Long-form question answering systems still struggle with factoid hallucinations—outputs that read fluently but are factually wrong. A common remedy is model alignment, where models are trained to prefer factual responses over non-factual ones. A new arXiv preprint proposes a more granular variant of this idea.
Atomic Consistency Preference Optimization (ACPO) shifts the focus from whole answers to atomic claims. Instead of comparing entire responses, the method trains on curated pairs of factual and non-factual statements, encouraging the model to maintain consistency at the level of individual verifiable facts.
Because the abstract is short, the paper's full experimental design and results are not described here. The contribution, as stated, is an alignment objective that targets the smallest units of factual correctness, which could offer a more precise way to reduce hallucinations in long-form generation.