Abstract
Reliable and robust multimodal crisis classification is challenged by degraded, corrupted, or missing image–text evidence, because reliability-blind fusion can propagate an unreliable modality and amplify local evidence failure into joint prediction error. To address this problem, we propose Clean-Preserving Residual Reliability Adaptation (CP-RRA), a preserve–diagnose–intervene framework for robust multimodal crisis classification. First, CP-RRA preserves the intact evidence decision pathway as a stable anchor; second, it estimates modality reliability and intervention strength to construct a reliability-aware target representation; finally, it applies bounded residual adaptation between the target and the preserved anchor, with the intervention strength controlling the residual update. Evaluation on a five-class CrisisMMD v2.0 subset covers image and text corruptions, training-unseen perturbations, and complete modality loss. Under the original three-seed protocol, the pre-specified locked CP-RRA configuration (λrel = 1.0) achieves 76.17% worst Weighted-F1; under complete image or text loss, it recovers more than 44 percentage points relative to the preserved pathway. Post-lock Train/Dev ablation shows that removing explicit reliability supervision (λrel = 0) increases worst Weighted-F1 to 79.70%, while substantially weakening reliability diagnostics, revealing a prediction–diagnostic trade-off. Reliability diagnostics further show that the learned scores identify the usable evidence channel. These results indicate that CP-RRA improves continuity and resilience under asymmetric evidence failure while providing auditable signals for crisis information screening, prioritization, and analyst verification.