Researchers propose a paraphrase generation framework, HALLUCINATION-R1, to improve factual consistency in AI models. This framework generates semantically faithful yet robustness-challenging paraphrases, which can help expose factual instability in downstream QA models.