Why AI inspectors may need a visual guide, not more instructions

A product can look perfectly ordinary at a glance and still break a rule: a compartment may be missing a small part, two similar components may be mismatched, or a liquid level may fall outside its normal range. Detecting these logical anomalies means checking several category-specific constraints against a small set of normal reference images. A single broken rule can make the whole image anomalous.

The study introduces CASCADE, a training-free approach that gives vision-language models (LVLMs) visual aids matched to the constraint being checked. Its central question is whether errors in existing checklist-based methods come mainly from weak reasoning—or from difficulty extracting the right visual evidence from an unchanged image.

A checklist is not the whole picture

Checklist methods turn each rule into a separate question for the model. For instance, the system might ask whether every compartment contains one pushpin, or whether a cable’s length is within the range shown by normal examples. This makes the checks easier to track, but the model still sees the same unmodified query image each time.

The authors identify two recurring difficulties. In binding checks, such as counting or matching, visually similar objects can blur together. In calibration checks, which involve measurements, a model’s sense of scale can shift between images. The team compared two possible fixes: text that encourages more deliberate reasoning, and visual overlays that make relevant evidence easier to read.

In their tests, chain-of-thought and self-refinement prompts made little difference: neither improved accuracy by more than a point on the selected constraints. Visual guides had a much larger effect in these probes. For a counting check, distinctly marking individual instances raised balanced accuracy from 56.7 to 97.4. (Balanced accuracy averages performance on detecting violations and correctly recognizing normal cases.) These results support the authors’ view that, in the tested settings, the bottleneck was perception rather than reasoning over the same image.

Match the guide to the rule

CASCADE chooses between two types of visual scaffold. For binding constraints, it uses segmented objects with distinct colors and labels, making them easier to tell apart and refer to. For calibration constraints, it places a grid on the target object, so measurements can be compared in a consistent frame across reference images and the query.

The basic CASCADE version uses default settings for these guides. CASCADE++ goes further: it derives labels from the relationships specified by a rule and sets the grid using the distribution of measurements in the normal references. That distinction matters. Labels that follow image position may not reflect which objects a rule says should be paired. A grid that is too narrow or too wide may misclassify normal examples or let an abnormal measurement pass.

The researchers tested the systems on MVTec-LOCO and VID-AD, two manufacturing-image benchmarks. In the reported experiments, CASCADE++ used three normal references and no task-specific training. On MVTec-LOCO, it reached 93.2 AUROC—an image-level measure of how well a system ranks anomalous images above normal ones. The paper reports that this was 3.9 points above the full-shot state-of-the-art result cited for that benchmark. On VID-AD, it scored 94.5 AUROC, 11.4 points above the strongest listed baseline. The approach was also evaluated with multiple LVLM backbones, though performance varied between them.

Importantly, the default-scaffold CASCADE did not perform as well: its MVTec-LOCO score was 81.4. The authors report that configuring the scaffolds from the rules and references raised performance substantially. This cautions against treating an overlay as automatically helpful; a poorly matched visual aid can mislead.

What the results do—and don’t—show

These are benchmark results, not evidence that CASCADE is ready for every factory or inspection setting. The study focuses on two kinds of constraints—discrete object relationships and continuous measurements—and says that future work should address constraints requiring both perceptual operations. It also depends on SAM3 to locate objects. In the authors’ tests, errors that split or merge instances reduced performance, with merging especially harmful to counting and matching. The paper reports that small boundary shifts were more easily tolerated than these instance errors.

The authors also note a practical cost: on their reported setup, CASCADE++ took about 71.8 seconds per image, largely due to model API latency. That makes it a poor fit for some rapid inspection lines, even if the accuracy is promising. Results may also depend on the chosen references, prompts, backbones, and benchmark conditions.

The paper is an arXiv preprint, so its findings should be treated as research claims rather than settled evidence. Its reported comparisons are specific to the methods and evaluation protocols described; the available source does not establish how the approach would perform across all products, factories, or real-world conditions. Still, the study offers a useful design lesson: when an AI system must check a visual rule, improving the view of the evidence may matter more than asking the model to think harder about an unchanged image.