The paper introduces Counterfactual Semantic Saliency (CSS), a black-box, model-agnostic method that measures how the removal of objects from a scene shifts the semantic description produced by vision‑language models (VLMs). Using CSS, the authors compare prominent VLMs to a large human psychophysics baseline (16,603 responses across 307 scenes) and find that VLMs over‑rely on large, centrally located, and highly salient objects, while under‑rely on people compared to humans. The study highlights a significant scene‑comprehension gap between VLMs and human perception, with size bias emerging as a key factor driving model‑human divergence.
arXiv:2607. 09654v1 Announce Type: cross Abstract: Vision language models (VLMs) have made remarkable progress in visual reasoning during the last decade.
arXiv:2308. 06035v4 Announce Type: replace Abstract: Humans routinely draw on visual context to predict upcoming words.
OpenVAM is a new framework for visual attention modeling that combines a dense saliency map with language‑based explanations. It uses a decoupled design: a visual pathway for precise localization and a vision‑language head that generates grounded what/why explanations. The method is trained in three stages to preserve localization while adding language grounding, and a scalable pipeline creates multi‑domain annotations for evaluation.
arXiv:2606. 23763v1 Announce Type: cross Abstract: Recent work typically assesses vision--language consistency using attention distributions of answer-side tokens.
arXiv:2605. 08245v4 Announce Type: replace-cross Abstract: Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imaging to autonomous systems, yet they routinely hallucinate, confidently describing content not present in the input.
The study investigates whether attention weights in Vision‑Language Models (VLMs) accurately reflect model reasoning for visual inputs. Using causal perturbation analysis, it identifies three distinct processing modes—Faithful‑Sufficient, Faithful‑Distributed, and Non‑Focal—indicating heterogeneous visual attention faithfulness. The research also shows that human‑annotated ground‑truth regions align with model attention in only about 60% of cases, highlighting a systematic divergence between model visual reliance and human intuition across VQA, document, and chart tasks.
The paper introduces visual adaptations of counterfactual tests—vCT and vCCT—to evaluate whether chain-of-thought explanations in vision‑language models faithfully reflect the visual evidence driving predictions. Using these tests, the authors benchmark eight open‑source VLMs on two datasets and find that CoTs often fail to track visual evidence, sometimes omitting removed objects or mentioning them inconsistently. They also release two new datasets, Counter‑SNLI‑VE and Counter‑A‑OKVQA, consisting of image pairs that differ by a single object to facilitate further research.
arXiv:2607. 16311v1 Announce Type: cross Abstract: Vision-language models (VLMs) often answer visual questions using learned language and category priors rather than grounding their predictions in the image itself.
The paper presents a quantitative study of how well the attention maps of CNNs, Vision Transformers, and Vision Mamba align with human visual perception, using eye‑tracking data as ground truth. A standardized normalization pipeline maps attribution maps and gaze heatmaps to spatial probability distributions, and alignment is measured with Jensen‑Shannon similarity, Chi‑square similarity, and Pearson correlation across fine‑grained classification, medical imaging, and emotion recognition tasks. Results show that CNNs exhibit more global, diffuse saliency that generally matches human perception, ViTs focus on fine‑grained, patch‑level features, and Vision Mamba performs well in medical tasks but diverges in natural images, with architectural and training choices affecting interpretability.
arXiv:2607. 24017v1 Announce Type: cross Abstract: The empirical success of attention mechanism in Multimodal Large Language Models (MLLMs) often obscures its inherent, subtle flaws.
The paper investigates how vision‑language models (VLMs) handle negation by comparing their attention patterns to human gaze when processing affirmative versus negated statements. Using synthetic and realistic datasets, the authors find that most VLMs do not mimic human-like attention and rely on mechanisms beyond attention, with PaliGemma2‑3B shifting focus to alternatives under negation and LLaVA‑1.5‑7B showing a left‑side spatial bias. The study highlights a performance gap between affirmative and negated queries across downstream tasks, underscoring the need for better negation processing in VLMs.
The paper introduces Salience-LLaVA, a vision‑language model that prioritizes scene elements based on their importance for low‑vision users. It presents three new salience‑aware datasets—Salience COCO, Salience Flickr, and Salience VizWiz—annotated with object‑level salience verified by low‑vision participants. The authors also propose the SCMI metric to evaluate caption ordering accuracy and demonstrate the system’s practicality by deploying it on assistive glasses.