Factual hallucination is usually defined as the generation of incorrect facts. This paper studies a paraphrase‑induced hallucination scenario: a model answers a factual question correctly in its original form but produces a wrong answer under a semantically equivalent paraphrase. Such inconsistencies reveal latent factual instability under semantic invariance. Generic paraphrases often fail to provide effective robustness supervision—near‑copy paraphrases give weak signals, while overly diverse ones may break semantic equivalence. To address this, we propose HALLUCINATION‑R1, a robustness‑oriented paraphrase generation framework that learns to produce paraphrases that preserve meaning yet challenge downstream QA models' factual consistency. The framework uses a two‑stage optimization: the first stage stabilizes meaning‑preserving and diverse paraphrasing via semantic constraints and diversity rewards; the second stage adds extra reward to paraphrases that cause factual consistency degradation in downstream models. Experiments on SimpleQuestions, PopQA, and TruthfulQA demonstrate that HALLUCINATION‑R1 achieves a strong consistency‑diversity trade‑off and exposes robustness failures across multiple model families and datasets. Further analysis shows these failures are not reducible to surface‑level artifacts or semantic drift, but reflect non‑trivial factual instability under meaning‑preserving variation. A lightweight fine‑tuning study indicates that training with HALLUCINATION‑R1‑generated data improves robust accuracy under paraphrase variations, highlighting its utility for robustness‑oriented training. Code and models are publicly available at https://github.com/yuwenhan07/Hallucination-R1.
Review: The work offers a systematic approach to probe and improve factual robustness by coupling controlled paraphrase generation with targeted rewards, providing reusable data for robustness‑focused model training.