Red Teaming Multimodal Language Models: Evaluating Harm Across Prompt Modalities and Models

The recent introduction of Video Retrieval-Augmented Generation (Video-RAG) addresses the challenges faced by large video-language models (LVLMs) in comprehending long videos due to limited context. This innovative approach utilizes visually-aligned auxiliary texts extracted from video data to enhance cross-modality alignment without the need for extensive fine-tuning or costly GPU resources.

A novel framework called Visual Contrast Exploitation (VCE) has been proposed to enhance the safety of autoregressive image generation models, which have gained attention for their ability to create highly realistic images. This framework aims to address concerns regarding the generation of Not-Safe-For-Work (NSFW) content and copyright infringement by introducing a method for constructing contrastive image pairs that effectively decouple unsafe content from the generated images.

Large Language Models (LLMs) like GPT-4o have been evaluated for their effectiveness in assessing the difficulty of programming tasks, specifically through a comparison with a Light-GBM ensemble model. The study revealed that Light-GBM achieved 86% accuracy in classifying LeetCode problems, while GPT-4o only reached 37.75%, indicating significant limitations in LLMs for structured assessments.

A new architecture called Structured Cognitive Loop (SCL) has been introduced to address fundamental issues in large language model agents, such as entangled reasoning and memory volatility. SCL separates cognition into five distinct phases: Retrieval, Cognition, Control, Action, and Memory, while employing Soft Symbolic Control to enhance explainability and controllability. Empirical tests show SCL achieves zero policy violations and maintains decision traceability.

Recent research has focused on the latent reasoning capabilities of large language models (LLMs), specifically through a study on two-hop question answering. The investigation revealed that LLMs, including Llama 3 and GPT-4o, struggle with this basic reasoning task without employing chain-of-thought (CoT) techniques, which are essential for complex agentic tasks.

A new approach called TRIM has been introduced to address the high inference costs associated with Large Language Models (LLMs). This method optimizes language generation by allowing LLMs to omit semantically irrelevant words during inference, followed by reconstruction of the output using a smaller, cost-effective model. Experimental results indicate an average token saving of 19.4% for GPT-4o with minimal impact on evaluation metrics.

Recent advancements in multimodal large language models (MLLMs) and video agent systems have led to the development of SciEducator, an innovative multi-agent system designed for scientific video comprehension and education. This system utilizes the Deming Cycle's iterative approach to enhance the understanding of complex scientific processes through tailored multimodal educational content.

A new framework called Perception Loop Reasoning (PLR) has been introduced to enhance video understanding by addressing the limitations of existing Video Reasoning LLMs, which often rely on a flawed single-step perception paradigm. This framework integrates a loop-based approach with an anti-hallucination reward system to improve the accuracy and reliability of video analysis.