Beyond Benchmarks: Using VLMs to Reveal Systematic Classification Failures Under Real World Conditions
Dieuwertje Alblas, Alma M. Liezenga, Jan Erik van Woerden, Fedor Taggenbrock, Dalia Aljawaheri, Klamer Schutte
cs.CV · cs.AI · cs.ET
Quick verdict
VLM error-slice pipeline recovers injected perturbations on dogs but degrades on military data—an honest, preliminary, baseline-free V&V feasibility study.
Summary
This conference paper from TNO (Netherlands) assesses whether vision-language models (VLMs) can automate error slice discovery for verification and validation (V&V) of image classifiers in defence applications. A five-step pipeline inspired by Domino embeds test images with CLIP PE-Core-L-14-336, subtracts mean class embeddings to shift representation toward context, reduces dimensionality with UMAP, clusters with HDBSCAN, and prompts GPT-5 mini to describe collages in five ranked words. On a Kaggle dog-breed dataset (8,040 images, 93 classes; YOLO26-cls at 90% test accuracy), the method yields nine semantically coherent clusters (silhouette 0.67) and, after five injected perturbations drop accuracy to 38%, groups them into three type-aligned clusters (0.87). On military vehicle data (an 18-class public set, 289 Ukraine video crops, 1,813 restricted EDF STORE crops), clusters instead track source videos and datasets, with overlapping, low-value descriptions. Embedding analyses show much lower variation in military embeddings (2B explains over 90% of variance in 30 principal components versus 140 for dogs). Key limitations: no baseline comparison (Domino itself is not run), hyperparameters tuned by maximizing the same silhouette score used as the headline metric, single-run evaluation with qualitative judgment, and partially restricted data. The authors conclude VLM-based ESD cannot yet automate V&V but may assist human analysts.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 5
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A five-step VLM pipeline (PE-Core embeddings, class-mean subtraction, UMAP+HDBSCAN clustering, GPT-5 mini captions) can produce human-interpretable clusters that surface systematic classification failure conditions.
- On a well-represented domain (dog breeds), the method recovers five artificially injected perturbation types into three type-aligned clusters (silhouette 0.87) regardless of severity, after accuracy falls from 90% to 38%.
- On military vehicle data, clustering is dominated by video- and dataset-specific characteristics and cluster descriptions overlap, yielding little information beyond existing metadata.
- Military-domain embeddings show much lower variation than dog embeddings (2B: over 90% of variance in 30 principal components versus 140 for dogs), consistent with underrepresentation in VLM training data and monotonous surroundings.
- VLM-based ESD is not yet suitable for fully automated V&V but can accelerate human analysts in identifying and characterizing systematic model errors.
Strengths
Candid negative-result reporting
The paper explicitly documents where its method fails: on 2B, clusters align with the three source videos (Fig. 6 bottom-right; Table 2 descriptors overlap) and the authors state the output 'could have been derived from existing metadata'. This honesty raises trust in the positive results.
Controlled perturbation validation
Section 3.1.2 injects five perturbation types at four severities (20–80%), drops dog-classifier accuracy from 90% to 38%, and shows three clusters cleanly separating snow/rain, blur/slide-blackout, and bars-blackout (silhouette 0.87, all samples assigned), with severity-independent assignment — direct evidence the embedding geometry encodes these systematic corruptions.
Mechanistic embedding analysis
Actionable failure-slice discovery in-domain
Complete implementation disclosure
Weaknesses
Hyperparameter selection circularity
No baseline comparison
Classifier-embedder entanglement in military runs
Underpowered quantitative evaluation
Restricted data limits external validation
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (TNO, EDF STORE project) asks whether contrastive vision-language models can automate part of the V&V workflow for image classifiers: discovering and labelling systematic error slices without manual inspection. It proposes a five-step, Domino-inspired slice-then-tag pipeline (Fig. 1): classify test images; embed with CLIP PE-Core-L-14-336; subtract per-class mean embeddings to shift representation from object to context; UMAP (10 dims) plus HDBSCAN; and GPT-5 mini captions 10-image collages with five ranked words (Appendix A). Claims are deliberately modest: the method (i) recovers artificially injected perturbations on a well-represented domain, (ii) degrades on military data where clusters track videos/datasets, and (iii) is not yet fit for fully automated V&V. This honesty is a strength; the contribution is a feasibility assessment, not a validated method.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does this paper propose?
A five-step pipeline that uses a contrastive vision-language model (CLIP PE-Core) to embed test images, subtracts mean class embeddings to highlight context, clusters the results with UMAP and HDBSCAN, and has GPT-5 mini describe each cluster in five ranked words — aiming to automatically surface systematic classification errors for AI verification and validation.
Does the method work on military imagery?
Only partially. On dog photos it recovered injected perturbations well, but on three military vehicle datasets the clusters mostly reflected the source videos and datasets rather than operationally meaningful failure modes, and the generated cluster descriptions overlapped heavily.
Do the authors claim VLMs can fully automate V&V?
No. The paper explicitly concludes the method is not yet suitable for fully automated verification and validation and positions it as a tool to help human analysts find and characterize systematic errors faster, especially for third-party black-box models.
Is the code or data available?
No code release is mentioned. The dog dataset is public (Kaggle), the Ukraine clips come from three public YouTube videos, and the hyperparameters plus the GPT-5 mini prompt are given in the paper, but the EDF STORE military dataset is restricted and cannot be redistributed.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1–11 are available only through their captions and in-text descriptions; visual claims about cluster structure could not be independently inspected from the extraction.
- No truncation observed: the extraction runs complete through the references and the Appendix A prompt.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.