Unleashing the Intrinsic Visual Representation Capability of Multimodal Large Language Models

Knowledge editing has emerged as a method for correcting or injecting specific facts into large language models (LLMs), while fine-tuning is used for adapting these models to new tasks. A critical question arises: do edits survive after fine-tuning? This inquiry is essential for both removing harmful edits and preserving beneficial ones, as the outcome affects the utility and safety of LLMs.

A recent survey on diffusion models for time series and spatio-temporal data highlights their extensive applications across various fields, including healthcare, climate, and traffic management. The study categorizes models based on task type and data modality, aiming to provide a structured perspective for researchers and practitioners.

A novel framework called X-Scene has been introduced for large-scale driving scene generation, focusing on achieving high geometric intricacy and visual fidelity while allowing flexible user control over scene composition. This framework utilizes diffusion models to enhance the realism of data synthesis and closed-loop simulations in autonomous driving contexts.

A novel framework called self-diffusion has been proposed for solving inverse problems, which operates without the need for pretrained generative models. This approach involves an iterative process of alternating noising and denoising steps, refining estimates of solutions using a self-denoiser that is a randomly initialized convolutional network.

A new study titled 'Video-R2: Reinforcing Consistent and Grounded Reasoning in Multimodal Language Models' addresses the challenges faced by multimodal large language models in reasoning over dynamic visual content. The research identifies issues of logical inconsistency and weak grounding in visual evidence, proposing a reinforcement learning approach to enhance reasoning consistency and temporal precision.

PPTArena has been introduced as a benchmark for PowerPoint editing, focusing on reliable modifications to real slides based on natural-language instructions. It encompasses 100 decks, 2125 slides, and over 800 targeted edits, including text, charts, and animations, evaluated through a dual VLM-as-judge pipeline for instruction adherence and visual quality.

The PosA-VLA framework has been introduced to enhance action generation in Vision-Language-Action (VLA) models by utilizing pose-conditioned anchor attention. This approach aims to improve the consistency and precision of target-oriented actions, addressing issues of redundancy and instability in motion generation that have limited the effectiveness of existing models in complex environments.

A new method for detecting policy violations in large language models (LLMs) has been proposed, addressing the urgent need for organizations to align these models with internal policies in sensitive sectors like legal support, finance, and medical services. This training-free approach treats policy violation detection as an out-of-distribution detection problem, enhancing the reliability of compliance mechanisms.