AAAI Conference on Artificial Intelligence · 2140 papers
Patho-AgenticRAG: Towards Multimodal Agentic Retrieval-Augmented Generation for Pathology VLMs via Reinforcement Learning
Wenchuan Zhang (Sichuan University), Hong Bu (Sichuan University)
CodeGenerationRetrievalReinforcement LearningAgentic AIVision Language ModelMultimodalityBiomedical DataRetrieval-Augmented Generation
🎯 What it does: This work proposes a multimodal retrieval-augmented generation framework called Patho-AgenticRAG for pathological vision-language models.
Patho-R1: A Multimodal Reinforcement Learning-Based Pathology Expert Reasoner
Wenchuan Zhang (Sichuan University), Hong Bu (Sichuan University)
CodeData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelContrastive LearningImageTextMultimodalityBiomedical DataChain-of-Thought
🎯 What it does: Constructed a high-quality multimodal dataset centered on pathology textbooks, and trained two pathology-specific vision-language models, Patho-CLIP and Patho-R1;
🎯 What it does: This paper proposes the PathRAG method, which enhances the generation quality of large language models by retrieving key relationship paths from graph indexes and converting them into text prompts.
Paths Not Taken: Structure-Based Pruning in PSDD Learning and Inference
Cory Butz (University of Regina), Camilla Lewis (University of Regina)
CodeComputational EfficiencyTextGraphBenchmark
🎯 What it does: This paper proposes a new method that utilizes the determinism of PSDD for parameter learning, structured pruning inference, and parallel circuit evaluation.
PC-CrossDiff: Point-Cluster Dual-Level Cross-Modal Differential Attention for Unified 3D Referring and Segmentation
Wenbin Tan (Xiamen University), Yanyun Qu (Xiamen University)
CodeSegmentationTransformerVision Language ModelPoint Cloud
🎯 What it does: Propose a unified two-layer cross-modal differential attention framework named PC-CrossDiff, which can simultaneously accomplish 3D object localization (3DREC) and segmentation (3DRES) tasks.
PCFormer: Accelerating Privacy-preserving Transformer Inference by Partition and Combination
Bo Zeng (Wuhan University), Run Wang (Wuhan University)
CodeSafty and PrivacyComputational EfficiencyTransformerTextBenchmark
🎯 What it does: Propose the PCFormer framework, which partitions and merges nonlinear redundancies in Transformers during privacy inference to reduce computational and communication costs in HE/MPC.
PCoKG: Personality-aware Commonsense Reasoning with Debate
Weijie Li (Soochow University), Guodong Zhou (Soochow University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: Constructed a Personality-aware Commonsense Knowledge Graph (PCoKG), systematically organizing quadruples composed of events, MBTI personality types, reasoning dimensions, and reasoning outcomes, while enhancing knowledge quality through a multi-round LLM debate mechanism; simultaneously verifying its effectiveness in personalized dialogue generation tasks.
PEFT-BoA: Parameter-Efficient Fine-Tuning with Bag-of-Adapters for Multi-Modal Object Re-identification
Hongchao Li, YongLong Luo (Anhui Normal University)
CodeRecognitionTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Proposes a parameter-efficient fine-tuning framework PEFT-BoA based on CLIP, achieving multi-modal object ReID through three lightweight adapters.
PEGNet: A Physics-Embedded Graph Network for Long-Term Stable Multiphysics Simulation
Can Yang (Xiamen University), Min Jiang (Xiamen University)
CodeGraph Neural NetworkMeshGraphBiomedical DataPhysics Related
🎯 What it does: Proposed a physics-embedded graph network called PEGNet for long-term stable multi-physics simulations, primarily addressing the issues of error accumulation and insufficient physical consistency in traditional data-driven models.
PEOD: A Pixel-Aligned Event-RGB Benchmark for Object Detection Under Challenging Conditions
Luoping Cui (Beijing University of Posts and Telecommunications), Chuang Zhu (Beijing University of Posts and Telecommunications)
CodeObject DetectionMultimodalityBenchmark
🎯 What it does: This paper proposes PEOD, a 1280×720 pixel-aligned event-RGB dataset designed for object detection under harsh conditions such as extreme lighting and high speeds.
PepCCD: A Contrastive Conditioned Diffusion Framework for Target-Specific Peptide Generation
Jun Zhang (Shenzhen University), Zexuan Zhu (Shenzhen University)
CodeGenerationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningDiffusion modelContrastive LearningBiomedical Data
🎯 What it does: Propose a contrastive learning conditional diffusion framework (PepCCD) that can generate targeted specific peptides based solely on the target protein sequence, eliminating the need for structural information.
Perceive, Act and Correct: Confidence Is Not Enough for Hyperspectral Classification
Muzhou Yang (Nanjing University of Aeronautics and Astronautics), Mingqiang Wei (Nanjing University of Aeronautics and Astronautics)
CodeClassificationImage
🎯 What it does: Propose the CABIN framework, which enhances hyperspectral image classification under low annotation scenarios through a perception-action-correction loop.
🎯 What it does: Proposed the VeteranAD framework, adopting the 'perception-in-plan' paradigm, embedding the perception module into the planning process, utilizing multi-modal anchor trajectories as planning priors, performing localization-aware and autoregressive trajectory planning, and constructing an end-to-end autonomous driving system.
🎯 What it does: Constructed a subjective quality assessment dataset specifically for 3D Gaussian Splatting (3DGS) called 3DGS-QA, and proposed a no-reference quality prediction model named GSOQA;
Personality-guided Public-Private Domain Disentangled Hypergraph-Former Network for Multimodal Depression Detection
Changzeng Fu (Northeastern University), Chaoran Liu (Northeastern University)
CodeClassificationGraph Neural NetworkTransformerLarge Language ModelGenerative Adversarial NetworkContrastive LearningMultimodalityTime Series
🎯 What it does: Proposed a Hypergraph Former network (P3HF) based on personalized guidance and public-private domain decoupling for multimodal depression detection.
🎯 What it does: This paper proposes an untrained differential transformer (DiT) personalized image generation framework called 'Personalize Anything,' achieving high-fidelity subject reconstruction while accommodating editing, layout, and multi-subject synthesis through resolution-agnostic token replacement at specific positions during the reverse process.
Personalize Before Retrieve: LLM-based Personalized Query Expansion for User-Centric Retrieval
Yingyi Zhang (Dalian University of Technology), Xiangyu Zhao (City University of Hong Kong)
CodeRetrievalTransformerLarge Language ModelTextGraph
🎯 What it does: This paper proposes the PBR (Personalize-Before-Retrieve) framework, which enhances user-centered retrieval effectiveness by generating personalized query expansions through LLM before retrieval.
🎯 What it does: Proposes PERFECT, a personalized federated graph-level clustering framework for multi-client structural heterogeneity, which utilizes privacy-preserving representative samples to enable cross-client information sharing and enhances local clustering performance through clustering gradient optimization.
Perspective from a Broader Context: Can Room Style Knowledge Help Visual Floorplan Localization?
Bolei Chen (Central South University), Jianxin Wang (Central South University)
CodeConvolutional Neural NetworkContrastive LearningSimultaneous Localization and MappingImagePoint Cloud
🎯 What it does: Propose an unsupervised room style discriminator trained with clustering constraints, leveraging room style information to enhance visual floor plan localization and alleviate localization uncertainty caused by repetitive structural layouts.
PerTouch: VLM-Driven Agent for Personalized and Semantic Image Retouching
Zewei Chang (Nankai University), Chongyi Li (Samsung Electronics)
CodeRestorationSegmentationAgentic AIVision Language ModelDiffusion modelImage
🎯 What it does: Proposes a unified diffusion model framework called PerTouch, which can achieve semantic-level fine-grained image retouching while maintaining aesthetic consistency globally.
🎯 What it does: The study uses perturbed best response (PBRO) in zero-sum games to improve the iterative complexity of Fictitious Play (FP) and Double Oracle (DO) algorithms, and proposes corresponding randomized versions: Stochastic Fictitious Play (SFP) and Stochastic Double Oracle (SDO).
Perturbing to Preserve: Defending Fragile Knowledge in Online Continual Learning
Dulan Zhou (National University of Defense Technology), Kele Xu (National University of Defense Technology)
CodeImage
🎯 What it does: Propose the PDFK framework, combining exponential moving average (EMA) smoothing and structured perturbation consistency regularization to address knowledge fragility in online continual learning.
🎯 What it does: This paper systematically evaluates the robustness of Vision-Language-Action (VLA) models against physical sensor attacks (such as laser, light projection, EMI, ultrasound, etc.), and proposes an automated 'Real-Sim-Real' framework to simulate these attacks in simulation and verify them on real robots; subsequently, adversarial training is applied to enhance the model's defensive performance.
Physical-regularized Hierarchical Generative Model for Metallic Glass Structural Generation and Energy Prediction
Qiyuan Chen (University of Wisconsin - Madison), Bu Wang (Stanford University)
CodeGenerationData SynthesisGraph Neural NetworkAuto EncoderGraphPhysics Related
🎯 What it does: A physics-regularized hierarchical graph variational autoencoder, GLASSVAE, was studied for structure generation and energy prediction in metallic glasses.
🎯 What it does: Propose a mesh-free soft policy iteration framework based on physics-informed neural networks (PINN) for solving the Hamilton-Jacobi-Bellman equation in entropy-regularized stochastic control problems.
Physics-Informed Deformable Gaussian Splatting: Towards Unified Constitutive Laws for Time-Evolving Material Field
Haoqin Hong (University of Science and Technology of China), Jingrun Chen (University of Science and Technology of China)
CodeGenerationGaussian SplattingOptical FlowVideoPhysics Related
🎯 What it does: Proposed the Physics-Informed Deformable Gaussian Expansion (PIDG) model, which treats 3D Gaussian particles as physical material points from a Lagrangian perspective, and reconstructs monocular dynamic scenes by jointly evolving time-varying material fields.
🎯 What it does: Propose a framework called Koopman-PINN, which combines ART nonparametric volatility filtering, temporal matching initialization, and neural network learning of the Koopman operator to estimate five parameters of the Heston model using high-frequency price data.
PhysicsCorrect: A Training-Free Approach for Stable Neural PDE Simulations
Xinquan Huang (University of Pennsylvania), Paris Perdikaris (University of Pennsylvania)
CodeComputational EfficiencyConvolutional Neural NetworkTransformerMeshTime SeriesPhysics Related
🎯 What it does: Proposes PhysicsCorrect, a physics-consistent correction framework that does not require retraining, to stabilize error accumulation in neural PDE solvers during long-time sequence simulations.
CodeAutonomous DrivingAdversarial AttackTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A physically realizable adversarial patch attack framework called PhysPatch is proposed for autonomous driving systems driven by multimodal large language models (MLLM), which exhibits good cross-model transferability.
Picking a Representative Set of Solutions in Multiobjective Optimization: Axioms, Algorithms, and Experiments
Niclas Boehmer (Hasso Plattner Institute, University of Potsdam), Maximilian T. Wittmann (Hasso Plattner Institute, University of Potsdam)
CodeOptimizationBenchmark
🎯 What it does: The paper studies the Pareto pruning problem in multi-objective optimization, systematically evaluates existing representative quality metrics, and proposes a new metric called directed coverage, combining axiomatic analysis, complexity proof, and experimental evaluation.
🎯 What it does: This paper proposes the PipeDiT framework, which significantly accelerates DiT-based video generation by leveraging task pipelining and module decoupling.
PKR-QA: A Benchmark for Procedural Knowledge Reasoning with Knowledge Module Learning
Thanh-Son Nguyen (Agency for Science, Technology and Research), Basura Fernando (Agency for Science, Technology and Research)
CodeExplainability and InterpretabilityLarge Language ModelVision Language ModelVision-Language-Action ModelContrastive LearningVideoTextMultimodalityGraphBenchmark
🎯 What it does: Constructed the PKR-QA question-answering benchmark and the corresponding Program Knowledge Graph (PKG), and proposed the Knowledge Module Learning (KML) neuro-symbolic framework, which uses learnable knowledge modules to execute reasoning programs generated by LLMs, achieving interpretable reasoning on procedural tasks.
🎯 What it does: Propose PlanB&B, a model-based reinforcement learning agent that improves MILP variable selection strategies by learning a Branch-and-Bound (B&B) dynamic model and performing planning.
Francesco Percassi (University of Huddersfield), Enrico Scala (University of Brescia)
CodeOptimizationBenchmark
🎯 What it does: Propose the PUMA (Planning with Uncertain Models of Actions) framework, where action models are revealed after the first execution and become deterministic in subsequent executions, and present two polynomial compilation methods: one compiled into FOND (COMP2FOND), and the other compiled into classical planning (COMP2FOD).
PLaST: Towards Paralinguistic-aware Speech Translation
Yi Li (Xiamen University), Yidong Chen (Huawei Translation Services Center)
CodeTransformerLarge Language ModelTextMultimodalityAudio
🎯 What it does: Proposed a dual-branch end-to-end speech translation framework called PLaST, specifically designed to simultaneously capture linguistic content in speech and non-linguistic cues such as emotion, stress, and integrate them into large language models (LLMs) for translation;
Plug-and-Play Clarifier: A Zero-Shot Multimodal Framework for Egocentric Intent Disambiguation
Sicheng Yang, Zhensong Zhang (Imperial College London)
CodeObject DetectionSegmentationDepth EstimationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Propose Plug-and-Play Clarifier, a zero-shot, modular framework that progressively eliminates modality ambiguity in front-view interactions through three clarification modules: text, visual, and cross-modal, thereby improving intent understanding in egocentric AI agents.
Plug-and-Play Optimization for 3D Gaussian Splatting Compression: Distribution Regularization, Probabilistic Pruning and Detail Compensation
Tian Bai (University of Science and Technology of China), Ziyang Dai (University of Science and Technology of China)
CodeCompressionGaussian Splatting
🎯 What it does: This paper proposes a three-module plug-and-play optimization framework (distribution regularization, probabilistic pruning, and high-frequency compensation) to significantly reduce the storage and computational costs of 3D Gaussian Splatting (3DGS) models, while seamlessly integrating into mainstream structured compression methods.
🎯 What it does: Propose a pluggable federated recommendation framework that adopts parameter-efficient fine-tuning (PEFT) to update only compressed item embeddings on the client side, thereby significantly reducing communication overhead;
PlugTrack: Multi-Perceptive Motion Analysis for Adaptive Fusion in Multi-Object Tracking
Seungjae Kim (Kyung Hee University), MyeongAh Cho (Kyung Hee University)
CodeObject TrackingRecurrent Neural NetworkVideo
🎯 What it does: Propose the PlugTrack framework, which achieves adaptive fusion of Kalman filters and data-driven motion predictors through multi-sensory motion analysis;
PMGS: Reconstruction of Projectile Motion Across Large Spatiotemporal Spans via 3D Gaussian Splatting
Yijun Xu (Wuhan University), Chu He (Chongqing University)
CodePose EstimationGaussian SplattingOptical FlowVideoPhysics Related
🎯 What it does: This work proposes the PMGS framework, which reconstructs the complete spatiotemporal trajectories of projected motion from monocular videos using 3D Gaussian splatting.
PoeTone: A Framework for Constrained Generation of Structured Chinese Songci with LLMs
Zhan Qu (TU Dresden), Michael Färber
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper studies the performance of large language models in generating Song dynasty lyrics under structured constraints, and constructs the PoeTone evaluation framework and the Generate-Critic fine-tuning method.
Point Cloud Quantization Through Multimodal Prompting for 3D Understanding
Hongxuan Li (Tianjin University), Pengfei Zhu (Zhejiang Normal University)
CodeClassificationRecognitionSegmentationTransformerPrompt EngineeringVision Language ModelTextMultimodalityPoint Cloud
🎯 What it does: Proposes a text-prompt based point cloud quantization framework, which discretizes continuous point cloud features into text-driven prototypes and fuses visual details with semantic information through cross-attention.
CodeSegmentationAnomaly DetectionTransformerPoint CloudPhysics Related
🎯 What it does: Developed a high-resolution point cloud dataset of surface defects on ceramic packaging substrates named CPS3D-Seg, and proposed a 3D segmentation network based on causal inference called CINet.
🎯 What it does: Propose ProtoFormer, a Transformer-based point cloud semantic scene completion method that simultaneously accomplishes point cloud reconstruction and semantic annotation by leveraging learnable semantic prototypes and top-K attention mechanisms.
🎯 What it does: This paper proposes a domain generalization model called PointDGRWKV for point cloud classification, addressing the issues of local geometry modeling and cross-domain attention drift in RWKV on point cloud tasks.
Points Meet Pixels: Bridging 2D Vision-Language Model and 3D Perception Gaps for Point Cloud Quality Assessment
Mingxuan Li (Beijing Institute of Technology), Runze Hu (Beijing Institute of Technology)
CodeVision Language ModelPoint Cloud
🎯 What it does: Proposed PMP-PCQA, a point cloud quality assessment framework based on vision-language models, which bridges 2D VLM and 3D point clouds through point-pixel fine-grained correspondence, and constructs three modules (SAE, FCA, TAM) to achieve spatial enhancement, cross-modal consistency, and quality-sensitive feature mining.
🎯 What it does: Propose the MOSAIC algorithm, which helps agents select, integrate, and fine-tune strategies learned by others in distributed environments to accelerate their own learning.
Avik Kar (Indian Institute of Science), Rahul Singh (Indian Institute of Science)
CodeOptimizationReinforcement Learning
🎯 What it does: Studied average reward reinforcement learning in continuous space Lipschitz MDPs, and proposed adaptive exploration algorithms PZRL-MF and PZRL-MB based on policy zooming, providing finite-time regret upper bounds;
Polysemic Semantic Instance Network for Cross-Modal Hashing
Shuo Han, Lei Huang (Qingdao University of Technology)
CodeRetrievalTransformerMultimodality
🎯 What it does: Propose a multi-instance multi-modal hashing framework called DPSIH, which generates various semantic embeddings through multi-head self-attention and residual learning to achieve cross-modal retrieval.
PortraitSR: Artist-Inspired Prior Learning for Progressive Face Super-Resolution
Miaoqing Wang (Chongqing University of Post and Telecommunication), Long Sun (Nanjing University of Science and Technology)
CodeSuper ResolutionTransformerImage
🎯 What it does: Proposed a facial super-resolution framework called PortraitSR inspired by human painting, which includes structure-priority, texture-priority, and holistic fusion modules to recover high-quality facial images from low-resolution inputs.
CodeClassificationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Systematic evaluation and quantification of positional information bias in multimodal representation models (e.g., CLIP and its variants) for image-text retrieval tasks, distinguishing positional bias from contextual importance.
Positional Cognitive Specialization: Where Do LLMs Learn to Comprehend and Speak Your Language?
Luis Frentzen Salim (Academia Sinica), Hsing-Kuo Kenneth Pao (Academia Sinica)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigated the specialization of perception (understanding) and production (generation) functions in the front and back layers of large language models (LLMs) when learning low-resource languages, and proposed a layer selection strategy called CogSym, which trains only the outer 25% of layers, significantly reducing adaptation costs.
Post-Hoc Refinement for Multitask Symbolic Regression via Consensus-Accelerated Shapley Analysis
Xinyue Li (University of Electronic Science and Technology of China), Yu Zhang (University of Electronic Science and Technology of China)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTabularBenchmark
🎯 What it does: This paper proposes an MTGP-BS framework that synthesizes higher quality expressions by refining and reconstructing sub-expressions of the entire population in the later stages of multi-task symbolic regression.
🎯 What it does: Propose a node classification method based on posterior label smoothing, which adaptively generates soft labels by leveraging neighbor labels and global statistics.
PosterVerse: A Full-Workflow Framework for Commercial-Grade Poster Generation with HTML-Based Scalable Typography
Junle Liu (South China University of Technology), Lianwen Jin (South China University of Technology)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Built a full-flow commercial-grade poster generation framework called PosterVerse, which supports automatically generating blueprints, background images, and implementing editable, scalable text layouts using HTML; simultaneously launched the PosterDNA Chinese poster generation dataset, which includes three subsets: blueprints, backgrounds, and HTML.
🎯 What it does: This paper proposes a stealthy profile pollution attack (CREAT) based on constrained reinforcement learning, achieving precise misguidance of sequential recommenders through dual-layer optimization and pattern balance reward strategies.
PPGPT: Transferring Next-Token Modeling from Language to PPG Signals
Zexing Zhang (Changchun University of Technology), Qingxin Zhao (Changchun University of Technology)
CodeClassificationTransformerLarge Language ModelMixture of ExpertsContrastive LearningBiomedical DataBenchmark
🎯 What it does: Built a PPG-based foundation model called PPGPT, pre-trained using Next-Feature Token Prediction to achieve multi-task PPG signal analysis.
Practical Global and Local Bounds in Gaussian Process Regression via Chaining
Junyi Liu (National University of Singapore), Stanley Kok (National University of Singapore)
CodeExplainability and InterpretabilityTabular
🎯 What it does: Proposed a framework based on the chaining method for estimating upper and lower bounds on the global expected maximum and local uncertainty intervals in Gaussian Process Regression (GPR).
Devon R. Graham (University of British Columbia), Kevin Leyton-Brown (University of British Columbia)
CodeOptimizationHyperparameter SearchBenchmark
🎯 What it does: This paper improves the COUP algorithm for configuration methods, significantly enhancing practical performance while maintaining theoretical optimality guarantees, and demonstrates its effectiveness in SAT solution configuration.
PRAGWORLD: A Benchmark Evaluating LLMs’ Local World Model Under Minimal Linguistic Alterations and Conversational Dynamics
Sachin Vashistha (Indian Institute of Technology Kharagpur), Somak Aditya (Indian Institute of Technology Kharagpur)
CodeExplainability and InterpretabilityTransformerLarge Language ModelTextBenchmark
🎯 What it does: Evaluate the implicit world model plasticity of LLMs in dialogues, construct the PRAGWORLD benchmark, and test model robustness through seven minimal language variants.
Pre-DPO: Improving Data Utilization in Direct Preference Optimization Using a Guiding Reference Model
Junshu Pan (Zhejiang University), Yue Zhang (Westlake University)
CodeOptimizationHyperparameter SearchReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Propose Pre-DPO, a paradigm that introduces a guiding reference model in DPO/SimPO training. It first performs a standard preference optimization on the policy, then uses the optimized model as a reference to retrain the original policy, thereby better weighting the training samples.
🎯 What it does: This paper proposes the RandSF.Q method, which uses random slot-feature pairs to improve query prediction in video unsupervised object-centric learning.
Preference Elicitation for Step-Wise Explanations in Logic Puzzles
Marco Foschini (KU Leuven), Tias Guns (KU Leuven)
CodeOptimizationExplainability and InterpretabilityReinforcement LearningText
🎯 What it does: The study applies interactive preference elicitation (Constructive Preference Elicitation) to the step-by-step explanation of logic puzzles, aiming to learn users' linear preferences for explanation quality and generate more understandable explanation steps.
🎯 What it does: This study proposes PrefixGPT, a model based on generative pre-trained Transformers, capable of directly generating optimal prefix adder topologies from scratch that satisfy design rules.
Preserving Topological and Geometric Embeddings for Point Cloud Recovery
Kaiyue Zhou (Tsinghua University), Shengjin Wang (Tsinghua University)
CodeRestorationTransformerPoint Cloud
🎯 What it does: Developed an end-to-end point cloud recovery framework called TopGeoFormer, which can simultaneously preserve topological and geometric features during the sampling and recovery process.
🎯 What it does: Built an end-to-end pipeline for multi-person human mesh recovery from pressure pad data (PressTrack-HMR) and released the multi-person interactive pressure dataset MIP.
PRGB Benchmark: A Robust Placeholder-Assisted Algorithm for Benchmarking Retrieval-Augmented Generation
Zhehao Tan (Ant Group), Jinjie Gu (Ant Group)
CodeRetrievalLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Placeholder-RAG-Benchmark (PRGB) to conduct multi-dimensional fine-grained evaluation of LLMs' ability to utilize retrieved documents within RAG systems.
PriAgent: A Collaborative Multi-Agent Framework for Auditing Android Privacy Compliance
Ziwei Zhang (Chinese Academy of Sciences), Qingyun Liu (Chinese Academy of Sciences)
CodeSafty and PrivacyExplainability and InterpretabilityComputational EfficiencyLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented PriAgent, a multi-agent AI framework that automatically completes compliance audits for Android app code and natural language privacy policies, forming a complete closed-loop system from static analysis alerts to interpretable compliance reports;
Pricing Online LLM Services with Data-Calibrated Stackelberg Routing Game
Zhendong Guo (Southeast University), Jiahui Jin (Southeast University)
CodeOptimizationData-Centric LearningTransformerTabularFinance Related
🎯 What it does: Proposes the PriLLM framework for real-time dynamic pricing of a single service provider in an LLM routing platform, modeling user-provider interactions based on Stackelberg game theory;
PriorRG: Prior-Guided Contrastive Pre-training and Coarse-to-Fine Decoding for Chest X-ray Report Generation
Kang Liu (Xidian University), Qiguang Miao (Xidian University)
CodeGenerationTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodalityBiomedical DataRetrieval-Augmented Generation
🎯 What it does: This paper proposes the PriorRG framework, which utilizes patients' prior imaging, clinical indicators, and medical history information to generate chest X-ray reports that align better with clinical workflows.
PRISM: Privacy-Aware Routing for Adaptive Cloud–Edge LLM Inference via Semantic Sketch Collaboration
Junfei Zhan (University of Pennsylvania), Tengjiao He (University of Hong Kong)
CodeSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the PRISM framework to achieve cloud-edge collaborative LLM inference, utilizing dynamic routing, two-layer local differential privacy, and semantic sketch collaboration to enhance privacy and efficiency.
Privacy Auditing of Multi-Domain Graph Pre-Trained Model Under Membership Inference Attacks
Jiayi Luo (Beihang University), Jianxin Li (Guangxi Normal University)
CodeSafty and PrivacyAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: Propose the MGP-MIA framework for membership inference attacks on multi-domain graph pre-training models, enhancing the ability to detect privacy leakage
Privacy Leaks by Adversaries: Adversarial Iterations for Membership Inference Attack
Jing Xue (Xi'an Jiaotong University), Guang Dai (Fudan University)
CodeSafty and PrivacyAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Proposes a Membership Inference Attack (IMIA) based on the number of iterations required to generate adversarial samples, determining whether a sample is a member of the model's training set by measuring the number of iterations needed to generate adversarial samples.
Privacy on the Fly: A Predictive Adversarial Transformation Network for Mobile Sensor Data
Tianle Song (Xi'an Jiaotong University), Chao Shen (Xi'an Jiaotong University)
CodeSafty and PrivacyComputational EfficiencyRecurrent Neural NetworkGenerative Adversarial NetworkTime Series
🎯 What it does: Proposes the Predictive Adversarial Transformation Network (PATN), a real-time privacy protection framework that predicts future adversarial perturbations using historical sensor signals;
Privacy Preserving In-Context-Learning Framework for Large Language Models
Bishnu Bhusal (University of Missouri), Susmit Jha (SRI International)
CodeData SynthesisSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposed a differential privacy-based private prediction framework that generates high-quality synthetic text without fine-tuning large language models, and applies it to in-context learning (ICL).
Privacy-protected Retrieval-Augmented Generation for Knowledge Graph Question Answering
Yunfeng Ning (Wuhan University), Tieyun Qian (Wuhan University)
CodeGenerationRetrievalSafty and PrivacyTransformerLarge Language ModelGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the ARoG framework to address retrieval challenges caused by entity anonymization in KGQA, achieving a retrieval-augmented generation (RAG) system under privacy-preserving conditions.
Private Frequency Estimation via Residue Number Systems
Héber Hwang Arcolezi (Inria)
CodeComputational EfficiencyTabular
🎯 What it does: Propose a local differential privacy frequency estimation protocol based on the residue number system—Modular Subset Selection (MSS)—which achieves efficient frequency estimation with a single message.
Proactive Constrained Policy Optimization with Preemptive Penalty
Ning Yang (Institute of Automation Chinese Academy of Sciences), Jun Wang (Microsoft Research)
CodeOptimizationReinforcement Learning
🎯 What it does: This paper proposes Proactive Constrained Policy Optimization (PCPO), achieving safe reinforcement learning through proactive penalties and constraint-aware intrinsic rewards.
🎯 What it does: Proposed an unsupervised shape matching framework called PDCMatch, based on a probabilistic deformation consistency model in the spectral domain, which improves point correspondence by jointly estimating deformation and correspondence probabilities.
Probabilistic Hash Embeddings for Online Learning of Categorical Features
Aodong Li (Amazon Web Services), Balakrishnan Murali Narayanaswamy (Amazon Web Services)
CodeClassificationRecommendation SystemTabularTime Series
🎯 What it does: This paper proposes a probabilistic hashing embedding (PHE) model to handle scalable and dynamically changing categorical feature vocabularies in a streaming online learning environment.
Probabilistic Hierarchical Goal Network Planning with UCT
David H. Chan (University of Maryland), Dana S. Nau (U.S. Naval Research Laboratory)
CodeOptimizationReinforcement LearningBenchmark
🎯 What it does: Proposed a probabilistic hierarchical goal network (Probabilistic HGN) planning framework and implemented two UCT-based solving algorithms.
ProBench: Benchmarking GUI Agents with Accurate Process Information
Leyang Yang (Zhejiang University), Yong Li (Ant Group)
CodeLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Propose ProBench, a mobile benchmark containing over 200 multi-step GUI tasks covering state-related tasks and process-related tasks, and design an automatic evaluation pipeline.
Probing Preference Representations: A Multi-Dimensional Evaluation and Analysis Method for Reward Models
Chenglong Wang (Northeastern University), Tong Xiao (Northeastern University)
CodeExplainability and InterpretabilityRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: By constructing a six-dimensional MRMBench, the preference representation of reward models is evaluated using probe methods, and interpretability is enhanced through inference-time probe analysis.
🎯 What it does: ProCAST proposes a projection framework that maps unconstrained multivariate time series predictions into feasible domains satisfying coupled aggregation constraints through orthogonal or oblique projections, ensuring prediction feasibility.
ProFuser: Progressive Fusion of Large Language Models
Tianyuan Shi (Sun Yat-sen University), Wu Kai (Alibaba Group)
CodeKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the ProFuser method, which progressively fuses multiple large language models using a dual evaluation mechanism combining the minimum cross-entropy of training modes and reward model voting in inference modes.
🎯 What it does: Designed a GNN-to-MLP knowledge distillation framework called ProGMLP, which adjusts inference cost and accuracy on demand through progressively trained multi-layer MLP students to achieve flexible inference.
ProgRAG: Hallucination-Resistant Progressive Retrieval and Reasoning over Knowledge Graphs
Minbae Park (Hanyang University), Hyunjoon Kim (Seoul National University)
CodeRetrievalGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringGraphRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the ProgRAG framework, which decomposes complex knowledge graph question answering problems into sub-questions, iteratively retrieves information, and constructs optimal context through LLM pruning and prefix enumeration re-ranking;
🎯 What it does: Designed a verifiable step-wise reward mechanism (VSRM) to enhance the inference efficiency of large reasoning models and reduce overthinking.
PromptEmo: Learning Emotion with Bilateral Textual Prompts in Multi-Domain Open-set Scenarios
Xinyi Zeng (Sichuan University), Yan Wang (Chengdu University of Information Technology)
CodeClassificationRecognitionDomain AdaptationTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Proposed a bidirectional text prompting framework called PromptEmo based on CLIP to address the multi-domain open-set facial expression recognition (MO-FER) task.