DiaVLo: Diagnosing Behaviours of Vision-Language Models
cs.CL · cs.AI
Quick verdict
DiaVLo matches human-curated behaviour specs to VLM self-explanations, adding DML causal estimates; small samples and inconclusive causal checks temper claims.
Summary
The paper introduces DiaVLo, a framework for diagnosing vision-language model (VLM) behaviours by comparing specifications of desired behaviours (SHOULD-KNOWs) against observed ones (REALLY-KNOWs). SHOULD-KNOWs are built by running scene graph generation (IETrans) on images and curating the resulting triplets through two crowdsourced tasks (verification and expansion; 520 Prolific workers), with expert checks finding at most about 11% needing corrections. REALLY-KNOWs come from VLM self-explanations: models verbalize structured rationales for their answers, formatted as concept-relation triplets. Cosine similarity between sentence embeddings of the two sets, with thresholds tuned via sensitivity analysis and manual inspection (tau_a>=0.75, tau_e>=0.35, tau_d<0.35), assigns each sample to Aligned, Expanded, or Divergent. A causal layer occludes image regions for concept subsets (powerset of REALLY-KNOW concepts), re-runs inference, and fits Double Machine Learning to estimate concept-level effects. On four open VLMs (InternVL2 8B, LLaVA-1.6 7B, Qwen2.5-VL 7B, ShareGPT4V 7B) and four datasets (60-150 curated samples each), 6.4% of behaviours were Aligned, 52.5% Expanded, 41.1% Divergent. Bootstrapped mutual information between behaviour similarity and accuracy was significant (p=0.001) for all 16 model-dataset pairs. Main limitations: reliance on possibly unfaithful self-explanations, small non-representative samples, single-annotator curation without agreement measures, a residual 17.7% misclassification rate among aligned pairs at the chosen threshold, inconclusive causal refutation tests, and no stated code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● 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
- DiaVLo's behaviour labels (Aligned/Expanded/Divergent) correlate with measured VLM accuracy, with bootstrapped mutual information significant at p=0.001 for all 16 model-dataset pairs.
- The four tested 7-8B open VLMs mostly exhibit Expanded (52.5%) or Divergent (41.1%) behaviours; only 6.4% are Aligned with human-curated specifications.
- Counterfactual image occlusion combined with a Double Machine Learning estimator yields concept-level causal effect estimates identifying which concepts steer VLM outputs.
- VLMs tend to describe objects rather than compose spatial relations, organise concept relationships differently from humans, and prioritise certain concepts when answering.
Strengths
Well-motivated gap between benchmarking and interpretability
Sections 1-3 formalise VLM diagnosis via SHOULD-KNOW/REALLY-KNOW triplet specifications, building directly on Balayn et al. (2021) and Sharifi Noorian et al. (2022); Figure 1 positions diagnosis as an addition to benchmarking rather than a replacement.
Careful, bias-aware human curation
520 Prolific workers across verification and expansion tasks (Section 4.1.2, Appendix B); the authors reject LLM-judges citing documented preferential biases (Ye et al. 2025) and avoid VLM-based detection to escape self-preference circularity (Section 4.1.1); expert post-checks quantified residual error at 5.9-11.0% (Table 1, Expert row).
Unusually thorough sensitivity and ablation reporting
Transparent counterfactual design
Honest limitations and ethics treatment
Weaknesses
Behaviour classification carries a large residual error
Causal claims exceed their validation
Small, narrow evaluation with fragile cells
No inter-annotator agreement for the human ground truth
No code release; curated SKs not explicitly promised
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
DiaVLo (Corti and Yang, TU Delft; arXiv 2609.22008) is a diagnostic framework positioned as a complement to benchmarking (Figure 1): instead of aggregating scores, it verifies whether observed VLM behaviours match specifications of desired behaviours, and quantifies which visual concepts steer outputs. Three claims anchor the paper: (1) behaviour labels (Aligned/Expanded/Divergent) correlate with measured accuracy; (2) the tested VLMs exhibit mostly Expanded (52.5%) or Divergent (41.1%) behaviours, with only 6.4% Aligned; (3) counterfactual occlusion plus Double Machine Learning (DML) identifies the most influential concepts. Scope: four open 7-8B VLMs, four datasets (LLaVa-Bench n=60; MMBench 139; SEED-Bench 2 150; VQA v2 150, two categories each), in classification and generation settings.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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 is DiaVLo and what does it do?
DiaVLo is a diagnostic framework for vision-language models. It builds specifications of desired behaviours from scene graphs refined by human crowd workers, captures each model's own structured rationales as observed behaviours, semantically compares the two to label every case as Aligned, Expanded, or Divergent, and estimates how strongly individual visual concepts causally influence outputs by occluding image regions and fitting a Double Machine Learning model.
What did DiaVLo find about VLM behaviour?
Across four open 7-8B VLMs (InternVL2, LLaVA-1.6, Qwen2.5-VL, ShareGPT4V) and four VQA datasets with 60-150 curated samples each, only 6.4% of behaviours were Aligned with the human-curated specifications, 52.5% were Expanded (partially matching plus extra behaviours), and 41.1% were Divergent. Behaviour-similarity scores carried information about task accuracy: bootstrapped mutual information was significant (p=0.001) for all 16 model-dataset pairs.
What are the main limitations of the DiaVLo paper?
The authors flag that VLM self-explanations may not be faithful to internal computation. Samples are small (60-150 per dataset) and drawn from only a few benchmark categories; each sample was curated by a single annotator with no agreement measure; about 17.7% of pairs classified as Aligned at the chosen threshold were found on inspection to be misclassified; causal refutation tests were inconclusive; and only open 7-8B models were tested, with no closed-source models.
Is DiaVLo's code or data available?
The paper does not state a code release. It says crowdsourced data will be released in processed form to protect participant privacy; the curated SHOULD-KNOW specifications and the pipeline code are not explicitly promised, which limits exact replication of the reported behaviour distributions.
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 are extracted as garbled axis/tick text; trend claims for Figures 3-5 and 27-30 rely on the paper's verbal descriptions rather than the plots themselves.
- Behaviour examples in Figures 19-26 are image-based and could not be analyzed from the extracted text.
- Table 2 colour coding for RK extraction success bands is lost in extraction; success rates were reconstructed from the prose.
- Crowdsourcing task screenshots (Figures 6-8) could not be inspected, so interface-driven bias assessments rely on the text alone.
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.