AAAI Conference on Artificial Intelligence · 2140 papers
MetaEval: Measuring the Discrimination of Benchmarks for Efficient LLM Evaluation
Zhuo Wang (East China Normal University), Zhenxiao Cheng (East China Normal University)
CodeComputational EfficiencyLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the MetaEval framework to measure the discriminability of individual questions in evaluation benchmarks and achieve efficient evaluation based on this.
MetaGameBO: Hierarchical Game-Theoretic Driven Robust Meta-Learning for Bayesian Optimization
Hui Li (Beijing Jiaotong University), Jian Yu (Beijing Jiaotong University)
CodeOptimizationMeta LearningBenchmark
🎯 What it does: Propose a hierarchical game-theoretic framework named MetaGameBO to explicitly optimize robustness for extremely difficult tasks in meta-learning with Bayesian optimization, and achieve efficient sample utilization through multi-level sample selection.
MetaGDPO: Alleviating Catastrophic Forgetting with Metacognitive Knowledge Through Group Direct Preference Optimization
Lanxue Zhang (Chinese Academy of Sciences), Yanan Cao (JIUTIAN Research)
CodeOptimizationComputational EfficiencyMeta LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Construct a 5K QA dataset incorporating metacognitive knowledge and propose Group Direct Preference Optimization (GDPO) to compress and fine-tune small LLMs, mitigating catastrophic forgetting.
METP: Multi-Granularity Integration of External Covariates for Temporal Point Processes
Boyang Li (Peking University), Xi Zhang (Peking University)
CodeTransformerMixture of ExpertsTime SeriesFinance Related
🎯 What it does: Propose the METP framework, which integrates the lag effects of multi-granularity external covariates into the intensity function of temporal point processes to enhance event time prediction.
🎯 What it does: Propose MFINet, a real-time LiDAR semantic segmentation network based on three-branch multi-perspective fusion and 2D-3D interaction enhancement.
🎯 What it does: Propose a multifunctional network called MFmamba, which can achieve image super-resolution, spectral recovery, and joint recovery of both tasks when only a single panchromatic image is input.
MGT-Prism: Enhancing Domain Generalization for Machine-Generated Text Detection via Spectral Alignment
Shengchao Liu (Xi'an Jiaotong University), Shuai Xiao (Xi'an Jiaotong University)
CodeAnomaly DetectionTransformerText
🎯 What it does: Propose a machine-generated text detection method called MGT-Prism based on frequency domain analysis, leveraging low-frequency filtering and spectral alignment to enhance cross-domain generalization performance.
MHA2MLA-VLM: Enabling DeepSeek’s Economical Multi-Head Latent Attention Across Vision-Language Models
Xiaoran Fan (Fudan University), Tao Gui (Fudan University)
CodeCompressionVision Language ModelMultimodality
🎯 What it does: This paper proposes a parameter-efficient, modality-aware MHA2MLA-VLM framework, which migrates existing MHA/GQA-based vision-language models to the Multi-Head Latent Attention (MLA) architecture, achieving substantial KV cache compression while maintaining the original performance.
MicroEvoEval: A Systematic Evaluation Framework for Image-Based Microstructure Evolution Prediction
Qinyi Zhang (Sichuan University), Hao Wang (Sichuan University)
CodeRecurrent Neural NetworkTransformerImageBenchmarkPhysics Related
🎯 What it does: This paper proposes and implements MicroEvoEval, a systematic evaluation framework for image-based microstructure evolution prediction, assessing the short-term and long-term prediction performance of 14 deep learning models across four representative PDE tasks.
Mind the Gap: The Divergence Between Human and LLM-Generated Tasks
Yi-Long Lu (State Key Laboratory of General Artificial Intelligence), Wei Wang (State Key Laboratory of General Artificial Intelligence)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Compare differences between humans and large language models (GPT-4o) in generating autonomous tasks, revealing that human tasks are driven by personal values and physical experiences, while LLM-generated tasks lack social and physical aspects, tending toward abstraction.
Mind the Third Eye! Benchmarking Privacy Awareness in MLLM-powered Smartphone Agents
Zhixin Lin, Dongliang Xu (Shandong University)
CodeSafty and PrivacyLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Constructed a large-scale privacy-aware benchmark named SAPA-Bench to evaluate the privacy identification and response capabilities of mobile agents.
CodeGenerationData SynthesisDomain AdaptationDiffusion modelVideoBiomedical Data
🎯 What it does: This paper proposes the MindCross framework, achieving cross-subject reconstruction of brain signal to video mapping under a single model, and can quickly adapt with only a small amount of new subject data.
Minimum-Length Conformal Prediction Sets for Ordinal Classification
Zijian Zhang (Washington State University), Yan Yan (Washington State University)
CodeClassificationImageTime Series
🎯 What it does: Proposed a model-free shortest covering confidence prediction method (min-CPS) and its length-regularized variant (min-RCPS) for distribution-agnostic uncertainty quantification in ordinal classification.
Zeqing Wang (National University of Singapore), Xinchao Wang (National University of Singapore)
CodeGenerationTransformerDiffusion modelVideo
🎯 What it does: Propose a distributed inference strategy named DualParal, combining video sequence parallelism with model layer parallelism, and achieving efficient generation of long videos through block-wise denoising.
MIRA: Evaluating Multimodal AI on Complex Clinical Reasoning in Interventional Radiology
Jingxiong Li (Nanjing University of Science and Technology), Liang Xiao (Nanjing University of Science and Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityBiomedical DataMagnetic Resonance ImagingComputed TomographyUltrasoundBenchmark
🎯 What it does: Built and released MIRA—a large-scale multimodal question-answering benchmark designed for interventional radiology, containing 184,479 medical images and approximately 1.2 million expert-generated question-answer pairs, covering open-ended, closed-ended, single-choice, and multiple-choice questions, along with expert-verified reasoning explanations.
CodeExplainability and InterpretabilityGraph Neural NetworkTransformerLarge Language ModelTextGraphBiomedical DataRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the MIRAGE framework, achieving scalable inference during testing in medical QA tasks through parallel multi-chain reasoning and structured knowledge graph retrieval;
MIRAGE: Towards AI-Generated Image Detection in the Wild
OuCheng Huang (Taobao & Tmall Group of Alibaba), Bo Zheng (Taobao & Tmall Group of Alibaba)
CodeAnomaly DetectionLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodalityBenchmark
🎯 What it does: This paper addresses the task of detecting AI-generated images in real-world scenarios by constructing MIRAGE, a wild detection benchmark containing human-planned and multi-model synthesized images, and proposes MIRAGE-R1, a vision-language model capable of adaptive reasoning.
MISF: MLLM Guided Iterative Sample Filtering for Data Fault Detection
Guoying Chen (Beijing Institute of Computer Technology and Application), Kunlong Wang (Beijing Institute of Computer Technology and Application)
CodeAnomaly DetectionData-Centric LearningConvolutional Neural NetworkLarge Language ModelDiffusion modelImage
🎯 What it does: Proposes the MISF (MLLM-Guided Iterative Sample Filtering) framework, which utilizes a multimodal large language model to generate synthetic images and a small number of real clean samples to initialize the detector. Then, it iteratively filters clean samples through Gini uncertainty and prediction consistency to achieve detection of label noise and backdoor attacks in image data.
Mitigating Endogenous Confirmation Bias in Noisy Label Learning for Vision-Language Models
Feiyang Ning (Harbin Institute of Technology), Xinyang Chen (Harbin Institute of Technology)
CodeData-Centric LearningPrompt EngineeringVision Language ModelMultimodality
🎯 What it does: To address the self-verification bias caused by pre-trained knowledge in vision-language models during noisy label learning, this paper proposes a multi-stage debiasing framework DKAF, which selects and corrects noisy samples through cross-modal pseudo labels, bimodal consistency, and debiased cross-entropy.
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataComputed Tomography
🎯 What it does: By constructing the dual-stream entity alignment network DEAR, the problem of entity misreporting in 3D CT report generation is addressed through fine-grained alignment of organ and lesion entities, significantly reducing hallucinations.
Mitigating Hallucinations in Large Language Models via Causal Reasoning
Yuangang Li (University of Southern California), Yue Zhao (University of Southern California)
CodeExplainability and InterpretabilityGraph Neural NetworkLarge Language ModelSupervised Fine-TuningTextGraphChain-of-Thought
🎯 What it does: This paper enhances causal inference capabilities and reduces logical hallucinations by first enabling LLMs to construct causal directed acyclic graphs (DAGs) and then performing reasoning on the graphs;
🎯 What it does: Proposed and studied the Edge-Preserving Training method (MPT) to address negative flips during model updates, which maintains the old class margins by adding bias in softmax and achieves balanced learning of new and old classes through dual-source focal distillation.
MMBERT: Scaled Mixture-of-Experts Multimodal BERT for Robust Chinese Hate Speech Detection Under Cloaking Perturbations
Qiyao Xue (University of Pittsburgh), Wei Gao (University of Pittsburgh)
CodeClassificationTransformerMixture of ExpertsVision Language ModelImageTextMultimodalityAudio
🎯 What it does: Proposes MMBERT, a BERT framework integrating text, speech, and visual modalities, utilizing Mixture-of-Experts (MoE) dynamic routing and achieving robust Chinese hate speech detection through a three-stage progressive training approach.
MME-SCI: A Comprehensive and Challenging Science Benchmark for Multimodal Large Language Models
Jiacheng Ruan (Shanghai Jiao Tong University), Yangyang Kang (Zhejiang University)
CodeLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Created the MME-SCI multimodal scientific evaluation benchmark, covering four disciplines (mathematics, physics, chemistry, biology), supporting five languages and three input modes, and providing 1,019 high-quality question-answer samples with 63 fine-grained knowledge point annotations.
Tao Zhang (Chinese Academy of Sciences), Weiming Hu (Tencent Inc)
CodeRetrievalReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose the MMhops dataset and the MMhops-R1 framework for evaluating and enhancing multimodal multi-hop reasoning capabilities.
MMMamba: A Versatile Cross-Modal in Context Fusion Framework for Pan-Sharpening and Zero-Shot Image Enhancement
Yingying Wang (Xiamen University), Haoxuan Che (Xiamen University)
CodeRestorationSuper ResolutionImage
🎯 What it does: Propose a cross-modal context fusion framework called MMMamba based on Mamba for high-resolution multispectral image fusion and zero-shot image enhancement.
🎯 What it does: Proposed a material-based physical backdoor attack (MOBA), which implants realizable triggers into LiDAR point cloud training data, forcing 3D object detection models to produce incorrect predictions when the triggers appear.
Mobile-Agent-RAG: Driving Smart Multi-Agent Coordination with Contextual Knowledge Empowerment for Long-Horizon Mobile Automation
Yuxiang Zhou (Sun Yat-sen University), Guanbin Li (OPPO Inc.)
CodeAutonomous DrivingRobotic IntelligenceTransformerLarge Language ModelAgentic AIVision-Language-Action ModelImageTextSequentialBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes Mobile-Agent-RAG, a hierarchical multi-agent mobile automation framework that employs dual-layer retrieval-enhanced (Manager-RAG and Operator-RAG) modules to address strategic hallucinations in high-level planning and precise execution errors in low-level UI operations separately.
MoCHA: Advanced Vision-Language Reasoning with MoE Connector and Hierarchical Group Attention
Yuqi Pang (Chinese Academy of Sciences), Chen He (Chinese Academy of Sciences)
CodeTransformerMixture of ExpertsVision Language ModelMultimodalityBenchmark
🎯 What it does: MoCHA achieves efficient multimodal reasoning by integrating four visual encoders (CLIP, SigLIP, DINOv2, and ConvNeXt) and introducing a sparse expert connector (MoEC) and hierarchical group attention (HGA).
🎯 What it does: This paper proposes the OSSDet framework for object detection tasks in multispectral aerial images and first constructs a large-scale multispectral object detection dataset called MODA.
🎯 What it does: For unsupervised visible-infrared person re-identification, a modality-aware Jaccard distance and split contrastive learning framework is proposed to enhance cross-modal association and representation learning.
🎯 What it does: This paper proposes the MBCD framework, which addresses the problem of over-convergence of dominant modalities caused by weight averaging in multi-modal domain generalization through three modules: adaptive modal dropout, gradient consistency constraints, and cooperative distillation, thereby achieving a flatter loss landscape and more balanced cross-modal fusion.
Model Counting for Dependency Quantified Boolean Formulas
Long-Hin Fung (National Taiwan University), Tony Tan (University of Liverpool)
CodeBenchmark
🎯 What it does: Studied the model counting problem for dependency quantified Boolean formulas (DQBF), proving that even #2-DQBF with only two existential quantifiers remains #EXP-complete, and implemented a specialized counter called sharp2DQR based on BDD.
Model Editing as a Double-Edged Sword: Steering Agent Behavior Toward Beneficence or Harm
Baixiang Huang (Emory University), Kai Shu (Cisco Research)
CodeTransformerLarge Language ModelTextBenchmark
🎯 What it does: Propose to treat the ethical behavior regulation of LLM base agents as a model editing task, and construct a three-tier BEHAVIORBENCH benchmark to systematically evaluate the effectiveness of behavior editing.
Model Whisper: Steering Vectors Unlock Large Language Models’ Potential in Test-Time
Xinyue Kang (Tsinghua University), Li Chen (Tsinghua University)
CodeOptimizationLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose Test-Time Steering Vectors (TTSV), which achieves test-time adaptation and enhances inference performance by adding learnable vectors before input embeddings without modifying LLM parameters.
Modeling Item-Level Dynamic Variability with Residual Diffusion for Bundle Recommendation
Dong Zhang (Wuhan University of Technology), Jimmy Huang (York University)
CodeRecommendation SystemDiffusion model
🎯 What it does: Proposed a model-agnostic residual diffusion framework RDiffBR, helping bundle recommendation models adapt to item-level dynamic changes.
🎯 What it does: The study utilizes Vision Transformer autoencoders combined with low-rank adapters (LoRA) to achieve rapid context learning in early visual layers, and analyzes their impact on representation geometry and attention mechanisms.
🎯 What it does: Developed a model called VNOIP based on variational neural ODEs for predicting the future popularity of information in social networks, combining bidirectional jump ODEs to capture sequential information and model macro trends.
Modeling Uncertainty Trends for Timely Retrieval in Dynamic RAG
Bo Li (Hebei University of Technology), Wei Ye (Hebei University of Technology)
CodeRetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Propose a training-free, entropy-trend-based dynamic retrieval trigger mechanism (ETC) to accurately determine when to retrieve external knowledge during the generation process.
Modulation-Based Backdoors: Leveraging Amplitude and Frequency Patterns to Attack Speaker Recognition
Hanbo Cai (Hohai University), Ying Luo (Sun Yat-sen University)
CodeRecognitionAdversarial AttackAudio
🎯 What it does: This paper proposes an acoustic backdoor attack based on frequency modulation (FSMA) and amplitude modulation (ASMA), which deceives speaker recognition models by embedding hidden frequency or amplitude variations into speech without altering semantics.
MoE^2: A Mixture-of-Mixtures of Experts for Ensemble-Free Domain Generalization
Ahmed Radwan (University of British Columbia), Mohamed S. Shehata (University of British Columbia)
CodeClassificationDomain AdaptationTransformerMixture of ExpertsImage
🎯 What it does: Propose the MoE² framework, which uses a single frozen ViT backbone and dynamically combines lightweight adapter experts to synthesize customized network parameters for each input, achieving domain generalization.
MoEA-Net: Modality-Incremental Expert Aggregation Network for Retinal Prognostic Prediction
Hua Wang (Beijing University of Technology), Xiaobing Yu (Academy of Medical Sciences)
CodeClassificationTransformerLarge Language ModelMixture of ExpertsVision Language ModelMultimodalityTime SeriesBiomedical Data
🎯 What it does: Designed and implemented a multimodal time-series retinal image prognosis prediction framework called MoEA-Net, capable of predicting long-term changes in retinal diseases related to Age-related Macular Degeneration (AMD) and Diabetic Macular Edema (DME).
MoETTA: Test-Time Adaptation Under Mixed Distribution Shifts with MoE-LayerNorm
Xiao Fan (Tongji University), Zhi Wang (Tsinghua University)
CodeDomain AdaptationConvolutional Neural NetworkTransformerMixture of ExpertsImageBenchmark
🎯 What it does: This paper proposes MoETTA, a LayerNorm structure based on Mixture-of-Experts (MoE), to achieve test-time adaptation (TTA) with adaptive updates under mixed distribution shifts, and introduces two more realistic benchmarks, potpourri and potpourri+, to evaluate model robustness under multi-source, mixed distribution conditions.
MOGO: Residual Quantized Hierarchical Causal Transformer for Real-Time and Infinite-Length 3D Human Motion Generation
Dongjie Fu (MogoAI), Hansung Kim (University of Southampton)
CodeGenerationTransformerLarge Language ModelTextSequentialChain-of-Thought
🎯 What it does: Proposes the MOGO framework, which realizes real-time, one-time generation from text to 3D human actions through hierarchical encoding with residual quantization (MoSA-VQ) and a single-pass forward residual quantization hierarchical causal transformer (RQHC-Transformer), supporting infinite-length generation.
MoHoBench: Assessing Honesty of Multimodal Large Language Models via Unanswerable Visual Questions
Yanxu Zhu (Beijing Jiaotong University), Xing Xie (Microsoft Research Asia)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Constructed the MoHoBench large-scale multimodal model honesty evaluation benchmark and systematically evaluated the honesty performance of 28 multimodal large language models on unanswerable visual questions.
MolSight: Optical Chemical Structure Recognition with SMILES Pretraining, Multi-Granularity Learning and Reinforcement Learning
Wenrui Zhang (Huazhong University of Science and Technology), Wenyu Liu (Huazhong University of Science and Technology)
CodeRecognitionDrug DiscoveryTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageText
🎯 What it does: Achieved optical chemical structure recognition (OCSR) through a three-stage training process (pre-training, hierarchical fine-tuning, reinforcement learning), with significant improvements in stereochemistry recognition.
🎯 What it does: Propose the MonoCLUE framework, which enhances the accuracy and robustness of 3D detection through object-level clustering and scene memory on monocular images.
More than Irrational: Modeling Belief-Biased Agents
Yifan Zhu (ELLIS Institute), Samuel Kaski (ELLIS Institute)
CodeExplainability and InterpretabilityReinforcement LearningSequential
🎯 What it does: Proposed a user model based on the computational rationality framework, utilizing a parameterizable memory decay function to explain irrational behaviors, and developed an online inference algorithm with nested particle filters to simultaneously estimate users' memory boundary parameters and internal biased beliefs from passively observed behavioral sequences;
Mosaic Pruning: A Hierarchical Framework for Generalizable Pruning of Mixture-of-Experts Models
Wentao Hu (Xi'an Jiaotong University), Jiayin Wang (Xi'an Jiaotong University)
CodeComputational EfficiencyMixture of ExpertsTextBenchmark
🎯 What it does: Propose a hierarchical 'clustering-reselection' framework called Mosaic Pruning (MoP) for structured pruning in large-scale sparse expert models (MoE), enabling cross-domain deployment after one-time pruning while maintaining functional diversity.
MosaicDoc: A Large-Scale Bilingual Benchmark for Visually Rich Document Understanding
Ketong Chen (South China University of Technology), Yang Xue (South China University of Technology)
CodeTransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmark
🎯 What it does: Propose the DocWeaver multi-agent automated generation framework and construct the first visually rich document understanding benchmark MosaicDoc for Chinese and English newspapers and magazines.
🎯 What it does: Propose MotionFlow, a training-agnostic test-time latent optimization framework that transfers motion from a source video to generate a new video under target text prompts.
🎯 What it does: Propose the MoToRec framework, which converts the multimodal recommendation task into a sparse regularized residual quantized variational autoencoder (RQ-VAE) to generate interpretable discrete semantic tokens, and achieves cold start recommendations through adaptive sparsity amplification and hierarchical multi-source graph encoding.
MovieGraph-ToM: Evaluating Long-Range Theory of Mind in Large Language Models via Implicit Social-Causal Graphs
Tingjiang Wei (East China Normal University), Liang He (East China Normal University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelTextMultimodalityGraphBenchmark
🎯 What it does: Constructed a large-scale multimodal benchmark, MovieGraph-ToM, to evaluate LLMs' long-range theory of mind (ToM) reasoning capabilities in full movies.
🎯 What it does: Designed and implemented a motion semantic contrastive learning framework (MovSemCL) to extract motion semantic features from raw GPS trajectories, employing hierarchical patch encoding and curvature-guided enhancement, ultimately generating efficient and robust trajectory embeddings.
MP-ISMoE: Mixed-Precision Interactive Side Mixture-of-Experts for Efficient Transfer Learning
Yutong Zhang (Beihang University), Jiaxin Chen (United Arab Emirates University)
CodeComputational EfficiencySupervised Fine-TuningMixture of ExpertsVision Language ModelTextMultimodality
🎯 What it does: Propose a Mixed Precision Interactive Side Expert Network (MP-ISMoE), which significantly improves the performance of memory-efficient transfer learning by efficiently quantizing the frozen backbone and introducing sparse MoE in the side network.
MPA: Multimodal Prototype Augmentation for Few-Shot Learning
Liwen Wu (Yunnan University), Bin Pu (Yunnan University)
CodeClassificationMeta LearningLarge Language ModelVision Language ModelMultimodality
🎯 What it does: This paper proposes a multi-modal prototype enhancement framework called MPA, which improves few-shot learning performance through LLM-generated diverse semantics, hierarchical multi-perspective enhancement, and adaptive absorption of uncertain categories.
CodeOptimizationComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelAgentic AIPrompt EngineeringTextGraph
🎯 What it does: Propose a node-level multi-agent system MPAS based on graph neural network message propagation mechanisms, breaking the traditional sequential topology constraints to enable parallel information exchange under arbitrary topologies.
MPI-Mamba: Latent Feature Fusion Mamba for Anisotropic Image Calibration and Deblurring in Magnetic Particle Imaging
Liwen Zhang (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences), Jie Tian (Southeast University)
CodeRestorationDiffusion modelBiomedical Data
🎯 What it does: Propose MPI-Mamba, an end-to-end framework integrating Mamba-based LFF-SSM with conditional latent diffusion models (CL-DM), for anisotropic image correction and deblurring in magnetic particle imaging (MPI) caused by unidirectional scanning.
MPMA: Preference Manipulation Attack Against Model Context Protocol
Zihan Wang (University of Electronic Science and Technology of China), Guowen Xu (University of Electronic Science and Technology of China)
CodeAdversarial AttackLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposed a novel security threat targeting the Model Context Protocol (MCP) called MPMA, and designed two attack strategies: Direct Preference Manipulation Attack (DPMA) and Genetic Algorithm-based Preference Manipulation Attack using Advertising Strategies (GAPMA); conducted systematic experiments on attack effectiveness and stealthiness.
🎯 What it does: Propose MRGeo, a systematic approach for robust cross-view geolocation (CVGL) under image distortion environments, primarily through multi-layer defense strategies to enhance feature quality and enforce geometric consistency.
MrM: Black-Box Membership Inference Attacks Against Multimodal RAG Systems
Peiru Yang (Tsinghua University), Tao Qi (Beijing University of Posts and Telecommunications)
CodeRetrievalSafty and PrivacyAdversarial AttackVision Language ModelImageRetrieval-Augmented Generation
🎯 What it does: This work proposes MrM, a black-box membership inference attack (MIA) framework targeting multi-modal retrieval-augmented generation (RAG) systems, which leverages visual object masking and adversarial mask selection to induce the system to leak information about whether data exists in the knowledge base during retrieval and generation stages.
🎯 What it does: Propose a hybrid RWKV-Transformer (MRT) architecture that achieves ultra-low bitrate image compression using 1-D sparse representations;
🎯 What it does: Proposed a reliable parking spot search algorithm called MS-PPO for structured parking environments, aiming to minimize a linear combination of the mean and standard deviation of parking search time.
🎯 What it does: This paper proposes MSAT-LDM, a high-fidelity and transferable watermark embedding framework implemented on a latent diffusion model (LDM).
MSCFL: Model Structure-Aware Clustered Federated Learning for System Heterogeneity and Data Drift
Yang Xu (Hunan University), Yaoxue Zhang (Tsinghua University)
CodeDomain AdaptationFederated LearningImage
🎯 What it does: Proposes the MSCFL framework, integrating model pruning with clustering federated learning to address system heterogeneity, data heterogeneity, and data drift issues.
MSME: A Multi-Stage Multi-Expert Framework for Zero-Shot Stance Detection
Yuanshuo Zhang (Minzu University of China), Xiaobing Zhao (Minzu University of China)
CodeClassificationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed the MSME (Multi-Stage Multi-Expert) framework, using three-stage decomposition (knowledge preparation, expert reasoning, decision aggregation) to address issues such as missing background knowledge, unclear target-label mapping, and rhetorical complexity in zero-shot stance detection.
MTAttack: Multi-Target Backdoor Attacks Against Large Vision-Language Models
Zihan Wang (Beihang University), Xiao Bai (Beihang University)
CodeAdversarial AttackSupervised Fine-TuningVision Language ModelImageText
🎯 What it does: The study addresses multi-target backdoor attacks in large-scale vision-language models (LVLMs), proposing the MTAttack framework to successfully implant backdoors.
MTP: Exploring Multimodal Urban Traffic Profiling with Modality Augmentation and Spectrum Fusion
Haolong Xiang (Nanjing University of Information Science and Technology), Wei Fan (University of Auckland)
CodeClassificationRepresentation LearningTransformerLarge Language ModelContrastive LearningMultimodalityTime Series
🎯 What it does: Proposed a multi-modal traffic flow portrait framework called MTP, which learns numerical, visual, and textual multi-modal features in the frequency domain and achieves fine-grained urban traffic state portraits through hierarchical contrastive learning.
🎯 What it does: Propose a multi-task reinforcement learning method, MTRL-CG, which groups related tasks using spectral clustering and reduces negative interference by employing group-specific policies.
🎯 What it does: Proposed a general pre-training method called MUG for heterogeneous graphs, which can be pre-trained on different heterogeneous graph datasets and directly transferred to downstream tasks;
🎯 What it does: This paper proposes a multi-agent parameter-free update context learning coordination framework MAICC, which achieves rapid collaborative adaptation through decentralized memory retrieval.
🎯 What it does: This paper proposes a multi-agent pointer Transformer (MAPT) framework based on Transformer, designed to solve the multi-vehicle dynamic random request delivery problem, generating joint action sequences through self-regressive decoding at each step;
Multi-Agent Undercover Gaming: Hallucination Removal Through Counterfactual Test for Multimodal Reasoning
Dayong Liang (South China University of Technology), Changmeng Zheng (Hong Kong Polytechnic University)
CodeAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Propose a Multi-Agent Covert Game (MUG) protocol that detects and eliminates hallucinations in LLMs by generating counterfactual images for counterfactual testing.
CodeRecommendation SystemTransformerLarge Language ModelAuto EncoderContrastive LearningMultimodality
🎯 What it does: This paper proposes the MACRec framework, which utilizes cross-modal quantization to generate hierarchical semantic IDs and introduces multi-dimensional cross-modal alignment in the generative recommendation stage, thereby significantly enhancing recommendation performance.
Multi-granularity Temporal Knowledge Editing over Large Language Models
Simiao Zhao (National University of Defense Technology), Xiang Zhao (National University of Defense Technology)
CodeTransformerLarge Language ModelTime SeriesSequentialBenchmark
🎯 What it does: This paper proposes a Multi-Granularity Temporal Knowledge Editing (MTKE) framework and constructs a corresponding benchmark dataset, MTKE.
Multi-Horizon Time Series Forecasting of Non-Parametric CDFs with Deep Lattice Networks
Niklas Erdmann (University of Oslo), Paal E. Engelstad (University of Oslo)
CodeRecurrent Neural NetworkTime Series
🎯 What it does: Achieved same-quantile quantile regression (SQR) prediction for multi-time domain non-parametric cumulative distribution functions (CDF) by combining deep lattice networks (DLN) with LSTM, and first applied DLN to multi-step forecasting of time series.
Multi-Level Domain Adaptation and Contrastive Domain Isolation with Bilinear Fusion for Patient Drug Response Prediction
Yuting Bai (Hunan University), Jiawei Luo (Hunan University)
CodeDomain AdaptationDrug DiscoveryContrastive LearningBiomedical Data
🎯 What it does: Designed and implemented a three-stage hierarchical domain adaptation framework, MACB-DRP, to transfer knowledge of cell line drug sensitivity to patient drug response prediction.
Multi-level Style Preference Optimization: An Adaptive Detection Framework for Human-Machine Hybrid Text
Zehao Wang (Northwestern Polytechnical University), Yaxiong Wang (Northwestern Polytechnical University)
CodeClassificationData-Centric LearningText
🎯 What it does: Propose the Multi-level Style Preference Optimization (MSPO) framework, which detects human-machine mixed text through multi-level (sequence, phrase, lexical) style preference optimization.
Multi-Modal Assistance for Unsupervised Domain Adaptation on Point Cloud 3D Object Detection
Shenao Zhao (Zhengzhou University), Zhoufan Yang (Zhengzhou University)
CodeObject DetectionDomain AdaptationAutonomous DrivingTransformerLarge Language ModelVision Language ModelMultimodalityPoint Cloud
🎯 What it does: Proposed the MMAssist framework, which achieves multi-modal assisted unsupervised domain adaptation for point cloud 3D object detection by introducing image and text features as bridges between source and target domain 3D detection models.
Multi-Task Test-time Adaptation via Gradient Consensus and Plasticity Constraint
Zhong Ye (Guangdong University of Technology), Zhenguo Yang (Guangdong University of Technology)
CodeDomain AdaptationImage
🎯 What it does: This paper proposes a multi-task test-time adaptation method called CoCo-MT-TTA to address gradient conflicts and catastrophic forgetting in multi-task scenarios.
Multi-Value Alignment for LLMs via Value Decorrelation and Extrapolation
Hefei Xu (Hefei University of Technology), Hao Liu (Hefei University of Technology)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Propose a multivalued alignment framework called MVA to address parameter interference issues in large language models (LLMs) when dealing with conflicting human values. The framework utilizes information-theoretic regularization and parameter extrapolation to generate diverse models with multiple values.
Multi-View Differential Mixing and Graph-Guided Structural Region Selection for Cross-Modal Alignment
Linlin Ji (Shandong Normal University), Li Liu (Shandong Normal University)
CodeRetrievalTransformerMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: This paper proposes the MG-Net framework to address the global representation information bottleneck and the lack of spatial structure in local features in image-text matching;
Multi-view Invariance Learning for 3D Scene Graph Pre-training via Collaborative Cross-Modal Regularization
Yucheng Huang (University of Electronic Science and Technology of China), Jiayuan Sun (University of Electronic Science and Technology of China)
CodeRepresentation LearningTransformerVision Language ModelPoint Cloud
🎯 What it does: This paper proposes a multi-view invariant learning framework for 3D scene graph pre-training, combining cross-modal regularization and knowledge filtering gate adapters to achieve self-supervised learning of object and predicate features.
Huabin Wang (Anhui University), Zilong Ling (Anhui University)
CodeRestorationGenerative Adversarial NetworkImagePhysics Related
🎯 What it does: Propose an MGT-Net based on multi-window Gabor transform and defect guidance, capable of automatically enhancing gain and defects in original GPR B-Scan images;
🎯 What it does: Built MULTIBENCH++, a unified large-scale multimodal fusion benchmark covering 30+ datasets, 15+ modalities, and 20+ tasks, along with an automated evaluation pipeline and standard implementations.
Multigranular Evaluation for Brain Visual Decoding
Weihao Xia (University of Cambridge), Cengiz Oztireli (University of Cambridge)
CodeExplainability and InterpretabilityGraph Neural NetworkLarge Language ModelVision Language ModelVideoMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Propose the BASIC framework to perform multi-grained structural, reasoning, and contextual three-dimensional evaluation of brain vision decoding results.
🎯 What it does: Proposes MultiMedBench, a VQA benchmark specifically designed for medical multimodal knowledge editing, covering two tasks: understanding and reasoning.
Multimodal Robust Prompt Distillation for 3D Point Cloud Models
Xiang Gu (Nanjing University of Science and Technology), Shuchao Pang (Nanjing University of Industry Technology)
CodeClassificationKnowledge DistillationPrompt EngineeringVision Language ModelContrastive LearningMultimodalityPoint Cloud
🎯 What it does: Propose a multi-modal robust prompt distillation framework MRPD, which enhances the adversarial robustness of point cloud models by distilling robust knowledge from image, text, and 3D teacher models into lightweight prompts;
🎯 What it does: Proposed MRiemGNN, a multi-modal heterogeneous graph neural network that co-learns in Euclidean and Riemannian spaces, using relation-aware kernelized message passing to capture composite relationships and enhancing representations through bidirectional mutual distillation.
🎯 What it does: Propose MultiTab-Net, a multi-task Transformer specifically designed for large-scale tabular data, and construct MultiTab-Bench as a synthetic data generation tool that allows adjustable task relevance, difficulty, and quantity.
CodeClassificationConvolutional Neural NetworkTransformerVision Language ModelImage
🎯 What it does: A multi-task deep evidence fusion network (DEFNet) was constructed, which achieves no-reference image quality assessment by simultaneously optimizing three tasks: BIQA, scene classification, and distortion type classification, and adopting cross-subregion and local-global two-level trustworthy information fusion.
MUSE: Multi-Scale Dense Self-Distillation for Nucleus Detection and Classification
Zijiang Yang (University of Science and Technology Beijing), Hui Jiang (Alibaba Group)
CodeClassificationObject DetectionTransformerBiomedical Data
🎯 What it does: This paper proposes a self-supervised learning framework called MUSE for nucleus detection and classification, employing a multi-scale self-distillation strategy and equipped with an expandable ViT encoder-decoder network.