EDCT-Bench: Uncovering Faithfulness Gaps in VLMs via Explanation-Driven Counterfactual Testing
Sihao Ding, Santosh Vasa, Aditi Ramadwar, Thomas Monninger
cs.CV · cs.AI
Quick verdict
EDCT-Bench tests whether VLM explanations stay consistent when their cited visual evidence is edited; a well-validated, adaptive protocol, but the benchmark is small, scores are judge-dependent, and failed edits are silently excluded.
Summary
The paper introduces EDCT, an intervention-based protocol testing whether a VLM's answer and natural-language explanation remain consistent when the visual evidence cited in the explanation is minimally edited. The pipeline extracts cited visual concepts with an LLM (k=2), generates counterfactual images with FLUX.2 or Gemini Image, verifies edits are localized (Moondream3 detection plus SAM2 masks with inside/outside change thresholds after alignment), re-queries the model, and has an LLM judge score Answer Consistency and Explanation Faithfulness, whose product is the Counterfactual Consistency Score (CCS); a 1-5 rubric variant is also reported. EDCT-Bench contains 300 manually curated image-question pairs from OK-VQA (160), DriveLM (80), and 3DSRBench (60). Six VLMs show large gaps: the best, Gemini 2.5 Flash, reaches only 0.600 binary CCS; others fall in 0.435-0.566, with 3DSRBench hardest and explanation consistency consistently below answer consistency. Validation includes human agreement (94.3% with human consensus, 83% three-way), five-run stability (SD about 0.03), and a judge/editor ablation showing judge choice shifts scores more than editor choice. A preliminary 48-sample fine-tuning study finds 4x higher initial loss and more localized attention on counterfactuals, but no post-finetuning CCS. Key limitations: small per-domain samples with wide CIs, model-specific edit difficulty complicating cross-model comparison, LLM-judge dependence, exclusion of failed edits, and an underpowered training study without a control arm.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- All six evaluated VLMs exhibit substantial counterfactual consistency failures, with even the best model (Gemini 2.5 Flash) reaching only 0.600 binary CCS and others between 0.435 and 0.566.
- EDCT's automated binary CCS approximates human judgment reliably, matching human-consensus labels on 94.3% of doubly-agreed interventions and remaining stable across five pipeline runs.
- Strong general multimodal capability does not imply explanation faithfulness: CCS does not uniformly track MMMU/MMLU/MathVista scores or CoT-Bias diagnostics.
- EDCT-generated counterfactuals are high-impact training examples for Qwen3-VL-8B, producing ~4x higher initial loss (0.7124 vs. 0.1792) and qualitatively more localized attention, though held-out faithfulness gains remain unverified.
Strengths
Principled treatment of valid answer invariance
Sec. 3.4 explicitly defines necessity/sufficiency w.r.t. the edited scene and accepts unchanged answers when remaining evidence is sufficient (headlights/tires/windshield example), avoiding the common flaw of naively demanding an answer flip after every edit; the judge prompt in Appendix B encodes this in the rubric.
Edits are verified, not assumed
Algorithm 1 quantifies edit localization (change fraction inside a SAM2 mask vs. spillover outside, with ORB alignment and optical-flow warping to absorb global distortion), and Sec. 4.4 reports automatic edit-validity agreement with humans (89.8-95.3%) comparable to inter-annotator agreement (86.7%).
Unusually thorough validation for a benchmark paper
Honest epistemic framing
Adaptive, model-specific interventions
Weaknesses
Silent exclusion of failed edits creates selection bias
Scores are judge-conditional, with a possible family-bias confound
Small per-domain samples yield wide intervals
Training-signal study is underpowered and uncontrolled
No code or data release stated; heavy reliance on proprietary APIs
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- 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 EDCT-Bench?
EDCT-Bench is a 300-item benchmark for testing whether vision-language models (VLMs) give explanations that are actually consistent with the images they describe. For each image-question pair, the system reads the model's explanation, edits the visual evidence the explanation cites (using AI image editors), verifies the edit changed only the intended region, re-asks the model, and scores whether the new answer and explanation stay logically consistent with the edited image. It draws 160 items from OK-VQA, 80 from DriveLM (driving scenes), and 60 from 3DSRBench (3D spatial reasoning).
What does faithfulness gap in VLM explanations mean?
A faithfulness gap means a model's natural-language explanation sounds plausible but does not reflect the visual evidence actually driving its answer - the explanation may be a post-hoc justification. EDCT tests this behaviorally: if a model says it identified a tennis racket because of its strung surface, replacing that surface with a solid paddle-like one should change the answer or force the explanation to cite different evidence; models that keep answering and citing strings are failing the counterfactual test.
How reliable is the automated scoring compared to humans?
On 128 human-annotated interventions with two annotators, the automated binary score agreed with the shared human label in 94.3% of the 88 cases where the two annotators agreed with each other; unconditional three-way agreement was 83.0%. Individual annotator-to-system agreement was 86.0% and 88.4%, close to the 88.0% human-human rate. Aggregate scores were also stable across five independent pipeline runs (standard deviation about 0.036).
Which VLM performed best on EDCT-Bench?
Gemini 2.5 Flash scored highest with a binary Counterfactual Consistency Score of 0.600 (rubric-graded 0.603), but this still means it failed roughly 40% of counterfactual consistency tests. The five open-weight models evaluated (Pixtral-12B, Gemma3-27B, Qwen3-VL-30B-A3B, Qwen3-VL-8B, GLM-4.6V-106B) scored between 0.435 and 0.566, and 3D spatial reasoning (3DSRBench) was the hardest domain for every model.
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.
- Review is based on the extracted text only; figures (pipeline diagrams, attention maps in Fig. 4/9, the CCS-vs-benchmark scatter plots in Fig. 5, and qualitative edit examples) could not be visually verified.
- Table values and confidence intervals are taken as extracted; minor OCR garbling is possible (e.g., subscript formatting, the boxed 0.435-0.566 range).
- No code/data release statement was found in the extracted text; if such a statement exists in a truncated portion (e.g., a footnote or supplementary link), codeAvailable would be understated.
- Appendix prompts are quoted in full, but judge calibration, rubric inter-judge variance, and retry bookkeeping details are not fully specified in the text.
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.