AI models get more consistent at handling rephrased or reworded inputs simply by getting better at the underlying task - no separate robustness training required.
A paper titled "Robustness as an Emergent Property of Task Performance" (arXiv:2602.03344v2, https://arxiv.org/abs/2602.03344) tested multiple models across a range of datasets, varying inputs through paraphrasing and temperature changes. The researchers found a strong positive correlation between task performance and robustness, defined as consistent behavior across semantically equivalent inputs. Their conclusion: robustness isn't a distinct skill layered on top of a model's abilities. It shows up as competence on a task matures.
That reframes a chunk of AI research as possibly premature. If consistency arrives automatically once a model masters a task, treating robustness as a benchmark to chase separately from raw performance starts to look like solving a problem that resolves itself. The paper's authors add a practical note for teams shipping products: models are already reliable on well-established tasks, even where newer benchmarks remain unstable.
Convenient framing for an industry that would rather not audit robustness on its own terms - though the paper's logic cuts both ways. On tasks models haven't mastered yet, don't expect much consistency either.