Vision-language models are usually stress-tested with degradation levels labelled 1–5, but those labels carry no physical units. A new arXiv preprint argues that such nominal scales make it hard to know whether a model's confidence actually tracks how much it can still see.

The paper proposes a physical visibility axis and examines accuracy, confidence, and abstention along it. The authors ask whether a VLM knows when it cannot see, rather than simply measuring how its accuracy degrades.

By tying robustness to a physical quantity, the work could give developers a clearer signal for when to trust a model's output or allow it to abstain. The abstract does not report specific results, but the framing points to a gap in current evaluation practice.