Collaborative LiDAR SLAM enables multiple robots to jointly explore and map large-scale environments. However, its performance degrades in dynamic scenes due to moving objects and suboptimal inter-robot map fusion. To address these challenges, we propose two lightweight and system-agnostic modules: Dynamic Point Removal and Reference Selection. The DPR module filters dynamic points in real time using multi-frame voxel consistency, without relying on semantic labels or deep learning. The RS module improves global map fusion by scoring submaps across six reliability metrics—including trajectory accuracy, structural completeness, and registration stability—to dynamically select the optimal reference map. We integrate both modules into two representative collaborative SLAM frameworks, DiSCo-SLAM and DCL-SLAM, and evaluate them on five multi-robot datasets, including real-world campus scenarios. Experimental results demonstrate that our approach consistently improves trajectory accuracy, map consistency, and loop closure robustness, while exhibiting strong generalizability and plug-and-play compatibility.