Multimodal Continual Learning with MLLMs from Multi-scenario Perspectives

The SPINE framework introduces a token-selective approach to test-time reinforcement learning, addressing the challenges faced by large language models (LLMs) and multimodal LLMs (MLLMs) during distribution shifts at test-time. By focusing on high-entropy tokens and applying an entropy-band regularizer, SPINE aims to enhance model performance and maintain exploration during reinforcement learning processes.

A new framework named Vision-Motion-Reference aligned Referring Multi-Object Tracking (VMRMOT) has been proposed to enhance the performance of referring multi-object tracking (RMOT) by integrating motion dynamics with visual and language references using multi-modal large language models (MLLMs). This addresses the limitations of conventional RMOT, which struggles to account for dynamic changes in object motion.

The introduction of ReEXplore marks a significant advancement in embodied exploration by utilizing a training-free framework that enhances the decision-making capabilities of multimodal large language models (MLLMs) through retrospective experience replay and hierarchical frontier selection. This approach addresses the limitations of existing MLLMs, which struggle with outdated knowledge and complex action spaces.

ReMatch has been introduced as a framework that utilizes the generative capabilities of Multimodal Large Language Models (MLLMs) for enhanced multimodal retrieval. This approach trains the embedding MLLM end-to-end, incorporating a chat-style generative matching stage that assesses relevance from diverse inputs, thereby improving the quality of multimodal embeddings.

PRISM-Bench has been introduced as a new benchmark for evaluating multimodal large language models (MLLMs) through puzzle-based visual tasks that assess both problem-solving capabilities and reasoning processes. This benchmark specifically requires models to identify errors in a step-by-step chain of thought, enhancing the evaluation of logical consistency and visual reasoning.

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.

A new algorithm named MM-Det++ has been proposed to enhance the detection of videos generated by diffusion models, addressing the growing concerns over synthetic media and information security. This algorithm integrates a Spatio-Temporal branch utilizing a Frame-Centric Vision Transformer and a Multimodal branch for improved detection capabilities.

A new benchmark called RoadBench has been introduced to evaluate the fine-grained spatial understanding and reasoning capabilities of multimodal large language models (MLLMs) in urban road scenarios, focusing on road markings as a critical element. This benchmark includes six tasks with 9,121 manually verified test cases, utilizing BEV and FPV image inputs to assess MLLMs' performance.