding more coherent geometry and lower trajectory errors across multiple datasets. Attention-guided reference point shifting, which introduces an attention-based reference point shifting layer to acquire transformation-invariant features for partial point set registration. MAFNet, which proposes a multi-frequency adaptive fusion network for real-time stereo matching, producing high-quality disparity maps using only efficient 2D convolutions. A dynamic memory assignment strategy, which optimizes the memory usage of the dilation operation in the VANICP algorithm, achieving over 97% reduction in memory consumption while preserving the original performance.