Adversarial Confusion Attack: Disrupting Multimodal Large Language Models

ReMatch has been introduced as a new framework that utilizes the generative capabilities of Multimodal Large Language Models (MLLMs) for enhanced multimodal retrieval. This approach trains the MLLM end-to-end, employing a chat-style generative matching stage that assesses relevance from various inputs, including raw data and projected embeddings.

A new benchmark called CaptionQA has been introduced to evaluate the utility of model-generated captions in supporting downstream tasks across various domains, including Natural, Document, E-commerce, and Embodied AI. This benchmark consists of 33,027 annotated multiple-choice questions that require visual information to answer, aiming to assess whether captions can effectively replace images in multimodal systems.

LLaVA-UHD v3 has been introduced as a new multi-modal large language model (MLLM) that utilizes Progressive Visual Compression (PVC) for efficient native-resolution encoding, enhancing visual understanding capabilities while addressing computational overhead. This model integrates refined patch embedding and windowed token compression to optimize performance in vision-language tasks.

A new training framework named Monet has been introduced to enhance multimodal large language models (MLLMs) by enabling them to reason directly within latent visual spaces, generating continuous embeddings as intermediate visual thoughts. This approach addresses the limitations of existing methods that rely heavily on external tools for visual reasoning.

CAPability has been introduced as a comprehensive visual caption benchmark designed to evaluate the correctness and thoroughness of captions generated by multimodal large language models (MLLMs). This benchmark addresses the limitations of existing visual captioning assessments, which often rely on brief ground-truth sentences and traditional metrics that fail to capture detailed captioning effectively.

A new framework named STVG-o1 has been introduced to enhance spatio-temporal video grounding (STVG) by enabling multimodal large language models (MLLMs) to achieve state-of-the-art performance without architectural changes. This framework employs a bounding-box chain-of-thought mechanism and a multi-dimensional reinforcement reward function to improve localization accuracy in untrimmed videos based on natural language descriptions.

Qwen3-VL has been introduced as the latest vision-language model in the Qwen series, showcasing enhanced capabilities across various multimodal benchmarks. It supports interleaved contexts of up to 256K tokens, integrating text, images, and video, with variants designed for different latency-quality trade-offs.

Andrej Karpathy, former director of AI at Tesla and a founding member of OpenAI, created a 'vibe code project' over the weekend, allowing multiple AI assistants to collaboratively read and critique a book, ultimately synthesizing a final answer under a designated 'Chairman.' The project, named LLM Council, was shared on GitHub with a disclaimer about its ephemeral nature.