PARI: Policy-Driven Active Residual Intervention for Weakly Supervised Point Cloud Segmentation

3D Weakly Supervised Semantic Segmentation (3D-WSSS) on point clouds aims to learn dense 3D semantics from extremely sparse point-level annotations. Existing methods follow a passive paradigm based on pseudo-label propagation or consistency regularization, leading to two key issues: 1) Lack of targeted intervention, where erroneous predictions are not explicitly identified and corrected, resulting in severe error accumulation; 2) Limited contextual awareness, where supervision is restricted to point- or view-level self-augmentation consistency, failing to capture the intricate contextual correlations embedded within the scene. To address these limitations, we propose a Policy-driven Active Residual Intervention (PARI) framework for 3D-WSSS, where an Active Intervention Agent (AIA) explicitly localizes high-risk points under the guidance of a Confidence-aware Reward Shaping (CRS) strategy, and a Bias-aware Refinement Module (BRM) subsequently executes targeted refinement by exploiting rich spatial and semantic scene contexts. Specifically, AIA actively localizes high-risk points from uncertainty maps and prediction priors, while BRM refines them through residual correction based on hybrid spatial-feature context. CRS introduces a high-confidence suppression reward that discourages unnecessary intervention on reliable predictions, and a low-confidence exploration reward that encourages intervention on uncertain predictions. Experiments on S3DIS and ScanNetV2 demonstrate state-of-the-art performance under various weak supervision settings with strong cross-backbone generalization, even surpassing the same-backbone fully supervised counterpart with 1% annotations.