AAAI 2025 Papers — Page 22
AAAI Conference on Artificial Intelligence · 3028 papers
PBECount: Prompt-Before-Extract Paradigm for Class-Agnostic Counting
Canchen Yang (Sichuan University), Chun Xu (Sichuan University)
Object DetectionConvolutional Neural NetworkPrompt EngineeringImage
🎯 What it does: Proposes the Prompt-Before-Extract paradigm and designs a pure CNN model PBECount based on it to achieve class-agnostic object counting.
PC-BEV: An Efficient Polar-Cartesian BEV Fusion Framework for LiDAR Semantic Segmentation
Shoumeng Qiu (Fudan University), Jian Pu (Fudan University)
SegmentationAutonomous DrivingComputational EfficiencyTransformerPoint Cloud
🎯 What it does: This paper proposes a LiDAR semantic segmentation framework called PC-BEV, which fuses polar and Cartesian coordinate segmentation only within the BEV space, and achieves efficient dense feature fusion through a pre-computed remap method.
PCM Selector: Penalized Covariate-Mediator Selection Operator for Evaluating Linear Causal Effects
Hisayoshi Nanmo (Chugai Pharmaceutical Co), Manabu Kuroki (Yokohama National University)
Tabular
🎯 What it does: In the case of multicollinearity or high-dimensional data with missing confounding variables, a two-stage penalized regression method called PCM Selector is proposed to estimate causal effects in linear structural equation models.
PDBs Go Numeric: Pattern-Database Heuristics for Simple Numeric Planning
Daniel Gnad (Linkoping University), Alexander Shleyfman (Bar-Ilan University)
OptimizationReinforcement LearningBenchmark
🎯 What it does: A heuristic for simple numerical planning (SNP) with integer variables is proposed, along with various methods to handle the infinite state space of numerical variables.
PDDM: Pseudo Depth Diffusion Model for RGB-PD Semantic Segmentation Based in Complex Indoor Scenes
Xinhua Xu (Peking University), Jinfu Liu (Peking University)
SegmentationDepth EstimationDiffusion modelImage
🎯 What it does: A framework for RGB semantic segmentation based on pseudo depth is proposed, utilizing a multi-source pseudo depth aggregation module (PDAM) and a diffusion model (PDDM) to achieve high-precision segmentation of indoor scenes.
PEARL: Input-Agnostic Prompt Enhancement with Negative Feedback Regulation for Class-Incremental Learning
Yongchun Qin (Southeast University), Hui Xue (Southeast University)
ClassificationRepresentation LearningTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes an input-agnostic Prompt mechanism called PEARL for class-incremental learning, addressing the knowledge confusion problem caused by traditional Query-Select mechanisms.
Pedestrian Attribute Recognition: A New Benchmark Dataset and a Large Language Model Augmented Framework
Jiandong Jin (Anhui University), Chenglong Li (Anhui University)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningImageBenchmark
🎯 What it does: A new large-scale cross-domain pedestrian attribute recognition dataset MSP60K is proposed, and based on this dataset, the LLM-PAR framework is developed, which combines visual Transformer, MEQ-Former, and large language models to achieve joint learning of attribute classification and text description; a systematic benchmark of 17 existing PAR methods is conducted under random and cross-domain splits; the performance improvement of LLM-PAR is validated on MSP60K and other public datasets.
PEIE: Physics Embedded Illumination Estimation for Adaptive Dehazing
Huaizhuo Liu (Beihang University), Yurui Liu (Beihang University)
RestorationData SynthesisConvolutional Neural NetworkGenerative Adversarial NetworkImagePhysics Related
🎯 What it does: This paper proposes a physics-based embedded illumination estimation framework (PEIE) for adaptive dehazing in real-world scenarios.
Perception-Guided Jailbreak Against Text-to-Image Models
Yihao Huang (Nanyang Technological University), Yang Liu (Nanyang Technological University)
GenerationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a perception-oriented jailbreak method based on large language models (PGJ), which uses secure phrases to replace unsafe words without querying any T2I models, thereby bypassing text security checkers to generate NSFW images.
PerReactor: Offline Personalised Multiple Appropriate Facial Reaction Generation
Hengde Zhu (University of Leicester), Siyang Song (University of Cambridge)
GenerationTransformerGenerative Adversarial NetworkVideo
🎯 What it does: This work proposes PerReactor, an offline personalized multi-appropriate facial response generation (PMAFRG) framework that can generate diverse and personalized facial responses for a given speaker's behavior.
Personalized Clustering via Targeted Representation Learning
Xiwen Geng (Renmin University of China), Mengdie Wang (Renmin University of China)
Representation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: This paper proposes a personalized clustering framework called PCL, which utilizes active querying of the most uncertain and difficult negative samples to learn user preference-oriented feature representations, achieving clustering results that align with user needs.
Personalized Dynamic Music Emotion Recognition with Dual-Scale Attention-Based Meta-Learning
Dengming Zhang (Zhejiang University), Pei Chen (Zhejiang University)
RecognitionMeta LearningRecurrent Neural NetworkTransformerContrastive LearningAudio
🎯 What it does: A dual-scale attention meta-learning model (DSAML) is proposed to achieve personalized dynamic music emotion recognition (PDMER).
Personalized Federated Collaborative Filtering: A Variational AutoEncoder Approach
Zhiwei Li (University of Technology Sydney), Chengqi Zhang (Hong Kong Polytechnic University)
Recommendation SystemFederated LearningSafty and PrivacyAuto EncoderTabular
🎯 What it does: A federated collaborative filtering framework named FedDAE is proposed, which utilizes Variational AutoEncoder (VAE) combined with dual encoders and a gating network to achieve personalized recommendations without uploading raw data.
Personalized Federated Learning for Spatio-Temporal Forecasting: A Dual Semantic Alignment-Based Contrastive Approach
Qingxiang Liu (Institute of Computing Technology Chinese Academy of Sciences), Jingjing Xue (Institute of Computing Technology Chinese Academy of Sciences)
Federated LearningRecurrent Neural NetworkContrastive LearningTime SeriesSequential
🎯 What it does: For the spatiotemporal prediction task in federated learning, a contrastive learning method based on adaptive semantic alignment (FUELS) is proposed, which achieves personalized model training for clients through dynamic hard negative sample filtering and lightweight prototypes.
Personalized Label Inference Attack in Federated Transfer Learning via Contrastive Meta Learning
Hanyu Zhao (Beijing Institute of Technology), Yu-an Tan (City University of Macau)
Federated LearningSafty and PrivacyAdversarial AttackMeta LearningContrastive LearningImage
🎯 What it does: This paper studies the privacy leakage risks of parameter decoupling strategies in federated transfer learning and proposes a label inference attack based on meta-classifiers and contrastive learning (CML).
Personalized Lip Reading: Adapting to Your Unique Lip Movements with Vision and Language
Jeong Hun Yeo (Korea Advanced Institute of Science and Technology), Yong Man Ro (Korea Advanced Institute of Science and Technology)
RecognitionDomain AdaptationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringVideoTextMultimodality
🎯 What it does: This paper proposes a personalized lip-reading method based on bidirectional adaptation of vision and language, aimed at improving the model's recognition performance for unseen speaker videos.
Personalized Sleep Staging Leveraging Source-free Unsupervised Domain Adaptation
Yangxuan Zhou (Zhejiang University), Gang Pan (Zhejiang University)
Domain AdaptationContrastive LearningTime SeriesBiomedical Data
🎯 What it does: A source-free unsupervised individual domain adaptation (SF-UIDA) framework is proposed for personalized rapid adaptation of sleep staging models for each new subject without accessing source data.
PersonaMagic: Stage-Regulated High-Fidelity Face Customization with Tandem Equilibrium
Xinzhe Li (Ocean University of China), Yong Du (Ocean University of China)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: Proposes the PersonaMagic method, which utilizes staged text control (dynamic embeddings and static supercategory embeddings) and Tandem Equilibrium loss to achieve high-fidelity facial personalization generation from a single portrait, and can seamlessly enhance existing pre-trained personalization models as a plugin.
Perturbating, Tuning, and Collaborating: Harnessing Vision Foundation Models for Single Domain Generalization on Medical Imaging
Chuang Liu (Beihang University), Haogang Zhu (Beihang University)
SegmentationDomain AdaptationTransformerSupervised Fine-TuningBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A collaborative reasoning framework called CollaSU-SDG is proposed, which integrates specialized models and visual foundation models to achieve cross-domain generalization for single-source medical image segmentation.
PFedCS: A Personalized Federated Learning Method for Enhancing Collaboration among Similar Classifiers
Siyuan Wu (Nanjing University), Wanchun Dou (Nanjing University)
Federated LearningKnowledge DistillationImage
🎯 What it does: This paper proposes a personalized federated learning framework PFedCS based on classifier parameter distance for adaptive clustering and distance-constrained aggregation, utilizing similar clients to collaboratively train customized classifiers to enhance local task performance.
pFedES: Generalized Proxy Feature Extractor Sharing for Model Heterogeneous Personalized Federated Learning
Liping Yi (Nankai University), Xiaoxiao Li (University of British Columbia)
ClassificationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: The pFedES framework is proposed, which adds a shared proxy homogeneous feature extractor in front of each heterogeneous client model in federated learning, and achieves bidirectional transfer of global knowledge and local personalized knowledge through alternating training and aggregation.
pFedGPA: Diffusion-based Generative Parameter Aggregation for Personalized Federated Learning
Jiahao Lai (Tsinghua University), Yang Li (Tsinghua University)
GenerationFederated LearningDiffusion modelImage
🎯 What it does: A personalized federated learning aggregation framework pFedGPA based on diffusion models has been developed. By training a diffusion model on the server, generative aggregation of uploaded model parameters is achieved, enabling automatic generation and initialization of personalized model parameters for clients.
Phoneme-Level Feature Discrepancies: A Key to Detecting Sophisticated Speech Deepfakes
Kuiyuan Zhang (Harbin Institute of Technology), Yifang Guo (Alibaba Group)
ClassificationAnomaly DetectionGraph Neural NetworkContrastive LearningAudio
🎯 What it does: The research utilizes adaptive phoneme pooling and graph attention networks to achieve speech deepfake detection by leveraging the inconsistency of phoneme-level features.
PHR-DIFF: Portrait Highlights Removal via Patch-aware Diffusion Model
Hongsheng Zheng (Wuhan University), Chunxia Xiao (Wuhan University)
Image TranslationRestorationDiffusion modelImage
🎯 What it does: A patch-based diffusion model PHR-DIFF is proposed for removing specular highlights in portraits and restoring details.
PhyCamo: A Robust Physical Camouflage via Contrastive Learning for Multi-View Physical Adversarial Attack
Ximin Zhang (Zhejiang University of Technology), Zhenguang Liu (Zhejiang University of Technology)
Adversarial AttackDiffusion modelContrastive LearningImage
🎯 What it does: This paper proposes the PhyCamo framework for multi-view physical adversarial attacks, utilizing diffusion models for data enhancement, contrastive learning to improve robustness, and achieving efficient attacks on encoders.
PhysAug: A Physical-guided and Frequency-based Data Augmentation for Single-Domain Generalized Object Detection
Xiaoran Xu (University of Chinese Academy of Sciences), Jian Liu (Institute of Microelectronics of the Chinese Academy of Sciences)
Object DetectionDomain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A frequency domain data augmentation method called PhysAug is proposed, based on an atmospheric optical physical model, for single-source generalized object detection.
PhysDiff: Physiology-based Dynamicity Disentangled Diffusion Model for Remote Physiological Measurement
Wei Qian (Hefei University of Technology), Meng Wang (Zhejiang University)
TransformerDiffusion modelBiomedical Data
🎯 What it does: A diffusion model combining physiological dynamics separation (PhysDiff) is proposed for remote photoplethysmography (rPPG) signal estimation.
Physical Marker: Revealing Invisible Hyperlinks Hidden in Printed Trademarks
Yuliang Xue (Fudan University), Xinpeng Zhang (Fudan University)
RecognitionImage TranslationData SynthesisTransformerAuto EncoderImage
🎯 What it does: A physical depth hiding scheme for embedding links in the transparent background area of logos is proposed for visually linking printable brand identifiers.
Physical-aware Neural Radiance Fields for Efficient Exposure Correction
Kai Xu (China University of Petroleum), Yan Wang (China University of Petroleum)
RestorationGenerationNeural Radiance FieldImage
🎯 What it does: A physical-aware neural radiance field (PHY-NeRF) is designed to achieve exposure correction for low-light and overexposed scenes through adaptive lighting particles and a global lighting consistency module, and it can generate naturally lit images from multiple viewpoints.
Pilot: Building the Federated Multimodal Instruction Tuning Framework
Baochen Xiong (Institute of Automation, Chinese Academy of Sciences), Changsheng Xu (Institute of Automation, Chinese Academy of Sciences)
Federated LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: The Federated Multimodal Instruction Tuning (FedMIT) framework Pilot is proposed, achieving collaborative fine-tuning of multimodal large language models on distributed clients.
Pinwheel-shaped Convolution and Scale-based Dynamic Loss for Infrared Small Target Detection
Jiangnan Yang (Southwest University of Science and Technology), Xueli Huang (Nanjing University of Science and Technology)
Object DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes two techniques: Pinwheel-shaped Convolution (PConv) and Scale-Based Dynamic Loss (SD Loss) to enhance feature extraction and regression performance for infrared small target detection, and constructs a large-scale single-frame infrared small target dataset SIRST-UAVB.
Pioneer: Physics-informed Riemannian Graph ODE for Entropy-increasing Dynamics
Li Sun (North China Electric Power University), Philip S. Yu (University of Illinois at Chicago)
Graph Neural NetworkGraphTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A physical information-based Riemannian graph ODE model called Pioneer is proposed to simulate dynamic interactive systems of entropy increase and achieve trajectory prediction on Riemannian manifolds.
Pixel Is Not a Barrier: An Effective Evasion Attack for Pixel-Domain Diffusion Models
Chun-Yen Shih (National Taiwan University), Jun-Cheng Chen (National Taiwan University)
GenerationAdversarial AttackDiffusion modelAuto EncoderImageStochastic Differential Equation
🎯 What it does: A framework for adversarial attacks on pixel domain diffusion models (PDM) called AtkPDM has been designed and implemented, utilizing the 2-Wasserstein distance loss of intermediate features from UNet and latent space optimization to enhance the effectiveness and naturalness of the attacks.
PixelMan: Consistent Object Editing with Diffusion Models via Pixel Manipulation and Generation
Liyao Jiang (University of Alberta), Di Niu (Huawei Technologies Canada)
Image TranslationGenerationDiffusion modelImage
🎯 What it does: A non-reverse, non-training Diffusion model consistency object editing method called PixelMan is proposed, which first directly copies the source object to the target position in pixel space, and then iteratively fuses and fills through a three-branch sampling method to maintain overall image consistency.
PIXELS: Progressive Image Xemplar-based Editing with Latent Surgery
Shristi Das Biswas (Purdue University), Kaushik Roy (Amazon Fashion)
Image TranslationGenerationDiffusion modelImageMultimodality
🎯 What it does: The PIXELS framework is proposed, enabling example-based image editing. This method utilizes a pre-trained text-image diffusion model during the inference phase, combined with user-defined pixel/region-level editing maps, to perform progressive editing of the source image with any number of example images, supporting multimodal prompts.
PlaNet: Learning to Mitigate Atmospheric Turbulence in Planetary Images
Yifei Xia (Peking University), Boxin Shi (Peking University)
RestorationData SynthesisConvolutional Neural NetworkImage
🎯 What it does: A model named PlaNet is proposed, capable of simulating turbulence with vertical distance perception on large-scale synthetic planetary images, and utilizing a variable frame convolutional network to remove turbulence from multi-frame planetary images.
PlanLLM: Video Procedure Planning with Refinable Large Language Models
Dejie Yang (Peking University), Yang Liu (Peking University)
SegmentationGenerationTransformerLarge Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes PlanLLM, a cross-modal joint learning framework that utilizes a trainable LLM for video program planning, capable of handling both closed-set single-step classification and open vocabulary free-text planning simultaneously.
Planning from Imagination: Episodic Simulation and Episodic Memory for Vision-and-Language Navigation
Yiyuan Pan (Shanghai Jiao Tong University), Hesheng Wang (Shanghai Jiao Tong University)
GenerationData SynthesisRobotic IntelligenceTransformerVision Language ModelGenerative Adversarial NetworkImageMultimodality
🎯 What it does: This paper proposes the SALImemory (SALI) model, which integrates human situational memory and situational simulation mechanisms into visual language navigation, enabling agents to plan paths in unseen environments by combining imagination and memory.
Planning in the Dark: LLM-Symbolic Planning Pipeline Without Experts
Sukai Huang (University of Melbourne), Trevor Cohn (Google)
TransformerLarge Language ModelText
🎯 What it does: A LLM-symbolic planning pipeline has been constructed that does not require expert intervention, capable of generating various executable action plans and automatically filtering and ranking them.
PLATYPUS: Progressive Local Surface Estimator for Arbitrary-Scale Point Cloud Upsampling
Donghyun Kim (Yonsei University), Seong Jae Hwang (Yonsei University)
GenerationData SynthesisOptimizationPoint Cloud
🎯 What it does: This paper presents the PLATYPUS system, which utilizes an advanced local surface estimator (PLSE) to upsample sparse point clouds, employing curvature-based sampling and a curriculum learning strategy to enhance the reconstruction quality in high-curvature areas.
Plug-and-Play Tri-Branch Invertible Block for Image Rescaling
Jingwei Bao (University of Electronic Science and Technology of China), Shuyuan Zhu (Kuaishou Technology)
RestorationSuper ResolutionCompressionFlow-based ModelImage
🎯 What it does: This paper proposes a three-branch invertible block (T-InvBlock) for image rescaling, which splits the low-frequency branch into luminance and chrominance channels, and uses a zero-high-frequency mapping during upsampling to achieve unified downsampling and upsampling.
PNVC: Towards Practical INR-based Video Compression
Ge Gao (University of Bristol), David Bull (University of Bristol)
CompressionAuto EncoderOptical FlowVideo
🎯 What it does: A practical implicit neural representation (INR) video compression framework PNVC is proposed, which combines the foundation of autoencoders with overfitting techniques, supporting both low-latency and random access coding modes.
POI Recommendation via Multi-Objective Adversarial Imitation Learning
Zhenglin Wan (Chinese University of Hong Kong), Maohao Ran (Hong Kong Baptist University)
Recommendation SystemGraph Neural NetworkReinforcement LearningGenerative Adversarial NetworkGraphSequential
🎯 What it does: A multi-objective adversarial imitation learning framework named MOAIR is proposed for next-step POI recommendation.
POI-Enhancer: An LLM-based Semantic Enhancement Framework for POI Representation Learning
Jiawei Cheng (Beihang University), Xiangyu Zhao (City University of Hong Kong)
Recommendation SystemRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Using large language models (LLM) to extract semantic information such as POI addresses, visitation patterns, and surrounding environments, and injecting the extracted high-quality text semantics into traditional POI representation models through modules like dual feature alignment, semantic fusion, cross-attention fusion, and multi-view contrastive learning, thereby enhancing the expressive capability of POI representations.
Point Cloud Mamba: Point Cloud Learning via State Space Model
Tao Zhang (Wuhan University), Xiangtai Li (Nanyang Technological University)
RecognitionSegmentationTransformerPoint Cloud
🎯 What it does: A point cloud learning framework called Point Cloud Mamba is proposed, which utilizes consistent traversal serialization, sequential prompts, and spatial coordinate mapping position encoding to achieve global modeling of point clouds.
Point Cloud Semantic Segmentation with Sparse and Inhomogeneous Annotations
Zhiyi Pan (Peking University), Ge Li (Tencent)
SegmentationPoint Cloud
🎯 What it does: In the weakly supervised point cloud semantic segmentation task, the AADNet model is proposed, which can adapt to any sparse annotation distribution, addressing the bias issue in gradient estimation caused by non-uniform sparse annotations.
PointCFormer: A Relation-Based Progressive Feature Extraction Network for Point Cloud Completion
Yi Zhong (National University of Defense Technology), Yingmei Wei (National University of Defense Technology)
TransformerPoint Cloud
🎯 What it does: A Point Cloud Completion Network based on Transformer, PointCFormer, is designed to achieve global structure preservation and local detail capture by utilizing relationship-aware local features and progressive feature extraction.
PointDGMamba: Domain Generalization of Point Cloud Classification via Generalized State Space Model
Hao Yang (Shanghai Jiao Tong University), Shuicheng Yan (Skywork AI)
ClassificationDomain AdaptationPoint CloudBenchmark
🎯 What it does: A PointDGMamba framework based on the State Space Model (SSM) is proposed for domain generalization in point cloud classification, with core modules including Masked Sequence Denoising, Sequence-wise Cross-domain Feature Aggregation, and Dual-level Domain Scanning.
PointRWKV: Efficient RWKV-Like Model for Hierarchical Point Cloud Learning
Qingdong He (Youtu Lab, Tencent), Chengjie Wang (Tencent)
ClassificationSegmentationGraph Neural NetworkTransformerPoint Cloud
🎯 What it does: A linear complexity point cloud learning model called PointRWKV based on RWKV is proposed, along with a multi-scale hierarchical structure and parallel local/global feature aggregation.
PointTalk: Audio-Driven Dynamic Lip Point Cloud for 3D Gaussian-based Talking Head Synthesis
Yifan Xie (Xi'an Jiaotong University), Fei Richard Yu (Shenzhen University)
GenerationData SynthesisContrastive LearningGaussian SplattingVideoPoint CloudAudio
🎯 What it does: This paper presents PointTalk, which utilizes a 3D Gaussian field and audio-driven dynamic lip point clouds to achieve high-quality, audio-synchronized talking head synthesis.
PokerBench: Training Large Language Models to Become Professional Poker Players
Richard Zhuang (University of California Berkeley), Gopala Anumanchipalli (University of California Berkeley)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkFinance Related
🎯 What it does: This paper introduces a benchmark called POKERBENCH, designed to evaluate the decision-making capabilities of LLMs in Texas No-Limit Hold'em poker, and fine-tunes various LLMs on this benchmark to enhance poker performance.
Polarization Guided Mask-Free Shadow Removal
Chu Zhou (National Institute of Informatics), Boxin Shi (Peking University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: A shadow removal method based on polarization information, PolShaRe, is proposed, which can restore images without the need for external shadow masks.
Political Actor Agent: Simulating Legislative System for Roll Call Votes Prediction with Large Language Models
Hao Li (Wuhan University), Hao Jiang (Wuhan University)
TransformerLarge Language ModelAgentic AIText
🎯 What it does: The role-playing political actor agent (PAA) built using large language models predicts the voting outcomes of U.S. House bills through scalable member profiles, multi-perspective planning, and influence mechanisms, providing interpretable decision-making bases.
Poplar: Efficient Scaling of Distributed DNN Training on Heterogeneous GPU Clusters
WenZheng Zhang (Peking University), Xiaoying Bai (Advanced Institute of Big Data)
TransformerLarge Language ModelText
🎯 What it does: Designed and implemented Poplar, a ZeRO-extended distributed training system that supports heterogeneous GPU clusters.
POPoS: Improving Efficient and Robust Facial Landmark Detection with Parallel Optimal Position Search
Chong-Yang Xiang (Southwest Jiaotong University), Xian-Sheng Hua (Zhejiang University)
RecognitionPose EstimationComputational EfficiencyConvolutional Neural NetworkImageBenchmark
🎯 What it does: A parallel optimal position search (POPoS) framework is proposed to enhance the accuracy and efficiency of facial keypoint detection.
Population Aware Diffusion for Time Series Generation
Yang Li (William and Mary), Haipeng Chen (William and Mary)
GenerationData SynthesisTransformerDiffusion modelTime SeriesFinance Related
🎯 What it does: A time series generation framework based on diffusion models, PaD-TS, is proposed, focusing on preserving the overall properties of the original dataset (such as value distribution and functional dependency distribution).
Portcullis: A Scalable and Verifiable Privacy Gateway for Third-Party LLM Inference
Jiangou Zhan (Tsinghua University), Ye Wu (Bytedance Inc)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: This paper presents Portcullis—a trusted, verifiable, and scalable privacy gateway that utilizes Intel TDX to ensure the confidentiality of input prompts during third-party LLM inference, while achieving desensitization and recovery of sensitive information without compromising output quality.
Pose as a Modality: A Psychology-Inspired Network for Personality Recognition with a New Multimodal Dataset
Bin Tang (East China Normal University), Shu-Guang Kuai (East China Normal University)
RecognitionPose EstimationVideoTextMultimodalityAudio
🎯 What it does: This paper constructs a multimodal personality dataset that includes full-body posture, facial expressions, frames, audio, and text, and proposes the PINet model to predict the five major dimensions of personality.
Pose Magic: Efficient and Temporally Consistent Human Pose Estimation with a Hybrid Mamba-GCN Network
Xinyi Zhang (Tsinghua University), Qingmin Liao (Tsinghua University)
Pose EstimationGraph Neural NetworkVideo
🎯 What it does: Proposes the Pose Magic architecture, combining Mamba and GCN to achieve attention-free 3D human pose estimation;
PoseLLaVA: Pose Centric Multimodal LLM for Fine-Grained 3D Pose Manipulation
Dong Feng (Intel Labs), Peng Wang (Nanjing University of Posts and Telecommunications)
Pose EstimationTransformerLarge Language ModelDiffusion modelContrastive LearningMultimodality
🎯 What it does: Design PoseLLaVA, which combines SMPL pose representation with a multimodal large language model to achieve language-based pose estimation, generation, and adjustment;
PoseMamba: Monocular 3D Human Pose Estimation with Bidirectional Global-Local Spatio-Temporal State Space Model
Yunlong Huang (Huazhong University of Science and Technology), Robert Caiming Qiu (Huazhong University of Science and Technology)
Pose EstimationConvolutional Neural NetworkVideo
🎯 What it does: A pure state space model-based 3D human pose estimation framework called PoseMamba is proposed, utilizing bidirectional global-local spatiotemporal scanning to capture skeletal relationships and temporal dependencies.
Position-Aware Guided Point Cloud Completion with CLIP Model
Feng Zhou (North China University of Technology), Junliang Xing (Tsinghua University)
RestorationGenerationData SynthesisTransformerVision Language ModelMultimodalityPoint Cloud
🎯 What it does: Rapidly expand the single-modal point cloud completion framework into a multi-modal framework and propose a position-aware module to enhance the localization accuracy of missing parts.
Power of Diversity: Enhancing Data-Free Black-Box Attack with Domain-Augmented Learning
Yang Wei (Chongqing University of Posts and Telecommunications), Bin Xiao (Jinan Inspur Data Technology Co., Ltd.)
Data SynthesisOptimizationAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Proposes a Domain-Augmented Learning method for data unsupervised black-box attacks, utilizing three main modules: adaptive semantic embedding, competition optimization, and heterogeneity excitation to enhance the diversity of synthetic data generated by the generator, thereby improving the training effectiveness of the substitute model and the success rate of the attack.
PowerMLP: An Efficient Version of KAN
Ruichen Qiu (University of Chinese Academy of Sciences), Xiao-Shan Gao (University of Chinese Academy of Sciences)
Computational EfficiencyImageText
🎯 What it does: PowerMLP is proposed, a ReLU power-based MLP network that replaces the spline activation function in KAN with a non-iterative approach, significantly improving training and inference speed while maintaining or exceeding the expressive power of KAN.
Practicable Black-Box Evasion Attacks on Link Prediction in Dynamic Graphs—a Graph Sequential Embedding Method
Jiate Li (Nanchang University), Binghui Wang (Illinois Institute of Technology)
Adversarial AttackRecurrent Neural NetworkGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A practical black-box evasion attack method for dynamic graph link prediction models is proposed, which can significantly reduce the prediction performance of the target model with only limited interactions and perturbations.
Practical Offloading for Fine-Tuning LLM on Commodity GPU via Learned Sparse Projectors
Siyuan Chen (Carnegie Mellon University), Phillip B. Gibbons (Carnegie Mellon University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a framework called LSPOffload for large-scale language model fine-tuning on consumer-grade GPUs. It utilizes learned sparse projectors to achieve high-dimensional subspace updates within limited GPU memory, thereby avoiding memory overflow while maintaining performance close to the original model.
PRAGA: Prototype-aware Graph Adaptive Aggregation for Spatial Multi-modal Omics Analysis
Xinlei Huang (Great Bay University), Ziyue Qiao (Great Bay University)
Graph Neural NetworkContrastive LearningMultimodalityBiomedical Data
🎯 What it does: The PRAGA framework is proposed, which uses a dynamically learnable omics-specific graph and dynamic prototype contrastive learning to achieve unified representation of multimodal spatial omics data.
Pragmatist: Multiview Conditional Diffusion Models for High-Fidelity 3D Reconstruction from Unposed Sparse Views
Songchun Zhang (Zhejiang University), Chunhui Zhao (Zhejiang University)
GenerationPose EstimationDiffusion modelNeural Radiance FieldPoint CloudMesh
🎯 What it does: This paper proposes a Pragmatist pipeline that transforms the problem of 3D reconstruction from pose-free sparse views into conditional new view synthesis, completing the reconstruction by generating complete observations and subsequently recovering view poses and refining textures.
Pre-Assignment Problem for Unique Minimum Vertex Cover on Bounded Clique-Width Graphs
Shinwoo An (POSTECH), Hyeonjun Shin (POSTECH)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies the pre-allocation problem of the unique minimum vertex cover on bounded clique-width graphs (PAU-VC), proposing a polynomial-time algorithm that solves the PAU-VC problem on trees and extends it to the class of bounded clique-width graphs.
Pre-Trained Vision-Language Models as Noisy Partial Annotators
Qian-Wei Wang (Tsinghua University), Shu-Tao Xia (Tsinghua University)
ClassificationKnowledge DistillationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: This paper studies the use of the pre-trained vision-language model CLIP to automatically generate partial labels for downstream image classification tasks and proposes the Co-Reg method to train dedicated small models in the noise partial label learning (NPLL) scenario.
Pre-training a Density-Aware Pose Transformer for Robust LiDAR-based 3D Human Pose Estimation
Xiaoqi An (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
Pose EstimationTransformerPoint Cloud
🎯 What it does: This study proposes a Density-Aware Pose Transformer (DAPT) and a comprehensive LiDAR human synthesis and occlusion enhancement scheme based on SMPL ray casting, achieving robust 3D human pose estimation from a single frame of LiDAR point cloud.
Pre-Training Graph Neural Networks on Molecules by Using Subgraph-Conditioned Graph Information Bottleneck
Van Thuy Hoang (Catholic University of Korea), O-Joun Lee (Catholic University of Korea)
Representation LearningDrug DiscoveryGraph Neural NetworkContrastive LearningGraphBiomedical Data
🎯 What it does: A self-supervised pre-training method S-CGIB is proposed, which can automatically identify molecular core subgraphs and important subgraphs without the need for labels or prior knowledge of functional groups, and use them to generate molecular-level representations.
Precision-Enhanced Human-Object Contact Detection via Depth-Aware Perspective Interaction and Object Texture Restoration
Yuxiao Wang (South China University of Technology), Qi Liu (South China University of Technology)
Object DetectionSegmentationDepth EstimationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A PIHOT framework based on depth perception and perspective interaction has been developed for accurately detecting human-object contact areas.
Predicting the Original Appearance of Damaged Historical Documents
Zhenhua Yang (South China University of Technology), Lianwen Jin (South China University of Technology)
Image TranslationRestorationDiffusion modelImage
🎯 What it does: This paper proposes the Historical Document Restoration (HDR) task, constructs the HDR28K large-scale dataset, and introduces the DiffHDR diffusion network to achieve high-quality restoration of damaged images.
Predicting User Behavior in Smart Spaces with LLM-Enhanced Logs and Personalized Prompts
Yunpeng Song (Xi'an Jiaotong University), Zhongmin Cai (Xi'an Jiaotong University)
Recommendation SystemAutonomous DrivingGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextSequential
🎯 What it does: A model is proposed that utilizes LLM to enhance logs and personalized prompts to predict user behavior in smart spaces (such as smart cars and smart homes).
Prediction-Based Adaptive Variable Ordering Heuristics for Constraint Satisfaction Problems
Jitao Xu (Northeast Normal University), Minghao Yin (Northeast Normal University)
OptimizationRecurrent Neural NetworkSupervised Fine-TuningTabular
🎯 What it does: This paper proposes a prediction-based adaptive variable ordering heuristic (PBAVOH), which encodes the topology of binary search trees into NPD sequences and uses LSTM and GBDT models to predict the total failure count of the search tree, thereby selecting the optimal variable ordering heuristic for each CSP instance before solving.
Prediction-Feedback DETR for Temporal Action Detection
Jihwan Kim (Sungkyunkwan University), Jae-Pil Heo (Sungkyunkwan University)
RecognitionObject DetectionTransformerVideo
🎯 What it does: This paper proposes Prediction-Feedback DETR (Pred-DETR), which alleviates the collapse of cross-attention by guiding cross-attention and self-attention through predicted results.
Preference-Oriented Supervised Fine-Tuning: Favoring Target Model over Aligned Large Language Models
Yuchen Fan (Zuoyebang Education Technology), Yang Song (Zuoyebang Education Technology)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A supervised fine-tuning method based on the Bradley–Terry preference model, called PoFT, is proposed to encourage the target model to outperform the aligned LLM on the same data.
PriFold: Biological Priors Improve RNA Secondary Structure Predictions
Chenchen Yang (Fudan University), Siqi Sun (Zelixir Biotech)
Protein Structure PredictionTransformerSupervised Fine-TuningBiomedical Data
🎯 What it does: This paper proposes PriFold, which improves RNA secondary structure prediction using biological priors.
Prior-Constrained Association Learning for Fine-Grained Generalized Category Discovery
Menglin Wang (Nanjing Normal University), Xiaojin Gong (Zhejiang University)
ClassificationRecognitionTransformerContrastive LearningImage
🎯 What it does: Proposes a prior-constrained greedy association learning combined with non-parametric prototype comparison and parametric classification to achieve fine-grained general category discovery.
Prior-guided Hierarchical Harmonization Network for Efficient Image Dehazing
Xiongfei Su (Zhejiang University), Xin Yuan (Tsinghua University)
Image HarmonizationRestorationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a hierarchical harmonious network PGH-Net based on the bright-dark channel prior and histogram equalization, achieving efficient single image dehazing.
Privacy-and-Utility-Aware Publishing of Schedules
Maike Basmer (Humboldt-Universitat zu Berlin), Matthias Weidlich (Humboldt-Universitat zu Berlin)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper studies the privacy leakage risks when releasing scheduling plans in parallel machine scheduling and proposes a mechanism for releasing schedules under the premise of meeting privacy and utility thresholds.
Privacy-Preserving Low-Rank Adaptation Against Membership Inference Attacks for Latent Diffusion Models
Zihao Luo (University of Auckland), Jingfeng Zhang (University of Auckland)
GenerationOptimizationSafty and PrivacyDiffusion modelImage
🎯 What it does: Under privacy leakage attacks, a low-rank adaptation for privacy (MP-LoRA) and a stable version (SMP-LoRA) are proposed for secure fine-tuning on the Latent Diffusion Model, reducing the success rate of membership inference attacks.
Privacy-Preserving V2X Collaborative Perception Integrating Unknown Collaborators
Bin Lu (Zhejiang University), Eryun Liu (Zhejiang University)
Object DetectionDomain AdaptationAutonomous DrivingSafty and PrivacyConvolutional Neural NetworkPoint Cloud
🎯 What it does: A privacy-preserving V2X collaborative perception framework is proposed, allowing vehicles to achieve collaboration through feature fusion and domain adaptation after independent training.
Private Blotto: Viewpoint Competition with Polarized Agents
Kate Donahue (Cornell University), Jon Kleinberg (Cornell University)
🎯 What it does: This paper proposes and analyzes the 'Private Blotto' game, studying how decentralized agents allocate limited effort across multiple projects and form pure Nash equilibria under two different aggregation functions (median and mean).
PrivDNFIS: Privacy-preserving and Efficient Deep Neuro-Fuzzy Inference System
Hao Ren (Sichuan University), Xingshu Chen (Sichuan University)
Safty and PrivacyComputational EfficiencyReinforcement LearningImage
🎯 What it does: Developed PrivDNFIS, a privacy-preserving deep neural fuzzy inference system based on lattice-based homomorphic encryption;
Proactive and Reactive Constraint Programming for Stochastic Project Scheduling with Maximal Time-Lags
Kim van den Houten (Delft University of Technology), Mathijs de Weerdt (Delft University of Technology)
OptimizationTabularBenchmark
🎯 What it does: Three new scheduling methods for SRCPSP/max are proposed (proactive SAA based on CP, complete rearrangement reactive method, and POS method using STNU), and they are systematically evaluated through statistical benchmarks.
Probabilistic Explanations for Linear Models
Bernardo Subercaseaux (Pontificia Universidad Catolica de Chile), Kuldeep S. Meel (Georgia Institute of Technology)
Explainability and InterpretabilityComputational Efficiency
🎯 What it does: This paper studies the efficient solution of probabilistic sufficient reasons (pδ,ε‑min‑SR) on linear models, providing theoretical proofs and a Monte Carlo sampling-based algorithm.
Probabilistic Forecasting of Irregularly Sampled Time Series with Missing Values via Conditional Normalizing Flows
Vijaya Krishna Yalavarthi (University of Hildesheim), Lars Schmidt-Thieme (University of Hildesheim)
Flow-based ModelTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: A model called ProFITi based on conditional normalizing flows is proposed for probabilistic prediction of missing values in irregularly sampled multivariate time series, capable of learning the multivariate joint distribution under query conditions.
Probabilistic Shielding for Safe Reinforcement Learning
Edwin Hamel-De le Court (Imperial College London), Alexander W. Goodall (Imperial College London)
Safty and PrivacyReinforcement Learning
🎯 What it does: A scalable safe reinforcement learning method based on state expansion and probabilistic shields is proposed, ensuring that probabilistic safety constraints are met during both training and testing phases.
Probabilistic Strategy Logic with Degrees of Observability
Chunyan Mu (University of Aberdeen), Brian Logan (University of Aberdeen)
🎯 What it does: This paper proposes a probabilistic strategy logic extension called oPSL, which introduces observability and observability operators to quantify and verify information transparency and concealment in multi-agent systems.
Probability-Density-aware Semi-supervised Learning
Shuyang Liu (East China Normal University), Shaohui Lin (East China Normal University)
ClassificationSegmentationContrastive LearningImage
🎯 What it does: A probability density-aware semi-supervised learning method is proposed, which improves similarity measurement using density information and constructs a new label propagation algorithm called PMLP.
Procedure Knowledge Decoupled Distillation Strategy for Procedure Planning in Instructional Videos
Xiaotian Pan (Harbin Institute of Technology), Weigang Zhang (Harbin Institute of Technology)
Knowledge DistillationTransformerDiffusion modelVideo
🎯 What it does: By allowing the teacher model to obtain real intermediate visual information and transferring its distributed knowledge to the student model through decoupled knowledge distillation (single action distillation and sequence distribution distillation), the program planning performance in teaching videos is improved.
ProcTag: Process Tagging for Assessing the Efficacy of Document Instruction Data
Yufan Shen (Zhejiang University), Cong Yao (Alibaba Group)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: The ProcTag method is proposed to evaluate the effectiveness of instruction data by labeling the execution process of document instructions with pseudocode, and DocLayPrompt is introduced to achieve a semi-structured layout-aware document representation.
Progressive Distribution Matching for Federated Semi-Supervised Learning
Dongping Liao (University of Macau), Cheng-Zhong Xu (DataStory Information Technology Co., Ltd)
Federated LearningImage
🎯 What it does: This paper proposes an advanced distribution matching method for federated semi-supervised learning called FedPDM.
Progressive Multi-granular Alignments for Grounded Reasoning in Large Vision-Language Models
Quang-Hung Le (Deakin University), Thao Minh Le (Deakin University)
RecognitionObject DetectionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: Proposes the PromViL framework, which employs hierarchical multi-granularity visual-language alignment and progressive reasoning to enhance the performance of large-scale visual-language models in complex visual-language reasoning tasks.
Progressive Self-Learning for Domain Adaptation on Symbolic Regression of Integer Sequences
Yaohui Zhu (Beijing University of Chemical Technology), Lingfeng Wang (Beijing University of Chemical Technology)
Domain AdaptationTransformerSequential
🎯 What it does: This paper proposes an advanced self-learning (PSL) framework that constructs source domain data using the initial terms of the target sequence and continuously generates and verifies formulas through a cyclic self-learning loop, thereby achieving domain adaptation for integer sequence symbolic regression.
Promising Multi-Granularity Linguistic Steganography by Jointing Syntactic and Lexical Manipulations
Chengfu Ou (Changsha University of Science and Technology), Yangfan Liu (Changsha University of Science and Technology)
GenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: A multi-granularity modified language steganography framework MMLS is proposed, which embeds the key into the syntactic space and symbolic space using both syntactic transformation and lexical replacement.
Promoting Knowledge Base Question Answering by Directing LLMs to Generate Task-relevant Logical Forms
Jianqi Gao (Shanghai Jiao Tong University), Hang Yu (Shanghai University)
GenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the FM-KBQA framework, which utilizes multi-task learning to simultaneously train the generation of logical forms (LF) and predict answer-related reasoning path indices on large language models (LLM), significantly improving the accuracy of knowledge base question answering (KBQA).
Prompt Compression with Context-Aware Sentence Encoding for Fast and Improved LLM Inference
Barys Liskavets (Alterra AI), Shane K. Luke (Workday Inc.)
CompressionTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: A sentence-level context-aware prompt compression (CPC) method is proposed, which uses a sentence encoder to evaluate the relevance of each sentence to the question, thereby trimming irrelevant sentences to shorten the input prompt.