The field of vision-language models is rapidly advancing, with a focus on improving their ability to reason and understand complex visual and linguistic concepts. Recent research has highlighted the importance of addressing hallucinations, which occur when models generate contradictory content to their input visual and text contents. Innovations in contrastive decoding, attention manipulation, and explainability analysis are being explored to mitigate this issue. Furthermore, studies are investigating the cognitive limits of vision-language models, including their ability to count objects, reason about physical dynamics, and understand compositional counting. Noteworthy papers include: MaskCD, which proposes a novel method for mitigating hallucinations in large vision-language models. Reasoning Riddles, which conducts a comprehensive explainability analysis to understand how vision-language models approach complex lateral thinking challenges. A Study of Rule Omission in Raven's Progressive Matrices, which investigates the generalization capacity of modern AI systems under conditions of incomplete training. Zero-Shot Fine-Grained Image Classification Using Large Vision-Language Models, which presents a novel method for zero-shot fine-grained image classification using large vision-language models. Your Vision-Language Model Can't Even Count to 20, which exposes the failures of vision-language models in compositional counting. More Than Meets the Eye, which uncovers the reasoning-planning disconnect in training vision-language driving models. Aligning Perception, Reasoning, Modeling and Interaction, which provides a comprehensive overview of physical AI and establishes clear distinctions between theoretical physics reasoning and applied physical understanding. Visual Representations inside the Language Model, which examines how popular multimodal language models process their visual key-value tokens. Does Physics Knowledge Emerge in Frontier Models, which benchmarks six frontier vision-language models on three physical simulation datasets. ChainMPQ, which proposes a training-free method for mitigating relation hallucinations in large vision-language models.