Automated tool finds and fixes AI safety systems wrongly rejecting safe questions

Original: DDOR: Delta Debugging for Explainable Overrefusal Testing and Repair

Source: arxiv.org ↗