Mask-to-Correct: AI Framework for Fact Correction Using Retrieval Diversity

Researchers propose Mask-to-Correct (M2C), a training-free, inference-only framework for automated fact correction that leverages diversity-aware masking to identify erroneous spans in claims and evaluate correction faithfulness using retrieved evidence. The method addresses limitations of existing supervised learning approaches that rely on scarce, biased manually annotated claim-evidence pairs. To mitigate retrieval bias, the ensemble-based M2C+ combines corrections across multiple rankers. The work is published on arXiv under ID 2605.18776.