arXiv:2610.27756v1 Announce Type: new Abstract: Vision-language models (VLMs) are exact on small sets and drift into estimation on larger ones, a profile reminiscent of children before they master co...
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.
arXiv:2606. 01287v1 Announce Type: cross Abstract: Recent latent visual reasoning methods achieve substantial gains by inserting continuous latent tokens into multimodal language models.
The paper shows that controlled visual in‑context learning can help vision‑language models (VLMs) overcome errors caused by training priors. By providing demonstrations, the models focus more on visual evidence, yielding significant accuracy gains on counterfactual examples while leaving prior‑aligned performance largely unchanged. This effect is observed across five VLM families and five vision‑centric tasks, indicating that many VLMs possess latent visual grounding abilities that standard tests may miss.
arXiv:2607. 12815v1 Announce Type: new Abstract: Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces.
Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces. We ask whether CoT requires continued access to image tokens, or whether it mainly operates over visual information already made available earlier in the forward pass.
arXiv:2604. 14888v3 Announce Type: replace-cross Abstract: Recent advances in vision language models (VLMs) offer reasoning capabilities, yet how these unfold and integrate visual and textual information remains unclear.
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.
The paper investigates why Vision Language Models (VLMs) struggle with abstract reasoning tasks. Using the Relational Match-to-Sample paradigm and mechanistic analysis, the authors identify four factors—capability tier, model scale, number of objects per scene, and lack of per-object noise—that influence VLM performance. They uncover two competing neural circuits: an early surface-feature circuit and a late relational circuit, and show that ablating relational heads degrades performance more than random heads, suggesting the relational circuit is broadly used.
arXiv:2609.22588v1 Announce Type: new Abstract: Vision-language models (VLMs) perform strongly on visual question answering benchmarks, yet often make decisions that contradict visual evidence they h...
ReWEIGH the Evidence is a training‑free decoding technique that calibrates token‑level ordinal visual evidence to reduce hallucinations in large vision‑language models. It aggregates vocabulary ranks across visual positions, compares candidates to a token‑specific reference derived from unlabeled images, and applies a bounded penalty only when evidence falls below this reference. Experiments on four 7B backbones show up to a 21.3% reduction in hallucinated object mentions while largely preserving or improving descriptive and general performance, with minimal added latency.
SpecialEduBench is a new benchmark for vision‑language models that evaluates their pedagogical competence in language intervention for autistic children across knowledge, skill, and attitude dimensions. It contains 4,537 knowledge items, 200 skill items, and 68 attitude items derived from recorded interventions, with 192 response cells for attitude items that combine pressure and monitoring. Eight leading vision‑language models were tested, none achieving full performance, especially in honesty cells, indicating that current models still struggle with situated teaching tasks.