Multimodal Large Language Models (MLLMs) face critical privacy challenges due to the indiscriminate memorization of sensitive data. Existing unlearning methods, largely adapted from Euclidean paradigms, suffer from a geometric mismatch: they fail to disentangle specific instances from general concepts, causing catastrophic forgetting or unsafe substitution. We introduce LOTUS (Lorentz Transport for Unlearning Strategies), a framework for surgical semantic pruning within the Lorentz manifold. Leveraging hyperbolic geometry’s hierarchical nature, LOTUS employs an Inverted Entailment Cone Loss to sever the inheritance of sensitive concepts and a Lorentz Transport mechanism to align pruned features within the tangent space, ensuring compatibility with Euclidean backbones via a safety refusal prior. Experiments on MLLMU-Bench with LLaVA and Qwen show that LOTUS significantly outperforms baselines, effectively erasing targeted visual data while preserving general utility.

Through preliminary studies, we observed that some figurative language detection tasks show consistent biases: a conservative preference for precision over recall and higher agreement within a model’s own judgments than between different models. To address these biases, we first introduce the Inter-Intra Agreement Ratio (IIAR), a novel metric that systematically quantifies LLM generalizability by comparing intra-and inter-model agreement patterns. Then, we propose Consensus-Based LLM Prediction Aggregation (CLPA), which leverages the collective strengths of multiple LLMs to mitigate individual model biases and achieve balanced precision and recall—without fine-tuning or compromising generality. Finally, we present a human–AI co-annotation framework that uses minimal expert input to efficiently determine optimal consensus thresholds, significantly reducing annotation effort. Experiments on the SemEval-2017 pun detection datasets demonstrate that our approach achieves a new state-of-the-art F1 score of 0.963 on heterographic pun detection through multi-LLM consensus aggregation; furthermore, our co-annotation framework requires only 10

Multimodal Recommendation (MMR) systems are crucial for modern platforms but are often hampered by inherent noise and uncertainty in modal features, such as blurry images, diverse visual appearances, or ambiguous text. Existing methods often overlook this modality-specific uncertainty, leading to ineffective feature fusion. Furthermore, they fail to leverage rich similarity patterns among users and items to refine representations and their corresponding uncertainty estimates. To address these challenges, we propose a novel framework, Similarity Propagation-enhanced Uncertainty for Multimodal Recommendation (SPUMR). SPUMR explicitly models and mitigates uncertainty by first constructing the Modality Similarity Graph and the Collaborative Similarity Graph to refine representations from both content and behavioral perspectives. The Uncertainty-aware Preference Aggregation module then adaptively fuses the refined multimodal features, assigning greater weight to more reliable modalities. Extensive experiments on three benchmark datasets demonstrate that SPUMR achieves significant improvements over existing leading methods.

The spread of hate speech is becoming increasingly rampant on social media, causing serious harm to both society and individuals. Most existing studies focus on detection techniques or explanation tasks, and there are few studies on unified detection and explanation models. In addition, existing large language models perform poorly on implicit hate speech. Therefore, we propose a unified detection and explanation model: Dethate-LLM. Through the LoRA fine-tuning technology, the LLaMA3-8B models were fine-tuned for hate detection, explanation, and minority target group detection tasks. Subsequently, these fine-tuned expert models are integrated with the original LLaMA3-8B into a comprehensive model, and a routing mechanism is designed to enable the model to automatically select and comprehensively utilize the most suitable expert modules. Experimental results demonstrate that Dethate-LLM outperforms state-of-the-art models in detecting implicit hate speech and excels in generating accurate explanations. On the IHC dataset, compared with the ToXCL baseline, Dethate-LLM increased the F1 score of the detection task by 0.05 points and the BERTScore of the interpretation task by 2.31 points. On the SBIC dataset, Dethate-LLM increased the F1 score of the detection task by 0.96 points, the ROUGE-L score of the interpretation task by 2.48 points, and the BERTScore by 0.26 points. These results mark a substantial advancement in the capabilities of large language models in this field.

Multilingual named entity recognition (NER) is especially challenging in low-resource and typologically diverse languages, where translation drift, morphological variation, and noisy alignments degrade performance. Existing encoder-based methods often rely on dense attention or uniform alignment, which tends to propagate irrelevant signals across languages. We present SEGA, a lightweight and typology-aware framework that incorporates sparse guided attention to select auxiliary signals, alongside a weighted fusion layer that balances representations between cross-lingual and monolingual contexts. Unlike prior approaches, SEGA requires no parallel corpora and supports fully monolingual inference. We evaluate SEGA on six multilingual NER benchmarks spanning over 60 languages, including CoNLL, WikiANN, MasakhaNER 2.0, XTREME-40, WikiNEuRal, and MultiNERD. SEGA achieves new state-of-the-art results on five datasets, with absolute gains of up to +24.2 F1 over strong encoder baselines, and outperforming prompt-based large language models by up to +18.9 F1 in low-resource scenarios. Efficiency analyses show that SEGA adds only similar to 30M parameters beyond a standard dual encoder, making it lightweight and deployable on a single GPU. Comprehensive ablation, visualization, and error analyses confirm that SEGA is robust to alignment noise, morphological complexity, and boundary ambiguity, offering a practical and scalable solution for real-world multilingual NER.

Enhancing the moral alignment of Large Language Models (LLMs) is a critical challenge in AI safety. Current alignment techniques often act as superficial guardrails, leaving the intrinsic moral representations of LLMs largely untouched. In this paper, we bridge this gap by leveraging Moral Foundations Theory (MFT) to map and manipulate the fine-grained moral landscape of LLMs. Through cross-lingual linear probing, we validate the shared nature of moral representations in middle layers and uncover a shared yet different moral subspace between English and Chinese. Building upon this, we extract steerable Moral Vectors and successfully validate their efficacy at both internal and behavioral levels. Leveraging the high generalizability of morality, we propose Adaptive Moral Fusion (AMF), a dynamic inference-time intervention that synergizes probe detection with vector injection to tackle the safety-helpfulness trade-off. Empirical results confirm that our approach acts as a targeted intrinsic defense, effectively reducing incorrect refusals on benign queries while minimizing jailbreak success rates compared to standard baselines.

Fine-tuning large language models (LLMs) with instruction sets has become an effective method to improve the performance of LLMs. However, current metaphor datasets suffer from issues such as in-consistent annotation methods, insufficient data, limited task diversity, and a narrow genre scope, making it difficult to form a unified instruction set. Additionally, existing fine-tuning strategies have not considered the logical sequence and difficulty differences among metaphor tasks. To address these issues, we have released the “Metaphor Understanding and Generation Instruction Fine-Tuning Dataset” (MetaIFD). This dataset contains 113,384 high-quality entries from over ten genres and supports five metaphor tasks and sentiment analysis tasks. Furthermore, MetaIFD introduces two key annotations: the metaphor identification analysis process and sentiment labels. In this paper, we also propose a three-stage fine-tuning strategy, following the principles of “explicit sentiment tasks preceding implicit sentiment tasks” and “understanding tasks preceding generation tasks,” along with a difficulty-based training approach that considers the logical sequence and difficulty differences between tasks. Experimental results show that the fine-tuned model significantly improves performance across all tasks. The dataset and instruction set are publicly available at: https://github.com/DUTIR-YSQ/MetaIFD .

Session-based recommendation aims to predict intents of anonymous users based on limited behaviors. With the ability in alleviating data sparsity, contrastive learning is prevailing in the task. However, we spot that existing contrastive learning based methods still suffer from three obstacles: (1) they overlook item-level sparsity and primarily focus on session-level sparsity; (2) they typically augment sessions using item IDs like crop, mask and reorder, failing to ensure the semantic consistency of augmented views; (3) they treat all positive-negative signals equally, without considering their varying utility. To this end, we propose a novel multi-modal adaptive contrastive learning framework called MACL for session-based recommendation. In MACL, a multi-modal augmentation is devised to generate semantically consistent views at both item and session levels by leveraging item multi-modal features. Besides, we present an adaptive contrastive loss that distinguishes varying contributions of positive-negative signals to improve self-supervised learning. Extensive experiments on three real-world datasets demonstrate the superiority of MACL over state-of-the-art methods.

We introduce VisualQuest, a novel dataset designed to rigorously evaluate multimodal large language models (MLLMs) on abstract visual reasoning tasks that require integration of symbolic, cultural, and linguistic knowledge. Unlike existing benchmarks focused on direct image captioning or classification of realistic images, VisualQuest comprises 3,551 non-photographic, stylized images spanning four categories—Public Figures, Popular Culture, Linguistic Expressions, and Literary Works—each paired with targeted questions to probe complex reasoning. We benchmark ten state-of-the-art MLLMs, revealing that only Gemini-2.5-flash and GPT-4o achieve strong overall performance, while 3.7 https://github.com/xkt88/VISUALQUEST .

Large language models (LLMs) are increasingly used for hate speech moderation, often within human–AI workflows in which reviewers provide feedback before a final decision. Such feedback introduces two manipulation directions: whitewashing hateful content as normal and smearing normal content as hateful. This study examines the susceptibility of initially correct model judgments to annotator-style rebuttals and analyzes whether attack effectiveness differs across manipulation directions. We introduce a rejudge protocol that extends direct contradiction with decision-boundary perturbations and adversarial rationales. Experiments with multiple LLMs on two hate speech datasets show that annotator-style rebuttals substantially degrade moderation performance, with stronger effects in multi-turn settings. The results further reveal stable, model-specific asymmetries between whitewashing and smearing across attack configurations, indicating distinct directional vulnerability patterns. Explicit reasoning prompts and defensive instructions reduce these effects but do not eliminate them. These findings highlight the need for direction-aware safeguards and dedicated feedback-robustness evaluation in human–AI moderation workflows.

Text-to-Image Models (T2IMs) excel at visualizing explicit instructions but struggle to translate high-level concepts, metaphors, and cultural entities into coherent visuals. Current evaluations focus on fidelity for concrete captions, leaving abstract visual reasoning unexplored. To address this gap, we introduce a multi-domain benchmark of 1,800 abstract concepts across six culturally diverse domains, rendered in Cartoon and Minimalist styles, paired with a two-task expert evaluation protocol covering graded semantic alignment and open-ended recognition. We further propose Iterative Self-Refinement (ISR), a closed-loop framework coordinating an LLM, T2IM, and MLLM that unifies evaluation and refinement in a single reasoning step, eliminating the diagnose-then-edit cascade of prior pipelines. Experiments across 10 LLMs and 10 MLLMs, validated by 18 domain experts (Fleiss' kappa = 0.69), show that ISR achieves substantial gains over T2IM-only baselines (+14.7% Cartoon, +12.9% Minimalist), identify MLLM abstract reasoning as the primary performance bottleneck, and reveal complementary strengths: MLLMs excel on visually-grounded domains while humans dominate imagination-based categories.

MER2026 marks the fourth edition of the MER series of challenges. The MER series provides valuable data resources to the research community and offers tasks centered on recent research trends, establishing itself as one of the largest challenges in the field. Throughout its history, the focus of MER has shifted from discriminative emotion recognition to generative emotion understanding. Specifically, MER2023 concentrated on discriminative emotion recognition, restricting the emotion recognition scope to fixed basic labels. In MER2024 and MER2025, we transitioned to generative emotion understanding and introduced two new tasks: fine-grained emotion recognition and descriptive emotion analysis, aiming to leverage the extensive vocabulary and multimodal understanding capabilities of Multimodal Large Language Models (MLLMs) to facilitate fine-grained and explainable emotion recognition. Building on this trajectory, MER2026 continues to follow these research trends and contains four tracks: MER-Cross shifts the focus from individual to dyadic interaction scenarios; MER-FG centers on fine-grained emotion recognition; MER-Prefer aims to predict human preferences regarding different emotion descriptions; MER-PS focuses on emotion recognition based on physiological signals. More details regarding the dataset and baselines are available at https://zeroqiaoba.github.io/MER-Challenge/.

Large language models for subjectivity analysis are typically trained with aggregated labels, which compress variations in human judgment into a single supervision signal. This paradigm overlooks the intrinsic uncertainty of low-agreement samples and often induces overconfident predictions, undermining reliability and generalization in complex subjective settings. In this work, we advocate uncertainty-aware subjectivity analysis, where models are expected to make predictions while expressing uncertainty that reflects human disagreement. To operationalize this perspective, we propose a two-phase Disagreement Perception and Uncertainty Alignment (DPUA) framework. Specifically, DPUA jointly models label prediction, rationale generation, and uncertainty expression under an uncertainty-aware setting. In the disagreement perception phase, adaptive decoupled learning enhances the model's sensitivity to disagreement-related cues while preserving task performance. In the uncertainty alignment phase, GRPO-based reward optimization further improves uncertainty-aware reasoning and aligns the model's confidence expression with the human disagreement distribution. Experiments on three subjectivity analysis tasks show that DPUA preserves task performance while better aligning model uncertainty with human disagreement, mitigating overconfidence on boundary samples, and improving out-of-distribution generalization.

Civil judicial cases are highly complicated, posing significant challenges for Large Language Models (LLMs) for Legal Judgment Prediction (LJP). While judges manage this complexity through the dispute focus—a mechanism distilling cases into core issues—existing research largely overlooks this tool in favor of generic reasoning frameworks that lack authentic judicial logic. To bridge this gap, we first introduce FocalLaw , the first dataset aligning full-process Chinese civil judicial data through the dispute focus, comprising 1,000 high-quality cases across six causes of action. Building on this dataset, we examine LLMs’ capability to utilize the dispute focus and uncover a counter-intuitive phenomenon: LLMs fail to leverage the dispute focus even with CoT and SFT, which we identify as the "Clerk Trap".To solve the problem, we propose FocalJudge , a novel framework that leverages the dispute focus to guide LLMs through a structured, judge-like cognitive workflow. Experimental results demonstrate the effectiveness of FocalJudge and offer valuable insights into the interpretability and reliability of LLMs in the legal domain.

With the rise of social media platforms, people increasingly resort to combining images and texts to convey complex emotions and thoughts, where the use of metaphors is particularly common. These multi-modal expressions often contain profound, indirect messages, requiring a more deep-going level of understanding from the receiver. To address this challenge, we develop an advanced multi-modal fusion model focused on the identification and understanding of metaphors. Distinct from traditional uni-modal analyses, our model is dedicated to exploring and analyzing the interplay between images and texts, especially in the source and target domains of metaphors. Through this approach, we can more accurately capture and compare features across modalities, unveiling hidden meanings. Additionally, the model employs a strategy of hierarchical contrastive learning, enhancing the depth of understanding and analysis of multi-modal data. Through extensive experiments, it has been demonstrated that our model is effective in metaphor recognition and understanding in multi-modal contexts, showing significant improvements in the field of metaphor interpretation and providing a more comprehensive and accurate analysis of multi-modal metaphors.

Figurative language detection remains a challenging task for large language models (LLMs) due to its inherent subjectivity and cultural dependency. Despite advancements in natural language processing (NLP), current LLMs still struggle with handling metaphors, idioms, hyperboles, and puns, exhibiting inconsistent predictions across different models and tasks. This study systematically analyzes these model-level behaviors, focusing on their biases and inconsistencies in figurative language detection. Our findings reveal that LLMs demonstrate a conservative bias, consistently favoring precision over recall, especially when dealing with ambiguous or subjective language. Additionally, we observe stronger intra-model agreement compared to inter-model agreement, uncovering systematic bias patterns within and across model families. To mitigate these biases and enhance model performance, we propose the Consensus-Based LLM Prediction Aggregation (CLPA) framework. Our method leverages stable agreement patterns across multiple models to improve the accuracy of figurative language predictions, achieving state-of-the-art results in heterographic pun detection without fine-tuning. This research highlights the potential for consensus-driven approaches in enhancing LLM generalization and reducing biases in subjective linguistic tasks.

Research on harmful meme detection has garnered significant attention, resulting in the development of numerous datasets and methods. However, progress in detecting Chinese harmful memes lags considerably, primarily due to two challenges: first, accurately assessing a meme's harmfulness depends heavily on understanding deep cultural context; second, many memes are semantically ambiguous, making harmfulness highly subjective. To address these issues, we focus on the interpretable detection of Chinese harmful memes by constructing the first Chinese harmful meme explanation dataset, Ex-ToxiCN-MM. This dataset offers opposing interpretations, categorized as "harmful" and "non-harmful", for each meme, aiming to rigorously evaluate a model's ability to discern and comprehend ambiguous, culturally grounded content. We built a specialized knowledge base of Chinese cultural concepts and offensive vocabulary to supply models with essential prior knowledge (C-HarmKB). To address the ambiguity and lack of background knowledge in meme attribution, we have developed a comprehensive attribution analysis framework, RIKE, which includes an Attribution Knowledge Enhancement module (AKE) and a Relative Intent Reasoning module (RIR). Extensive quantitative and qualitative experiments demonstrate that our method outperforms mainstream baseline models across multiple metrics in the task of attributing harmful memes in Chinese. The code, Ex-ToxiCN-MM dataset, and Chinese Harmful Semantic Knowledge Base (C-HarmKB) involved in this study have been open-sourced at https://github.com/wimiw123/Ex-ToxiCN-MM