Research map / Vision-language and multimodal / Visual question answering
Consistency and Uncertainty: Identifying Unreliable Responses From Black-Box Vision-Language Models for Selective Visual Question Answering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes black-box selective visual question answering by measuring answer consistency over model-generated question rephrasings from a small trained VQG probing model, showing it identifies unreliable VLM answers in ID, OOD, and adversarial settings.
Problem
Identifying unreliable answers from API-only black-box vision-language models for selective VQA, where internals, retraining, and validation sets are unavailable, and confidence scores are unreliable under distribution shift and adversarial inputs.
Approach
Use answer consistency over rephrasings of a visual question, generated by a smaller visual question generation model conditioned on the image and the black-box model's predicted answer, as a neighborhood-sampling proxy for predictive uncertainty.
Key results
- On OK-VQA, BLIP answers consistent over 5 rephrasings reach 0.19 coverage at 10% risk versus 0.11 for the all-questions baseline (n≥0).
- Consistency defines a different ordering from confidence: low-confidence answers remain highly consistent, and confidence alone fails on OOD (OKVQA) and adversarial (AdVQA) data.
- The approach still works when the probing VQG model (~200M) is about 55x smaller than the black-box BLIP-2 (11B).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use consistency across generated question rephrasings to select answers | BLIP selective VQA · VQAv2, OK-VQA, and AdVQA | risk at equivalent coverage | Helped |
| Use model confidence alone to select answers | BLIP selective VQA · OK-VQA and AdVQA | error rate: ⇡40% | Hurt |
| Increase the consistency threshold across rephrasings | BLIP selective VQA · AdVQA | coverage at fixed risk | Mixed |
| Use a 200m-parameter VQG probe with the 11B black-box model | BLIP-2 selective VQA · VQAv2, OK-VQA, and AdVQA | risk at equivalent coverage | Helped |
Surprising findings
- Consistency works even though many generated rephrasings are not literally valid paraphrases of the original question.
- Higher consistency is not necessarily rarer; on AdVQA the consistency distribution is nearly flat across levels.
Limitations
- Black-box models studied are relatively small (up to 11B), and the VQG probing model is small (<700M).
- Generated rephrasings are not guaranteed to be valid paraphrases, and the correlation with accuracy is weakest on adversarial data.
Experimental setup
Datasets: VQAv2 Visual Genome OK-VQA AdVQA
Benchmarks: VQAv2 OK-VQA AdVQA
Metrics: risk-coverage accuracy vs consistency coverage at specified risk
Baselines: model confidence (maximum answer probability) BLIP ALBEF BLIP-2
Builds on: Attribution-Based Confidence BLIP Selective classification
Closest papers by meaning
- Unveiling the Tapestry of Consistency in Large Vision-Language Models (NeurIPS 2024 · similarity 0.91)
- Rephrase, Augment, Reason: Visual Grounding of Questions for Vision-Language Models (ICLR 2024 · similarity 0.90)
- SIEVES: Selective Prediction Generalizes through Visual Evidence Scoring (NeurIPS 2026 · similarity 0.89)
- NaturalBench: Evaluating Vision-Language Models on Natural Adversarial Samples (NeurIPS 2024 · similarity 0.89)
- Language Models as Black-Box Optimizers for Vision-Language Models (CVPR 2024 · similarity 0.89)
- AdaptVision: Efficient Vision-Language Models via Adaptive Visual Acquisition (CVPR 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.