arXiv Computer Vision

Did I See It or Think It? Probing Reality Monitoring in Vision-Language Models with Concept Injection

The paper investigates whether vision‑language models can distinguish between concepts that are internally injected and content actually present in an image. Using a new paradigm that independently varies the image and the injected concept, the authors test six VLMs and find that only Gemma‑3‑27B reliably identifies injected concepts, yet its source attribution is fragile. The study shows that injected concepts often appear in generated captions, highlighting the need to evaluate both content and source in multimodal self‑reports.

arXiv Computation and Language
Nov 6

Vision-Default, Prior-Override: Causal Mechanisms of Perception-Knowledge Conflict in Vision-Language Models

The paper investigates how vision‑language models resolve conflicts between visual evidence and memorized knowledge. Using activation patching and ablation across residual streams, attention heads, and MLP sublayers, the authors find that visual grounding is the default behavior, while prior knowledge relies on a small, causally essential set of attention heads concentrated in the network’s second half. Ablating these heads flips knowledge‑grounded predictions to visually grounded ones in most cases, revealing an asymmetric causal structure and a sparse circuit of routing and writing heads that underlies perception‑knowledge conflict.

By Niclas Lietzow, Danielle Bitterman, Carsten Eickhoff, William Rudman, Michal Golovanevsky
arXiv AI
Oct 21

CounterCount: Diagnosing Knowledge Bias in How Vision-Language Models Count Attributes

CounterCount is a diagnostic dataset that pairs factual and counterfactual images to test whether Vision‑Language Models (VLMs) rely on visual evidence or on world knowledge when counting attributes. Humans achieve near‑perfect accuracy on both types of images, whereas VLMs drop from 91.6% to 49.5% on counterfactuals, often defaulting to canonical counts. The study shows that VLMs exhibit a System‑1‑like bias, underutilize visual evidence, and that targeted input manipulations can mitigate this bias.

By Reem Alzahrani, Bushra Bin Hemid, Hassan Alshanqiti, Zaid Alyafeai, Paul Gavrikov, Bernard Ghanem, Abdelrahman Eldesokey
arXiv AI
Sep 10

Counterfactual Tests for Measuring Chain-of-Thought Faithfulness in Visual Language Models

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.

By Bayar Menzat, Maximilian S\"uss, Ruizhi Wang, Benno Steinegger, Thomas Lukasiewicz, Oana-Maria Camburu
arXiv AI
Aug 10

Probing Visual Concepts in Lightweight Vision-Language Models for Automated Driving

arXiv:2603. 06054v2 Announce Type: replace-cross Abstract: The use of Vision-Language Models (VLMs) in automated driving applications is becoming increasingly common, with the aim of leveraging their reasoning and generalisation capabilities to handle long-tail scenarios.

By Nikos Theodoridis, Reenu Mohandas, Ganesh Sistu, Anthony Scanlan, Ciar\'an Eising, Tim Brophy
arXiv Computation and Language
Oct 12

When VLMs Destroy What They Know: Spatial Representations Fall Below Lexical Priors

The paper introduces EICS, a metric for assessing whether Vision‑Language Models (VLMs) consistently encode concepts under controlled one‑concept interventions, and CASP‑VL, a benchmark of causally controlled spatial perturbations derived from Visual Genome scene graphs. Experiments across multiple VLM architectures reveal that spatial relations are encoded far less faithfully than lexical information, even when models perform well on compositional benchmarks. The study shows that relational information is retained at the predicate token level but disappears when sentences are aggregated into a global representation, highlighting a systematic failure invisible to output‑based evaluation.

By Shubh Garg, Chen Shani
arXiv Computation and Language
Oct 19

Would you still call this Dax? Novel Visual References in VLMs and Humans

The paper introduces the Novel Visual References Dataset (NVRD), comprising 19,176 images of 90 new visual concepts with up to 20 perturbed versions each, designed to test how vision‑language models and humans map novel visual references to language. The authors evaluate three open‑source and two closed‑source VLMs against 2,400 human judgments, finding that models struggle to learn new concepts that contradict pre‑training knowledge and tend to overgeneralize labels to stimuli humans reject. NVRD is released as a benchmark for studying visual concept learning in both humans and machines.

By Ada Defne T\"ur, Gaurav Kamath, Joyce Chai, Siva Reddy, Benno Krojer
arXiv Machine Learning
Sep 10

I Don't Miss You, but I Do: Self-Explanation Faithfulness of Modality Missingness in Vision-Language Models

The paper introduces an interventional protocol to assess how vision‑language models (VLMs) explain the impact of missing modalities on their predictions. By comparing the models’ self‑explanations with actual changes observed after restoring missing inputs, the study finds that VLMs routinely overstate the sufficiency of available evidence and underestimate the effect of adding back missing modalities. Across eight open‑weight VLMs and four tasks, the discrepancy between predicted and realized changes is substantial, revealing systematic mischaracterization of modality dependence.

By Aydin Javadov, Daniel Schoess, Florian von Wangenheim