AAAI Conference on Artificial Intelligence · 1442 papers
MaFeRw: Query Rewriting with Multi-Aspect Feedbacks for Retrieval-Augmented Large Language Models
Yujing Wang (Beihang University), Zhiming Zheng (Beijing Academy of Blockchain and Edge Computing)
CodeGenerationRetrievalTransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Developed a multi-dimensional feedback query rewriting method called MaFeRw, which enhances the retrieval and generation quality of multi-turn dialogue RAG systems.
MAGIC: Generating Self-Correction Guideline for In-Context Text-to-SQL
Arian Askari (Leiden University), Xinye Tang (Microsoft)
CodeGenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Proposes the MAGIC multi-agent framework, which automatically generates text-to-SQL self-correction criteria and uses them during the inference phase;
Maintaining Fairness in Logit-based Knowledge Distillation for Class-Incremental Learning
Zijian Gao (National University of Defense Technology), Huaimin Wang (National University of Defense Technology)
CodeClassificationKnowledge DistillationImage
🎯 What it does: This paper proposes a preprocessing method for Z-score normalization of the logit of student and teacher models, eliminating the conflict between KD and CE objectives, thereby maintaining fairness between old and new classes in class-incremental learning, and further enhancing internal fairness within old classes through internal class relationship distillation.
🎯 What it does: A method for incremental learning using domain shift and contrastive learning, called DisCo, is proposed to reduce catastrophic forgetting in class-incremental learning.
Making Large Vision Language Models to Be Good Few-Shot Learners
Fan Liu (Hohai University), Jun Zhou (Hong Kong University of Science and Technology)
CodeClassificationMeta LearningTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: By performing meta-learning-based instruction fine-tuning on large-scale visual language models, and incorporating character perturbation label enhancement and attribute description-based candidate filtering, the performance in few-shot classification tasks is significantly improved.
🎯 What it does: Developed the MalCL system, which uses GAN to generate replay samples for continual learning to mitigate catastrophic forgetting in malware classification.
MalDetectFormer: Leveraging Sparse SpatioTemporal Information for Effective Malicious Traffic Detection
Shuai Zhang (Zhongguancun Laboratory), Bo Li (Beihang University)
CodeAnomaly DetectionTransformerTime Series
🎯 What it does: Proposes the MalDetectFormer model, which enhances malicious traffic detection and identification using sparse spatiotemporal attention and subgraph convolutional networks;
Mamba YOLO: A Simple Baseline for Object Detection with State Space Model
Zeyu Wang (Zhejiang Normal University), Hongbo Li (Beijing Geekplus Technology)
CodeObject DetectionImage
🎯 What it does: A real-time object detection framework Mamba YOLO based on a linear complexity state space model (SSM) is proposed, achieving high-performance detection models trained from scratch without large-scale pre-training.
🎯 What it does: The MambaPro framework is proposed for the task of multi-modal person re-identification, combining the CLIP pre-trained model, parallel feedforward adapters, synergistic residual prompts, and the Mamba aggregation module to achieve efficient feature learning.
MAMS: Model-Agnostic Module Selection Framework for Video Captioning
Sangho Lee (Sungkyunkwan University), Hogun Park (Sungkyunkwan University)
CodeGenerationTransformerVideoTextMultimodality
🎯 What it does: This paper proposes the MAMS (Model-Agnostic Module Selection) framework, which can dynamically select appropriately sized generation modules based on the visual information of each video segment and filter important visual tokens at both the frame and word levels.
🎯 What it does: Proposes the Manta framework, adapting Mamba for few-shot action recognition of long subsequences, and incorporates Matryoshka Mamba and mixed contrastive learning;
Mapping from Meaning: Addressing the Miscalibration of Prompt-Sensitive Language Models
Kyle Cox (University of Texas at Austin), Ying Ding (University of Texas at Austin)
CodeTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Using semantic-preserving paraphrase perturbation sampling and embedding variance uncertainty measurement, the uncertainty of LLM in question-answering tasks is evaluated and calibrated for improvement.
Rishabh Agrawal (University of Southern California), Ashutosh Nayyar (University of Southern California)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningFlow-based ModelSequential
🎯 What it does: In a strictly offline batch imitation learning environment, an algorithm based on the Markov balance equation is proposed, which estimates the state-action transition density using conditional normal flows and combines it with behavior cloning to achieve learning without rewards and interactions.
Masked Language Modeling Becomes Conditional Density Estimation for Tabular Data Synthesis
Seunghwan An (University of Seoul), Jong-June Jeon (University of Seoul)
CodeData SynthesisTransformerTabular
🎯 What it does: A conditional density estimation method based on a masked language model, MaCoDE, is proposed to generate synthetic data that retains the statistical features of tabular data while achieving high machine learning performance.
CodeRetrievalTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A point mutual information-based image-text matching scoring framework called MASS is proposed, which requires no additional training and aims to reduce the language bias of visual language models.
Massively Parallel Continuous Local Search for Hybrid SAT Solving on GPUs
Yunuo Cen (National University of Singapore), Xuanyao Fong (National University of Singapore)
CodeOptimizationBenchmark
🎯 What it does: A parallel algorithm based on FFT heuristics, FastFourierSAT, is proposed for efficiently executing continuous local search to solve mixed SAT formulas on GPUs.
🎯 What it does: This paper proposes the SemaGlue framework, which injects semantic features from a pre-trained semantic segmentation model into the image feature matching process to improve matching accuracy in sparse textures and occluded scenes.
Math-PUMA: Progressive Upward Multimodal Alignment to Enhance Mathematical Reasoning
Wenwen Zhuang (University of Chinese Academy of Sciences), Jin Zeng (University of Chinese Academy of Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodalityBenchmark
🎯 What it does: A three-stage training framework called Math-PUMA is proposed and implemented to enhance the performance of multimodal large language models in mathematical reasoning tasks.
MathSpeech: Leveraging Small LMs for Accurate Conversion in Mathematical Speech-to-Formula
Sieun Hyeon (Seoul National University), Jaeyoung Do (Seoul National University)
CodeRecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningTextAudio
🎯 What it does: Proposes the MathSpeech pipeline, which first corrects ASR output using a small LLM, and then directly generates LaTeX formulas, supporting the generation of subtitles for math lectures.
MCGAN: Enhancing GAN Training with Regression-Based Generator Loss
Baoren Xiao (University College London), Weixin Yang (University of Oxford)
CodeGenerationData SynthesisGenerative Adversarial NetworkImageVideoTime SeriesFinance Related
🎯 What it does: An algorithm named MCGAN is proposed, which enhances the training of Generative Adversarial Networks (GANs) through a regression loss function, addressing the lack of supervision in generator training.
🎯 What it does: This paper first constructs the first multimodal Chinese idiom reading comprehension dataset MChIRC (44,433 image-text pairs, 2,926 idioms), and then proposes a multimodal idiom reading comprehension model DCIGN based on dual contrastive learning and graph convolutional networks, using images to assist text in enhancing idiom understanding and selection.
MDD-5k: A New Diagnostic Conversation Dataset for Mental Disorders Synthesized via Neuro-Symbolic LLM Agents
Congchi Yin (Theta Health Inc.), Xun Jiang (Theta Health Inc.)
CodeGenerationData SynthesisTransformerLarge Language ModelTextElectronic Health Records
🎯 What it does: A neural-symbolic multi-agent framework was constructed to synthesize diverse psychological disorder diagnostic dialogues using anonymized real medical records, and the largest annotated diagnostic dialogue dataset in Chinese, MDD-5k, was released.
🎯 What it does: This paper studies the problem of multi-dimensional fine-grained fatigue detection and proposes the MDFG method, which automatically generates multi-dimensional fine-grained labels (time, type, level) using eye movement wavelet features, and constructs intermediate fatigue states through base class extraction and synthesis to achieve accurate recognition of real fatigue videos.
Measuring Cross-Modal Interactions in Multimodal Models
Laura Wenderoth (University of Cambridge), Mateja Jamnik (University of Cambridge)
CodeExplainability and InterpretabilityMultimodalityTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes and implements an interaction measurement for multimodal models called InterSHAP, which quantifies the interaction effects between different modalities and provides explanations at both global and local levels.
Measuring Human and AI Values Based on Generative Psychometrics with Large Language Models
Haoran Ye (Peking University), Guojie Song (Peking University)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A generative psychological measurement framework based on large language models (GPV) is proposed and implemented, which can automatically extract 'perceptions' from unstructured text and convert them into value vectors for both humans and LLMs.
MEATRD: Multimodal Anomalous Tissue Region Detection Enhanced with Spatial Transcriptomics
Kaichen Xu (Zhongnan University of Economics and Law), Xiaobo Sun (Zhongnan University of Economics and Law)
CodeAnomaly DetectionGraph Neural NetworkTransformerMultimodalityBiomedical Data
🎯 What it does: This paper proposes a multi-modal abnormal tissue region detection method called MEATRD, which combines histological images with spatial transcriptomic data to achieve precise localization and identification of abnormal tissue regions (ATR).
CodeClassificationSegmentationGenerationData SynthesisSafty and PrivacyTransformerSupervised Fine-TuningDiffusion modelGenerative Adversarial NetworkImageBiomedical Data
🎯 What it does: A MeMa dataset (42,307 synthetic patient facial images) was constructed using a Stable Diffusion generative model based on real patient photos, and MedSem-DeID was proposed on this dataset, achieving facial identity removal while retaining medical diagnostic information with reversible decryption.
Medical MLLM Is Vulnerable: Cross-Modality Jailbreak and Mismatched Attacks on Medical Multimodal Large Language Models
Xijie Huang (Beihang University), Chengwei Pan (Beihang University)
CodeAdversarial AttackTransformerLarge Language ModelMultimodalityBiomedical Data
🎯 What it does: This paper studies the security vulnerabilities of medical multimodal large language models (MedMLLM), proposing cross-modal jailbreaking and mismatch attacks, and constructing the 3MAD dataset.
🎯 What it does: Developed MegActor-Sigma, a hybrid modal conditional diffusion transformer (DiT) that allows for flexible control of portrait animation using both visual and audio signals.
🎯 What it does: The MEMatte framework is proposed, introducing a router and a lightweight token refinement module before the global attention in the Transformer, significantly reducing memory usage and inference speed for high-resolution image matting.
🎯 What it does: A memory-enhanced re-completion framework MARE is proposed to address the issue of insufficient visibility in 3D semantic scene completion from camera inputs.
Memory-Reduced Meta-Learning with Guaranteed Convergence
Honglin Yang (Xiamen University), Xiao Yu (Xiamen University)
CodeOptimizationMeta LearningImage
🎯 What it does: A meta-learning algorithm is designed that does not use historical lower-level parameters/gradients, significantly reducing memory consumption while ensuring sublinear convergence.
🎯 What it does: A large-scale multimodal dataset of Chinese adolescent anxiety and depression, MMPsy, has been constructed, and a Mental-Perceiver model based on Transformer full attention has been proposed to estimate psychological disorders from both speech and text simultaneously.
🎯 What it does: The MENTOR framework is proposed in multimodal recommendation, utilizing multi-level self-supervised learning for cross-modal alignment while preserving interaction information.
Mesh Watermark Removal Attack and Mitigation: A Novel Perspective of Function Space
Xingyu Zhu (Southern University of Science and Technology), Xuetao Wei (Southern University of Science and Technology)
CodeRestorationAdversarial AttackMesh
🎯 What it does: This paper proposes a function space-based attack method called FUNCEVADE and a corresponding defense method called FUNCMARK, aimed at achieving robustness against topological transformations in 3D mesh watermarking.
🎯 What it does: This paper proposes the Mesorch architecture, which uses a parallel CNN and Transformer dual-path encoder to simultaneously extract micro details and macro semantics, achieving precise localization of image tampering areas.
MetaNeRV: Meta Neural Representations for Videos with Spatial-Temporal Guidance
Jialong Guo (Zhejiang University), Haishuai Wang (Shanghai Jiaotong University)
CodeCompressionOptimizationMeta LearningVideo
🎯 What it does: This paper proposes MetaNeRV, a meta-learning based NeRV video implicit representation framework that can quickly adapt to new videos and supports multi-scale spatial supervision and advanced temporal guidance.
MetaSymNet: A Tree-like Symbol Network with Adaptive Architecture and Activation Functions
Yanjie Li (Institute of Semiconductor, Chinese Academy of Sciences), Meilan Hao (Institute of Semiconductor, Chinese Academy of Sciences)
CodeOptimizationTabularBenchmarkPhysics Related
🎯 What it does: This paper proposes MetaSymNet, a tree-structured adaptive symbolic network that transforms symbolic regression from combinatorial optimization to numerical optimization, and achieves automatic symbolic selection through an evolvable PANGU meta-function.
🎯 What it does: The MacFRL method is proposed in the context of continual learning, utilizing task reordering and flexible representation learning to achieve group fairness in continual learning with single-label initial tasks and unlabeled subsequent tasks.
MFL-Owner: Ownership Protection for Multi-modal Federated Learning via Orthogonal Transform Watermark
Keke Gai (Beijing Institute of Technology), Qi Wu (Australian Institute of Machine Learning)
CodeFederated LearningSafty and PrivacyImageTextMultimodality
🎯 What it does: The MFL-Owner framework is proposed, which uses orthogonal transformations to embed watermarks, protecting the ownership of multimodal federated learning models and enabling traceability.
MHBench: Demystifying Motion Hallucination in VideoLLMs
Ming Kong (Zhejiang University), Qiang Zhu (Zhejiang University)
CodeClassificationRecognitionTransformerVision Language ModelContrastive LearningVideoBenchmark
🎯 What it does: This paper first introduces the concept of 'Motion Hallucination' in video LLMs and systematically evaluates it by constructing the MHBench benchmark.
🎯 What it does: MimiQ is proposed, a data-free quantization method for Vision Transformers that generates synthetic samples and performs distillation by aligning multi-head attention similarities to achieve low-bit quantization.
🎯 What it does: This paper proposes a primary graph learning framework (PGL) that extracts primary subgraphs from the user-item interaction graph and performs message passing on them to better utilize local structural information to enhance multimodal recommendation performance.
Mind the Uncertainty in Human Disagreement: Evaluating Discrepancies Between Model Predictions and Human Responses in VQA
Jian Lan (Ludwig Maximilian University of Munich), Barbara Plank (Ludwig Maximilian University of Munich)
CodeClassificationRecognitionTransformerVision Language ModelImageMultimodality
🎯 What it does: This paper studies human uncertainty in answers (HUD) in visual question answering (VQA), evaluates the differences between model prediction distributions and human response distributions, and explores whether temperature scaling calibration can improve alignment with humans.
🎯 What it does: This paper proposes a multi-view learning framework MVOL, which utilizes the few in-distribution (ID) features contained in auxiliary outlier samples during training to improve OOD detection performance.
Mitigating Hallucinations in Large Vision-Language Models by Adaptively Constraining Information Flow
Jiaqi Bai (Guangzhou University), Zhihong Tian (Guangzhou University)
CodeGenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: The ADAVIB method is proposed, which injects noise into the visual-language projector of LVLM using the Variational Information Bottleneck (VIB) and adaptively adjusts the noise intensity through entropy to suppress the overconfidence of visual tokens in the LLM embedding space, thereby reducing object hallucinations in generated text.
Mitigating Social Bias in Large Language Models: A Multi-Objective Approach Within a Multi-Agent Framework
Zhenjie Xu (Sun Yat-sen University), Zhichao Lu (City University of Hong Kong)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A multi-objective, multi-agent framework (MOMA) is proposed to mitigate social biases in large language models, primarily through causal interventions on the representations of social groups in the input using two agents: masking and balancing.
Mixture of Experts Based Multi-Task Supervise Learning from Crowds
Tao Han (Zhejiang Gongshang University), Yili Fang (Zhejiang Gongshang University)
CodeMixture of ExpertsImageText
🎯 What it does: A new paradigm of multi-task supervised learning - crowds, called MLC, is proposed, and a Mixture-of-Experts model MMLC is constructed on it. Two aggregation strategies are further provided: utilizing oracle worker discovery (MMLC-owf) and data filling based on (MMLC-df) to achieve high-quality true labeling.
🎯 What it does: A Multi-Layer Jump Augmented Auxiliary Network (MLAAN) is proposed to address the issues of local learning lacking global information and having a narrow perspective;
MM-CamObj: A Comprehensive Multimodal Dataset for Camouflaged Object Scenarios
Jiacheng Ruan (Shanghai Jiao Tong University), Yuzhuo Fu (Institute for Advanced Algorithms Research)
CodeObject DetectionSegmentationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodalityBenchmark
🎯 What it does: Constructed the MM-CamObj two subsets (CamObj-Align and CamObj-Instruct) and trained a specialized CamObj-Llava based on this, proposing CamObj-Bench for evaluating multimodal models in camouflage scenarios.
🎯 What it does: Proposes the MOCID system, which improves the detection of small moving infrared targets by bidirectionally fusing clip-level motion context with frame-level displacement information.
🎯 What it does: A MASRC framework based on the multimodal (entity and scene) interrelationship is proposed, which captures long-term and short-term associations using graph networks and detects video scene boundaries through multi-shot comparison.
🎯 What it does: A multimodal recommendation framework based on modality-independent receptive fields in graph neural networks and a sampling-based global Transformer is proposed, which can simultaneously utilize text, visual, and learnable embedding information to enhance the expressiveness of user-item representations.
🎯 What it does: This paper proposes a method for measuring model change lineage, which can quantify the degree of modification between two models (Lineage closeness).
🎯 What it does: A new graph federated learning method called FedIIH is proposed, which can simultaneously consider the inter-heterogeneity of graph data across subgraphs and the intra-heterogeneity within subgraphs, enhancing the performance of distributed graph models.
🎯 What it does: This paper proposes the MOL-Mamba framework, which enhances molecular representation by integrating structural and electronic information on molecular graphs.
CodeGenerationTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageText
🎯 What it does: A training-free decoding method called MoLE is proposed, which significantly reduces the phenomenon of hallucination in text generation by selecting experts from different layers of a large visual language model and generating collaboratively.
🎯 What it does: Proposes the Momentum Pseudo Labeling (MPL) framework, which utilizes a momentum model to achieve global pseudo label updates and considers false negative samples in non-matching image-text pairs during contrastive learning, thereby enhancing weakly supervised phrase localization performance.
MonoBox: Tightness-Free Box-Supervised Polyp Segmentation Using Monotonicity Constraint
Qiang Hu (Wuhan National Laboratory for Optoelectronics), Zhiwei Wang (Wuhan National Laboratory for Optoelectronics)
CodeObject DetectionSegmentationImage
🎯 What it does: A box-supervised polyp segmentation method named MonoBox is proposed, which can achieve pixel-level segmentation without satisfying the box compactness assumption.
More Text, Less Point: Towards 3D Data-Efficient Point-Language Understanding
Yuan Tang (Huazhong University of Science and Technology), Min Chen (South China University of Technology)
CodeClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityPoint Cloud
🎯 What it does: The GreenPLM model is proposed, achieving language understanding and generation for 3D objects with only 12% of 3D training samples or even without 3D data, through a three-stage training process, 0M-Pooling, and large-scale text data.
🎯 What it does: Proposes the MoRe framework, which utilizes class-patch attention from Vision Transformer and generates more refined pseudo-labels for weakly supervised semantic segmentation through graph regularization and CAM-guided attention regularization.
MORE: Molecule Pretraining with Multi-Level Pretext Task
Yeongyeong Son (Pusan National University), Sunyoung Kwon (Pusan National University)
CodeRepresentation LearningDrug DiscoveryGraph Neural NetworkGraphBiomedical Data
🎯 What it does: A multi-level molecular graph pre-training framework called MORE is proposed, which combines self-supervised tasks of nodes, subgraphs, graphs, and 3D perspectives;
🎯 What it does: The MOS (Model Surgery) method is proposed, utilizing task-specific adapters, adapter fusion, and a self-correcting retrieval mechanism to address the catastrophic forgetting problem encountered by pre-trained models in class-incremental learning.
🎯 What it does: This paper proposes a pedestrian re-identification method based on 3D skeletons called MoCos, which integrates graphical synapses and combinatorial skeleton prototype learning to enhance skeleton representation.
🎯 What it does: This paper proposes an unsupervised motion artifact removal method called PFAD, which utilizes pixel-frequency alternating masks to guide a diffusion model in recovering clean MRI images.
🎯 What it does: A dynamic 3D Gaussian rendering framework M5D-GS is proposed, which decouples motion and deformation, achieving high-quality reconstruction and real-time rendering of significantly moving objects using a monocular camera.
Motion Prior Knowledge Learning with Homogeneous Language Descriptions for Moving Infrared Small Target Detection
Shengjia Chen (University of Electronic Science and Technology of China), Mao Ye (University of Electronic Science and Technology of China)
CodeObject DetectionConvolutional Neural NetworkRecurrent Neural NetworkGraph Neural NetworkVision Language ModelVideoTextMultimodality
🎯 What it does: This paper proposes a motion prior knowledge learning framework MoPKL based on audiovisual fusion, which guides the visual channel to learn the motion characteristics of small infrared targets through homogeneous language descriptions, thereby achieving fine detection of moving small targets.
Motion-adaptive Transformer for Event-based Image Deblurring
Senyan Xu (University of Science and Technology of China), Yan Chen (University of Science and Technology of China)
CodeRestorationTransformerImage
🎯 What it does: A Motion-Adaptive Transformer (MAT) is proposed, which improves the fusion and attention mechanism of events and images by utilizing the spatial motion information provided by event cameras, achieving high-quality motion blur removal.
🎯 What it does: A method called MPQ-DM is proposed, which implements mixed precision quantization and time-smooth distillation for extremely low-bit quantization of diffusion models.
MPTSNet: Integrating Multiscale Periodic Local Patterns and Global Dependencies for Multivariate Time Series Classification
Yang Mu (Technical University of Munich), Xiao Xiang Zhu (Munich Center for Machine Learning)
CodeClassificationConvolutional Neural NetworkTime Series
🎯 What it does: Proposed and implemented MPTSNet for multivariate time series classification, combining multi-scale periodic decomposition, CNN, and attention mechanisms to simultaneously extract local and global features.
MRBTP: Efficient Multi-Robot Behavior Tree Planning and Collaboration
Yishuai Cai (National University of Defense Technology), Ji Wang (National University of Defense Technology)
CodeOptimizationRobotic IntelligenceLarge Language ModelTabular
🎯 What it does: A provably sound and complete behavior tree planning algorithm for multi-robot systems, MRBTP, has been designed, supporting cross-tree expansion, backup structures, intention sharing, and providing an optional LLM subtree pre-planning plugin.
mRNA2vec: mRNA Embedding with Language Model in the 5'UTR-CDS for mRNA Design
Honggen Zhang (University of Hawaii), Lipeng Lai (XtalPi Inc)
CodeOptimizationDrug DiscoveryTransformerLarge Language ModelBiomedical Data
🎯 What it does: This paper proposes an mRNA embedding method called mRNA2vec based on data2vec, which jointly utilizes 5' UTR and CDS sequences for contextual pre-training. It enhances mRNA translation efficiency and expression level prediction through hard masking and auxiliary tasks (MFE regression, secondary structure classification).
MSAmba: Exploring Multimodal Sentiment Analysis with State Space Models
Xilin He (Shenzhen University), Zitong Yu (University of Exeter)
CodeClassificationTransformerMultimodality
🎯 What it does: This paper proposes a multimodal sentiment analysis framework called MSAmba based on Mamba, which mainly implements modeling of internal sequence information of each modality and cross-modal interaction through two major modules (ISM and CHM);
MSE-Adapter: A Lightweight Plugin Endowing LLMs with the Capability to Perform Multimodal Sentiment Analysis and Emotion Recognition
Yang Yang (South China University of Technology), Yupeng Qiang (South China University of Technology)
CodeClassificationRecognitionTransformerLarge Language ModelTextMultimodality
🎯 What it does: A lightweight plugin MSE-Adapter is proposed, enabling large language models to perform multimodal sentiment analysis and emotion recognition tasks with approximately 2.6M–2.8M trainable parameters while maintaining their original general capabilities.
MSSDA: Multi-Sub-Source Domain Adaptation for Diabetic Foot Neuropathy Recognition
Yan Zhong (South China University of Technology), Peiru Zhou (Jinan University)
CodeRecognitionDomain AdaptationConvolutional Neural NetworkContrastive LearningTime SeriesBiomedical Data
🎯 What it does: This paper proposes a Multi-Source Subdomain Adaptation framework (MSSDA) that utilizes continuous plantar pressure data collected from wearable shoes to identify diabetic foot neuropathy.
🎯 What it does: This paper proposes a multi-perspective event camera lip-reading framework called MTGA, which integrates features from both event frames and time-segmented voxel graph lists, achieving high-precision recognition of lip movements through spatiotemporal granularity alignment fusion and a backend implementation using position encoding, Bi-GRU, and Self-Attention.
🎯 What it does: A multi-teacher bytecode vulnerability detection framework called MTVHunter is designed, combining instruction denoising and semantic completion from two major teachers to enhance the security detection of smart contract bytecode.
🎯 What it does: This paper proposes a 3D human reconstruction method called MUC for uncalibrated multi-view cameras, which utilizes a single-view pre-trained encoder to generate human models and camera position information for each camera. It dynamically fuses joint and surface information from different views through a Joint Re-weighting Network (JRN) and a Surface Re-weighting Network (SRN) to obtain a detailed SMPL-X human mesh.
🎯 What it does: A distributed task allocation framework called REDA based on multi-agent reinforcement learning is proposed, which addresses time-related allocation problems and achieves global optimality.
Multi-Agent Motion Planning for Differential Drive Robots Through Stationary State Search
Jingtian Yan (Carnegie Mellon University), Jiaoyang Li (Carnegie Mellon University)
CodeOptimizationRobotic IntelligenceSimultaneous Localization and MappingBenchmark
🎯 What it does: A three-layer framework called MASS is proposed, which uses MAPF to resolve conflicts, SSIPP to search for resting states and generate feasible trajectories, and SPS to optimize speed, extending the method to lifelong multi-agent motion planning.
Multi-aspect Self-guided Deep Information Bottleneck for Multi-modal Clustering
Shizhe Hu (Zhengzhou University), Yangdong Ye (Zhengzhou University)
CodeRepresentation LearningImageTextMultimodality
🎯 What it does: A multi-modal clustering method is proposed - Multi-Scale Self-Guided Deep Information Bottleneck (MSDIB), which extracts discriminative features and completes clustering through information compression and retention as well as a self-guided mechanism.
Multi-Attribute Multi-Grained Adaptation of Pre-Trained Language Models for Text Understanding from Bayesian Perspective
You Zhang (Yunnan University), Xuejie Zhang (Yunnan University)
CodeDomain AdaptationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A multi-attribute multi-granularity adaptation framework M2A is proposed, which integrates IID and non-IID features of pre-trained language models (PLM) from a Bayesian perspective, and achieves adaptive enhancement of text understanding tasks through efficient fine-tuning using parameters like LoRA.
Multi-axis Prompt and Multi-dimension Fusion Network for All-in-one Weather-degraded Image Restoration
Yuanbo Wen (Chang'an University), Ting Chen (Australian National University)
CodeRestorationTransformerPrompt EngineeringImage
🎯 What it does: A multi-axis prompt and multi-dimensional fusion network, MPMF-Net, is proposed for the one-time removal of various weather noise in traffic images.
Multi-Branch Self-Drafting for LLM Inference Acceleration
Zipeng Gao (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes the Self-Draft method, which utilizes the multi-branch reasoning of LLM itself to generate and validate drafts in parallel during a single forward pass, thereby accelerating large language model inference.
Multi-clue Consistency Learning to Bridge Gaps Between General and Oriented Object in Semi-supervised Detection
Chenxu Wang (Nanjing University of Science and Technology), Zhen Cui (Nanjing University of Science and Technology)
CodeObject DetectionImage
🎯 What it does: This paper proposes a Multi-Channel Consistency Learning framework (MCL) aimed at addressing the issue of label assignment and confidence inconsistency in semi-supervised directed object detection.
Multi-Focus Image Fusion via Explicit Defocus Blur Modelling
Yuhui Quan (South China University of Technology), Hui Ji (National University of Singapore)
CodeImage TranslationRestorationConvolutional Neural NetworkRecurrent Neural NetworkImageBiomedical Data
🎯 What it does: This paper proposes a multi-focus image fusion network named DMANet, which generates pixel-level defocus descriptors and initial focused images through an explicit defocus blur model, and based on these results, generates soft decision maps and an uncertainty-aware fusion module, ultimately producing a panoramic focused image.
🎯 What it does: A unified framework called MQSPN based on multi-granularity query sets and set prediction is proposed to address the Grounded Multimodal Named Entity Recognition (GMNER) task based on image-text pairs.
🎯 What it does: A multi-granularity open intent classification method MOGB is proposed, which combines adaptive granularity spherical clustering and the nearest sub-center classifier for hierarchical representation learning, and constructs multi-granularity decision boundaries to distinguish known intents from unknown intents.
Multi-Label Ranking Loss Minimization for Matrix Completion
Jiaxuan Li (Xi'an Jiaotong University), Jiayin Wang (Xi'an Jiaotong University)
CodeRecommendation SystemDrug DiscoveryTabular
🎯 What it does: A matrix completion method based on multi-label ranking loss minimization (MLRM) is proposed, which uses the relative ranking among column vectors to constrain the completion results, thereby addressing the ambiguity caused by the low-rank assumption.
Xucong Wang (University of Science and Technology of China), Yang Wang (Suzhou Institute for Advanced Research, University of Science and Technology of China)
CodeObject DetectionKnowledge DistillationImage
🎯 What it does: The paper proposes a multi-label self-distillation method called MSKD, which implements self-teacher distillation using three spatial decoupling mechanisms.
🎯 What it does: By combining RGB and Depth visual modalities and introducing a multi-scale local-global spatial understanding mechanism, the M2SE-VTTS model is proposed to achieve immersive visual text-to-speech (VTTS) synthesis based on environmental images.
🎯 What it does: This paper proposes MMRecUn, a no-learning (unlearning) framework for multimodal recommendation systems (MMRS) that allows for selective forgetting of user and item interactions.