Consolidating Diffusion-Generated Video Detection with Unified Multimodal Forgery Learning

A new framework called AdaTok has been introduced to enhance the efficiency of Multimodal Large Language Models (MLLMs) by employing an object-level token merging strategy for adaptive token compression. This approach significantly reduces the number of tokens used, achieving approximately 96% of the performance of traditional models while utilizing only 10% of the tokens, addressing computational and memory burdens associated with patch-level tokenization.

A new method for enhancing social interaction understanding in videos has been proposed, focusing on the alignment of verbal and non-verbal cues in multi-speaker scenarios. This approach addresses the limitations observed in existing Multimodal Large Language Models (MLLMs), which struggle with cross-modal attention consistency in such contexts.

The introduction of ChineseVideoBench marks a significant advancement in the evaluation of Multimodal Large Language Models (MLLMs) specifically for Chinese Video Question Answering. This benchmark provides a comprehensive dataset and tailored metrics, addressing the need for culturally-aware evaluation frameworks in video analysis.

A new study has introduced a multimodal visual understanding dataset (MSVQA) aimed at addressing catastrophic forgetting in Multimodal Large Language Models (MLLMs) by adapting to various scenarios such as high altitude, underwater, low altitude, and indoor settings. The proposed method, UNIFIER, seeks to enhance visual learning by decoupling visual information into distinct branches within each vision block.

The introduction of DocPTBench marks a significant advancement in the benchmarking of end-to-end photographed document parsing and translation, addressing the limitations of existing benchmarks that primarily focus on pristine scanned documents. This new benchmark includes over 1,300 high-resolution photographed documents and eight translation scenarios, with human-verified annotations for improved accuracy.

A new benchmark called CFG-Bench has been introduced to evaluate fine-grained action intelligence in Multimodal Large Language Models (MLLMs) for embodied agents. This benchmark includes 1,368 curated videos and 19,562 question-answer pairs, focusing on cognitive abilities such as physical interaction and evaluative judgment.

The introduction of AdaVideoRAG marks a significant advancement in the field of long video understanding by utilizing an adaptive Retrieval-Augmented Generation (RAG) framework. This innovative approach addresses the limitations of existing models, which struggle with fixed-length contexts and long-term dependencies, by dynamically selecting retrieval schemes based on query complexity.

A new framework named OrdMoE has been introduced to enhance preference alignment in Multimodal Large Language Models (MLLMs) by utilizing intrinsic signals from Mixture-of-Experts (MoE) architectures, eliminating the need for costly human-annotated preference data. This approach constructs an internal preference hierarchy based on expert selection scores, enabling the generation of responses with varying quality levels.