Multimodal Large Language Model (MLLM) Personalization is a critical research problem that facilitates personalized dialogues with MLLMs targeting specific entities (known as personalized concepts). However, existing methods and benchmarks focus on the simple, context-agnostic visual identification and textual replacement of the personalized concept (e.g., "A yellow puppy" -> "Your puppy Mochi"), overlooking the ability to support long-context conversations. An ideal personalized MLLM assistant is capable of engaging in long-context dialogues with humans and continually improving its experience quality by learning from past dialogue histories. To bridge this gap, we propose LCMP, the first Long-Context MLLM Personalization evaluation benchmark. LCMP assesses the capability of MLLMs in perceiving variations of personalized concepts and generating contextually appropriate personalized responses that reflect these variations. As a strong baseline for LCMP, we introduce a novel training-free and state-aware framework TAME. TAME endows MLLMs with double memories to manage the temporal and persistent variations of each personalized concept in a differentiated manner. In addition, TAME incorporates a new training-free Retrieve-then-Align Augmented Generation (RA2G) paradigm. RA2G introduces an alignment step to extract the contextually fitted information from the multi-memory retrieved knowledge to the current questions, enabling better interactions for complex real-world user queries. Experiments on LCMP demonstrate that TAME achieves the best performance, showcasing remarkable and evolving interaction experiences in long-context scenarios.

The proliferation of harmful memes on online media poses significant risks to public health and stability. Existing detection methods heavily rely on large-scale labeled data for training, which necessitates substantial manual annotation efforts and limits their adaptability to the continually evolving nature of harmful content. To address these challenges, we present ALARM, the first lAbeL-free hARmful Meme detection framework powered by Large Multimodal Model (LMM) agent self-improvement. The core innovation of ALARM lies in exploiting the expressive information from "shallow" memes to iteratively enhance its ability to tackle more complex and subtle ones. ALARM consists of a novel Confidence-based Explicit Meme Identification mechanism that isolates the explicit memes from the original dataset and assigns them pseudo-labels. Besides, a new Pairwise Learning Guided Agent Self-Improvement paradigm is introduced, where the explicit memes are reorganized into contrastive pairs (positive vs. negative) to refine a learner LMM agent. This agent autonomously derives high-level detection cues from these pairs, which in turn empower the agent itself to handle complex and challenging memes effectively. Experiments on three diverse datasets demonstrate the superior performance and strong adaptability of ALARM to newly evolved memes. Notably, our method even outperforms label-driven methods. These results highlight the potential of label-free frameworks as a scalable and promising solution for adapting to novel forms and topics of harmful memes in dynamic online environments.

Traditional multimodal learning approaches often assume that all modalities are available during both the training and inference phases. However, this assumption is often impractical in real-world scenarios due to challenges such as sensor failures, data corruption, or privacy concerns. While recent efforts focus on enhancing the robustness of pre-trained Multimodal Transformers (MTs) under missing modality conditions, mainstream work in this field often overlook reconstructing the missing modalities and rely on static, sample-agnostic prompt-tuning techniques, undermining their efficacy in severe modality missing scenarios. To address these limitations, we propose REDEEM, a novel REtrieval-guiDEd conditional gEnerative fraMework that largely alleviates the modality missing problems on pre-trained MTs. REDEEM consists of a new adaptive retrieval mechanism to identify relevant instances for both modality-complete and -incomplete samples. It then conditions on the remaining modalities and utilizes the retrieved data as experts to effectively recover the missing ones in modality-incomplete instances through a within-modal reconstruction manner. Finally, REDEEM generates sample-aware inter-modal prompts from the retrieved instances to guide MTs in tackling severe modality missing challenges. Comprehensive experiments on three diverse multimodal classification benchmarks demonstrate that REDEEM significantly outperforms competitive baselines.

The rapid spread of rumor content on online micro-video platforms poses significant threats to public health and safety. However, existing Micro-Video Rumor Detection (MVRD) methods are generally black-box, which lacks transparency and makes it difficult to understand the reasoning behind classification decisions. In this work, we introduce ExMRD, a novel Explainable Micro-video Rumor Detection framework designed to generate detailed and coherent explanations for enhancing MVRD. Inspired by the powerful reasoning capacity of Chain-of-Thought (CoT), we introduce a novel inference mechanism called R3CoT-- consisting of Refining, Retrieving, and Reasoning on MVRD. This mechanism enables Multimodal Large Language Models (MLLMs) to reorganize the original video content, retrieve domain knowledge related to rumors, and generate explainable conclusions regarding whether the micro-video contains rumor information. Instead of directly fine-tuning MLLMs for MVRD, which is computationally expensive, we propose a Small Language Reviewer (SLReviewer), which distills the outputs of R3CoT guided MLLMs to ensure efficient and reliable predictions. Extensive experiments on three real-world benchmarks demonstrate that ExMRD significantly outperforms competitive baselines while providing high-quality rationales.

Short Video Hate Detection (SVHD) is increasingly vital as hateful content -such as racial and gender-based discriminationspreads rapidly across platforms like TikTok, YouTube Shorts, and Instagram Reels.Existing approaches face significant challenges: hate expressions continuously evolve, hateful signals are dispersed across multiple modalities (audio, text, and vision), and the contribution of each modality varies across different hate content.To address these issues, we introduce MoRE (Mixture of Retrievalaugmented multimodal Experts), a novel framework designed to enhance SVHD.MoRE employs specialized multimodal experts for each modality, leveraging their unique strengths to identify hateful content effectively.To ensure model's adaptability to rapidly evolving hate content, MoRE leverages contextual knowledge extracted from relevant instances retrieved by a powerful joint multimodal video retriever for each target short video.Moreover, a dynamic sample-sensitive integration network adaptively adjusts the importance of each modality on a per-sample basis, optimizing the detection process by prioritizing the most informative modalities for each instance.Our MoRE adopts an end-to-end training strategy that jointly optimizes both expert networks and the overall framework, resulting in nearly a twofold improvement in training efficiency, which in turn enhances its applicability to real-world scenarios.Extensive experiments on three benchmarks demonstrate that MoRE surpasses state-of-the-art baselines, achieving an average improvement of 6.91% in macro-F1 score across all datasets.

Temporal heterogeneous networks (THNs) investigate the structural interactions and their evolution over time in graphs with multiple types of nodes or edges. Existing THNs describe evolving networks as a sequence of graph snapshots and adopt mechanisms from static heterogeneous networks to capture the spatial-temporal correlation. However, these works are confined to the discrete-time setting and the implementation of stacked mechanisms often introduces a high level of complexity, both conceptually and computationally. Here, we conduct comprehensive examinations and propose STHN, a simplifying THN for continuous-time link prediction. Concretely, to integrate continuous dynamics, we maintain a historical interaction memory for each node. A link encoder that incorporates two components - type encoding and relative time encoding - is introduced to encapsulate implicit heterogeneous characteristics of interaction and extract the most informative temporal information. We further propose to use a patching technique that assists with Transformer feature extractor to support the interaction sequence with long histories. Extensive experiments on three real-world datasets empirically demonstrate that STHN outperforms state-of-the-art methods with competitive task accuracy and predictive efficiency on both transductive and inductive settings.