Logical Anomaly Detection (LAD) requires verifying whether all category-specific constraints, such as object counts, pairing relations, and measurement ranges, are satisfied by a query image given only a few normal references. Recent LAD methods based on Large Vision-Language Models (LVLMs) formulate this as checklist verification, decomposing the decision into independent textual checks. This decomposition, however, is text-only. Every item still reads the same unmodified query image, so the LVLM must extract different forms of visual evidence from one raw image. We identify this text-image mismatch as a constraint-specific perception gap, the source of two failure modes: instance conflation for binding constraints and scale drift for calibration constraints. To diagnose whether these failures stem from insufficient reasoning or insufficient perception, we compare textual reasoning guides and visual perception guides in a training-free setting. Our analysis shows that reasoning over the same image provides little benefit, whereas adapting the visual input to each constraint substantially improves verification accuracy. Based on this finding, we propose CASCADE, a training-free framework that extends checklist decomposition using constraint-specific visual scaffolds. We further introduce CASCADE++, which adapts these scaffolds at the object level by deriving their configurations from the constraint structure and reference distribution. Across MVTec-LOCO and VID-AD, CASCADE++ surpasses the full-shot state of the art by 3.9 AUROC points with only three references, and improves over the strongest VID-AD baseline by 11.4, transferring across multiple LVLM backbones without any training.
Figures & tables
Fig. 1 : Checklist verification on a raw query versus a scaffolded query—per-instance coloring for counting, an object-anchored grid for measurement. Both are the same image verified against the same checklist item, and both violate their constraint, yet only the scaffolded query yields the correct verdict. The Measurement condition is defined by the K=3 references shown in the band rather than stated explicitly (Section VII-B ).
Fig. 2 : Method comparison between the baseline and CASCADE. Both methods use checklist generator and checklist verifier for verdicts. Unlike the baseline, ours overlays a constraint-specific scaffold on the query image, so each constraint is verified with an adapted visual input.
Fig. 3 : Comparison of textual reasoning and visual perception guides. Bars show balanced accuracy on a binding constraint for Counting and the average over calibration constraints for Measurement, with the dashed lines indicating the text-only baselines. Each overlay is applied to both constraint types, not only the one it was designed for.
Fig. 4 : Comparison of raw and per-instance colored images on a counting probe. We query the count of n identical circles on a blank canvas, varying only n . Raw-image accuracy (red) collapses past n≈10 , while distinctly coloring each circle (gray) keeps it far higher throughout.
Fig. 5 : Comparison of within-image and cross-image length readings on a scale probe. We render a bar of fixed true length with a ground truth of 320 px and read its length from the LVLM. A within-image reading repeats the same image, while a cross-image reading spans separate renderings of the same length.
Fig. 6 : Comparison of readout strategies on a range-membership probe. Balanced Accuracy across violation margins for three VLM backbones. A test bar is judged against the range of K=3 references, calibrated in code as in CASCADE , so the backbone reads only the test: Comparison sees the references too and answers directly, Pixel reports raw pixels, Patch units of an imagined 12×12 grid, and the object-centric readout an index off a scale drawn on the bar; the last three are range-checked in code.
Channel
Numbering
Count (PP)
Linking (SC)
None
—
56.7
61.4
Color only
—
79.2
74.3
ID only
offset
75.8
38.6
ID only
constraint-derived
79.3
86.2
SoM (color + ID)
sequential
97.4
66.7
Visual Sketchpad
self-scaffolding
49.8
69.5
TABLE I : Disambiguation channel decomposition for binding constraints on Balanced Accuracy. We vary the two components of the overlay, color and a numeric ID, together with the ID’s numbering scheme, holding the question fixed so that only the overlay changes. Count is the Pushpins (PP) count constraint, and Linking the Splicing-Connector (SC) slot-matching constraint.
Constraint
Fixed
Dyn-size
Dyn- 2×
aadv
CASCADE++
BB-Cereal
77.6±9.9
71.4±6.7
78.6±10.1
63.8±8.6
82.9±4.0
JB-Fruit
90.0±11.3
95.9±1.9
76.9±6.1
85.7±9.5
94.4±3.9
JB-Text
83.3±6.6
81.4±10.6
82.6±2.3
84.3±10.9
96.9±0.1
JB-Liquid
97.8±1.0
25.6±7.9
72.1±0.9
84.0±4.3
95.9±0.2
SB-Length
92.2±0.4
50.0±2.7
96.4±0.5
86.7±2.0
92.4±4.9
Avg.
88.2
64.9
81.3
80.9
92.5
TABLE II : Calibration rule ablation on Balanced Accuracy (mean ± std over five random reference selections). Grid parameterization per strategy: Fixed (w=100,a∗) , Dyn-size (w=tgt/2,a∗) , Dyn- 2×(w=tgt/4,a∗) , aadv(w∗,adversariala) , and CASCADE++ (w∗,a∗) . Bold marks the best entry in each row.
Method
Backbone
Avg
BB
JB
PP
SB
SC
Full-shot
PatchCore [ 2 ]
CNN
74.0
74.8
93.9
63.6
57.8
79.2
GCAD [ 1 ]
CNN
86.0
87.0
100.0
97.5
56.0
89.7
EfficientAD [ 3 ]
CNN
86.8
85.5
98.4
97.7
56.7
95.5
AST [ 25 ]
CNN
79.7
80.0
91.6
65.1
80.1
81.8
LogSAD [ 17 ]
GPT-4o
89.3
—
—
—
—
—
TABLE III : Image-level AUROC (%) on the MVTec-LOCO logical subset. GPT-5 denotes GPT-5-image-mini and Gemini-3.1-FL denotes Gemini-3.1-Flash-Lite. Rows without a marker are quoted from the original papers; † marks results we produced under each method’s own inference procedure, and * marks a method whose design admits only a single reference, where the other few-shot methods use K=3 . Bold marks the best entry in each column within each of the two regimes; dashes mark entries the source paper does not report.
Method
Seg (s)
LVLM (s)
Total (s)
LogicAD [ 5 ]
0.2
52.6
52.8
LogicQA [ 4 ]
0.1
186.5
186.6
LogSAD [ 17 ]
2.7
0.0
3.4
CASCADE
3.2
51.4
55.0
CASCADE++
3.2
68.3
71.8
TABLE IV : Per-image inference cost (s) on MVTec-LOCO, all on GPT-5-image-mini. Seg covers each method’s segmentation stage and LVLM sums the per-call API latencies; Total is the full per-image wall clock, which exceeds Seg + LVLM by the remaining compute. LogSAD’s checklist runs once per category, so its per-image LVLM is zero. The feature-based detectors run locally in milliseconds and are not comparable on this axis.
Method
W
C
M
L
B
Mean ± Std
Full-shot
PatchCore [ 2 ]
40.4
45.0
48.0
46.9
42.0
44.5±2.9
EfficientAD [ 3 ]
52.6
60.2
60.8
47.9
48.0
53.9±5.7
PaDiM [ 28 ]
34.8
35.4
42.8
35.4
40.3
37.8±3.2
VAE [ 29 ]
44.8
49.4
45.4
47.2
39.3
45.2±3.4
AnoGAN [ 30 ]
48.9
54.8
49.2
47.4
46.8
49.4±2.8
TABLE V : Main results on VID-AD (image-level AUROC, %). W, C, M, L, and B denote white, cable, mesh background, low-light, and blurry conditions, respectively. Full-shot methods use the complete training set, whereas CASCADE++ uses K=3 references with no training. Bold marks the best entry in each column within each of the two regimes.
Configuration
Disambig.
Alignment
Binding
Calibration
Overall
No scaffold
—
—
72.6
62.7
74.5
Disambiguation only
✓
—
80.9
59.9
81.8
Alignment only
—
✓
72.9
70.0
79.5
CASCADE++ (both)
✓
✓
83.8
74.5
93.2
TABLE VI : Decomposition ablation on MVTec-LOCO (image-level AUROC, %). All rows use the CASCADE++ configuration rules and differ only in which scaffolds are enabled; “No scaffold” verifies the same checklist on the raw query image. The Binding and Calibration columns score each image by the violated fraction of binding-only and calibration-only sub-constraints, and so miss anomalies of the other type; Overall uses all sub-constraints, which is why it exceeds either column alone.
Large Vision Language Models (VLMs) excel at visual question answering and semantic grounding, but their capacity for autonomous logical reasoning remains underexplored. Existing anomaly benchmarks emphasize visual errors or direct prompting rather than the physical and social common sense needed for open-world deployment. To address this, we introduce LAD-bench, a benchmark of more than 1,000 curated synthetic images with logical anomalies across four domains: Residential, Urban, Collaborative, and Nature. We further propose a Tiered Prompting Protocol based on progressive disclosure, which measures how much explicit assistance a model needs to localize and reason about a logical fault. Evaluating leading foundation models reveals substantial weaknesses: even the best achieves only 70.11% overall accuracy, showing that implicit logical fault detection remains unsolved. Crucially, models often fail to identify anomalies even after receiving explicit hints in deeper tiers. By surfacing these limitations in sequential multimodal reasoning, LAD-Bench offers a rigorous framework for advancing the safety, reliability, and cognitive alignment of autonomous visual systems. Dataset and Code: https://huggingface.co/datasets/SahasraK/LADBench
Logical anomalies arise when every component looks normal but their combination violates the scene rules. In practical inspection, these rules must hold under varying backgrounds, illumination, and imaging quality, with only a few dozen normal images per scene, whereas existing methods couple normality with appearance and merge different constraints into one reference. We model logical normality as decomposable constraints and propose EviRef, a training-free method that builds a separate reference memory for each family of evidence. Component prototypes built once per scene parse images under any capture condition into one label space, on which occupancy and relation constraints are verified. Semantic constraints are verified with whole-image embeddings independent of parsing, and additional evidence such as text is admitted through a task-level gate without modifying existing memories. On VID-AD, EviRef attains a mean image-level area under the receiver operating characteristic curve (AUROC) of 92.1% (88.9% with visual evidence alone), the highest among the compared methods under each of five capture conditions, with a standard deviation of 0.5. Even with memories built from only 20 normal images per task, it surpasses all compared methods that use 50 normal images per task.
Yaoting Huang, Wenqi Cai, Yijin Wei +4
University of Toyama · NingboTech University · University of Fukui +1
Logical anomaly detection requires recognizing scene components and measuring whether their configuration satisfies task constraints. Correct identities can coexist with inaccurate geometry; clean mask fragments can yield misleading lengths. We propose Semantic-Geometric Decoupling and Evidence Binding (SGDB). A semantic decoder learns object identities from component-annotated normal sources and factual edits; a separate reader supplies final measurements. Partial one-to-one association binds geometric witnesses to identities while retaining unmatched objects in the count. A support-certified reader selects mask-derived axes using line quality and complete component ownership. Explicit task requirements and normal-data references define executable checks, with decision thresholds fixed by independent normal calibration. Execution-grounded explanation traces rule contributions to source objects and measurements, extracting compact sufficient alarm certificates. A frozen language model expresses selected outcomes as cited claims, with numeric evidence retained by the executor. Reports preserve observed deviations and unavailable evidence. SGDB achieves 98.19% AUROC and 94.52% fixed-threshold F1 on five MVTec LOCO logical categories, and 93.56% AUROC and 92.76% F1 across all 50 VID-AD tasks. Frozen-output controls evaluate geometric readout and explanatory sufficiency and compactness.
Chenrui Ji, Jiafan Lu, Zilu Luo +3
School of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237, China