This paper investigates how real-world shortcuts manifest across different layers of MedCLIP, a medical Vision-Language Model, and its vision encoder. By attaching linear probes and analyzing calibration and layer-wise confidence, the study reveals that despite high performance, the model is vulnerable to shortcuts, underscoring the critical need for high-quality and well-annotated datasets in medical AI.

This paper introduces MemJack, a memory-augmented multi-agent framework designed to systematically expose visual-semantic vulnerabilities in Vision-Language Models (VLMs). It orchestrates automated jailbreak attacks using unmodified natural images by dynamically mapping visual entities to malicious intents and leverages a persistent memory to transfer successful strategies across different images.

This paper investigates how to effectively utilize Vision-Language Models (VLMs) for image forgery detection and localization (IFDL). It introduces IFDL-VLM, a novel two-stage pipeline that decouples the core IFDL task from VLM-based explanation generation, leveraging localization masks to significantly enhance VLM interpretability and achieving state-of-the-art results across multiple benchmarks.

This paper investigates the effectiveness of Explainable Artificial Intelligence (XAI) methods in making Vision-Language Models (VLMs), specifically CLIP, interpretable within art-historical contexts. It evaluates seven XAI methods through zero-shot localization experiments and human interpretability studies, concluding that their effectiveness depends on the conceptual stability and representational availability of the examined categories.

This paper introduces Stage-Aware Reinforcement (StARe), a novel module that decomposes long-horizon robotic manipulation tasks into semantically meaningful stages, providing dense, interpretable reinforcement signals. Integrated into the Imitation → Preference → Interaction (IPI) fine-tuning pipeline, StARe significantly improves the performance and robustness of Vision-Language-Action (VLA) models on complex manipulation tasks.

This paper introduces Chameleon, an adaptive adversarial framework that exploits image downscaling vulnerabilities in Vision-Language Models (VLMs) to inject hidden malicious visual prompts. By employing an iterative, feedback-driven optimization mechanism, Chameleon can craft imperceptible perturbations that hijack VLM execution and compromise agentic decision-making systems.

This paper introduces MemLoRA, a novel memory system that enables efficient, on-device deployment of memory-augmented language models by equipping small models with specialized LoRA adapters. It also presents MemLoRA-V, extending these capabilities to multimodal contexts with native visual understanding.

This paper introduces SpaceTools, a vision-language model trained with Double Interactive Reinforcement Learning (DIRL) to achieve precise spatial reasoning and real-world robot manipulation by effectively coordinating multiple external tools. It demonstrates state-of-the-art performance on various spatial understanding benchmarks.

AdaptVision introduces an efficient VLM paradigm that autonomously determines the minimum number of visual tokens required for each sample by employing a coarse-to-fine visual acquisition strategy, leading to superior performance with significantly reduced computational overhead.

DeepSeek-OCR explores the innovative approach of compressing long text contexts into visual tokens to overcome computational challenges in Large Language Models. This vision-language model, featuring a novel DeepEncoder and DeepSeek3B-MoE decoder, demonstrates impressive OCR precision with significant compression ratios and state-of-the-art performance on document parsing benchmarks.