π― What it does: Construct the EDyGS model to achieve dynamic scene deblurring and novel view synthesis by leveraging event cameras and blurry monocular videos.
Efficient Dynamic Ensembling for Multiple LLM Experts
Jinwu Hu (South China University of Technology), Mingkui Tan (South China University of Technology)
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIPrompt EngineeringMixture of ExpertsText
π― What it does: Proposed an efficient dynamic ensemble reasoning (DER) framework that integrates complementary knowledge from multiple large language model (LLM) experts by training an agent to sequentially invoke them on input questions, generating higher-quality answers.
Efficient Dynamic Graphs Learning with Refined Batch Parallel Training
Zhengzhao Feng (Zhejiang University), Mingli Song (Zhejiang University)
CodeComputational EfficiencyGraph Neural NetworkGraphTime Series
π― What it does: Propose the RBT framework to address the memory staleness problem in memory-based temporal graph neural networks (MTGNN) during batch parallel training;
Efficient Inter-Operator Scheduling for Concurrent Recommendation Model Inference on GPU
Shuxi Guo (Beijing University of Posts and Telecommunications), Jingyu Wang (Beijing University of Posts and Telecommunications)
CodeRecommendation SystemComputational Efficiency
π― What it does: Propose RecOS, a scheduling system for concurrent inference of recommendation models (RM) on GPUs, addressing issues caused by traditional single-stream or topology-based multi-stream scheduling, such as operator queuing, cache conflicts, and low GPU resource utilization.
π― What it does: Proposed a visual representation learning framework HcNet based on the heat conduction equation, designed the Heat Conduction Layer and Refinement Approximation Layer, and constructed an interpretable visual backbone network.
π― What it does: Propose the EfficientPIE framework, which uses single-frame images to predict in real-time whether pedestrians have the intention to cross the street;
Endowing Interpretability for Neural Cognitive Diagnosis by Efficient Kolmogorov-Arnold Networks
Shangshang Yang (Anhui University), Ye Tian (Anhui University)
CodeExplainability and InterpretabilityTabularSequential
π― What it does: Proposed KAN2CD, which utilizes Kolmogorov-Arnold networks to enhance the explainability of neural cognitive diagnostic models while maintaining or even improving prediction accuracy.
Enhancing Counterfactual Estimation: A Focus on Temporal Treatments
Xin Wang (University of Science and Technology of China), Chunyan Miao (Nanyang Technological University)
CodeComputational EfficiencyRepresentation LearningRecurrent Neural NetworkAuto EncoderTabularTime SeriesBiomedical DataElectronic Health Records
π― What it does: Constructed the CTD-NKO model based on the Koopman operator and RNN for multi-timepoint causal counterfactual estimation, focusing on capturing interactions in time series treatments.
CodeClassificationRecognitionMixture of ExpertsContrastive LearningImage
π― What it does: A framework for independent and collaborative learning (ICL) is proposed to address long-tailed visual recognition tasks, achieving independent optimization of each expert while maintaining their advantages within a Mixture of Experts (MoE) model to enhance overall performance.
π― What it does: Propose the Memory-Driven Prompt Learning framework, using generative prompts and shared prompts to compensate for missing modalities, enabling the unified Transformer to maintain robustness under missing modalities.
Enhancing Multimodal Protein Function Prediction Through Dual-Branch Dynamic Selection with Reconstructive Pre-Training
Xiaoling Luo (Shenzhen University), Junsong Wang (Shenzhen Technology University)
CodeDrug DiscoveryTransformerLarge Language ModelMixture of ExpertsMultimodalityBiomedical Data
π― What it does: This paper proposes a dual-branch, reconstructive pre-training based multi-modal protein function prediction framework called DSRPGO.
Enhancing Table Recognition with Vision LLMs: A Benchmark and Neighbor-Guided Toolchain Reasoner
Yitong Zhou (University of Science and Technology of China), Xin Li (iFLYTEK Co., Ltd)
CodeRecognitionTransformerVision Language ModelImageBenchmarkRetrieval-Augmented Generation
π― What it does: Proposed a Vision LLM-based table recognition evaluation benchmark and designed the Neighbor-Guided Toolchain Reasoner (NGTR) framework to enhance table recognition performance in no-training scenarios.
π― What it does: Propose the ESBN framework by analyzing the estimation bias of batch normalization (BN), selectively replacing BN layers with batch-free normalization (BFN), and combining instance statistics and Gaussian Mixture Model (GMM) to achieve source unsupervised domain adaptation.
Evaluating and Mitigating Linguistic Discrimination in Large Language Models: Perspectives on Safety Equity and Knowledge Equity
Guoliang Dong (Singapore Management University), Xinyu Wang (Zhejiang University)
CodeSafty and PrivacyAdversarial AttackLarge Language ModelTextBenchmark
π― What it does: Evaluate the safety and quality bias of large language models in multilingual scenarios, and propose a lightweight consistency voting method called LDFighter to mitigate this bias.
Exact Algorithms with New Upper Bounds for the Maximum k-plex Problem
Jiongzhi Zheng (Huazhong University of Science and Technology), Kun He (Huazhong University of Science and Technology)
CodeOptimizationGraph
π― What it does: In the branch-and-bound framework for the maximum k-plex problem (MKP), the authors propose two new upper bounds: RelaxGCB (a relaxed upper bound based on graph coloring) and RelaxPUB (a combination of RelaxGCB and the latest partitioning upper bound), using them to improve eight existing exact solvers.
Exploiting Label Skewness for Spiking Neural Networks in Federated Learning
Di Yu (Zhejiang University), Shuiguang Deng (Zhejiang University)
CodeFederated LearningSafty and PrivacyKnowledge DistillationSpiking Neural NetworkImageTime Series
π― What it does: This paper proposes the FedLEC framework for federated learning on label-skewed edge devices to train deep spiking neural networks (SNN) while preserving data privacy.
Exploring Efficient and Effective Sequence Learning for Visual Object Tracking
Dongdong Li (National University of Defense Technology), Rui Chen (National University of Defense Technology)
CodeObject TrackingTransformerVideo
π― What it does: Proposes the FastSeqTrack framework, achieving efficient visual tracking by generating target bounding box coordinates in a single parallel step
π― What it does: Propose the AniSora system, which realizes functions such as animation video generation, frame interpolation, and local guidance, and provides a large-scale dataset and evaluation benchmark.
π― What it does: Investigated the over-smoothing problem in graph neural networks for graph classification tasks, and proposed an SDE method based on high-entropy node sampling and discretization.
Exploring the Trade-Offs: Quantization Methods, Task Difficulty, and Model Size in Large Language Models From Edge to Giant
Jemin Lee (Electronics and Telecommunications Research Institute), Yongin Kwon (Electronics and Telecommunications Research Institute)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: This paper evaluates systematically quantized instruction-tuned LLMs with parameters ranging from 1B to 405B on 13 benchmark datasets;
Exploring Transferable Homogenous Groups for Compositional Zero-Shot Learning
Zhijie Rao (Hong Kong Polytechnic University), Mengzhu Wang (Hong Kong Polytechnic University)
CodeClassificationRepresentation LearningTransformerPrompt EngineeringMixture of ExpertsVision Language ModelContrastive LearningMultimodality
π― What it does: Proposed Homogeneous Group Representation Learning (HGRL), which balances transferability and discriminability by learning representations of states and objects within homogeneous groups (clusters of the same attribute category), and introduced multi-expert networks, distributed group prompts, and intra-group enhancement in visual, prompt, and pair branches;
Xiren Zhou (University of Science and Technology of China), Huanhuan Chen (University of Science and Technology of China)
CodeAnomaly DetectionRecurrent Neural NetworkTransformerTime Series
π― What it does: Propose a model space fault diagnosis framework based on multi-path reservoir REDNet, fitting each time-series multivariate data into the REDNet model and performing classification in the model space.
π― What it does: Propose FedAPA under the federated learning framework, which generates personalized models through server-side gradient adaptive aggregation and reduces communication by partially sharing models.
π― What it does: This paper proposes a few-shot incremental multi-modal learning framework based on Tactile-Guided and Imaginary Visual Synthesis (TIFS), addressing the problems of catastrophic forgetting and modality imbalance in multi-modal incremental learning.
π― What it does: Built a neuro-symbolic system, FGeo-HyperGNet, to automatically generate readable and verifiable solution processes from geometric images and text.
Fine-Grained and Efficient Self-Unlearning with Layered Iteration
Hongyi Lyu (Macquarie University), Lianyong Qi (China University of Petroleum (East China))
CodeClassificationSafty and PrivacyComputational EfficiencyKnowledge DistillationImage
π― What it does: Designed and proposed the Self-Unlearning with Layered Iteration (SULI) method, achieving efficient fine-grained forgetting in machine learning models through layered iteration and soft label selective probability adjustment.
π― What it does: Developed a new recommendation framework FCSRec that explicitly models unobserved confounding factors and their time-varying effects, jointly learning them with sequential dependencies.
Free Lunch of Image-mask Alignment for Anomaly Image Generation and Segmentation
Xiangyue Li (Soochow University), Mingjie Sun (Soochow University)
CodeSegmentationGenerationAnomaly DetectionDiffusion modelImageBiomedical Data
π― What it does: This paper proposes a dual-branch training strategy, enabling the generative model to simultaneously generate anomalous images and their masks, while improving the correspondence between images and masks through alignment regularization. Subsequently, the pre-trained generative model provides high-quality data and generates feedback loss to further enhance segmentation performance.
FreqLLM: Frequency-Aware Large Language Models for Time Series Forecasting
Shunnan Wang (Chongqing University), Guansong Pang (Singapore Management University)
CodeTransformerLarge Language ModelPrompt EngineeringTime Series
π― What it does: Propose the FreqLLM framework, which embeds dual-scale frequency domain signals into LLMs and aligns them with the semantic space through soft prompts to enhance time series prediction performance.
π― What it does: Proposed the Abductive Abstract Reinforcement Learning (A2RL) framework, which integrates neural networks with symbolic reasoning to directly learn abstract high-level steps from raw perceptual inputs and represent them in the form of an abstract state machine (ASM);
π― What it does: Propose a CS-GBSBF method that converts facial expression images into a graph structure via granular spherical representation, separately extracting visual and spatial features. A component separation network is used to obtain visual/spatial representations of key facial regions, and in the fusion network, spatial representations adaptively guide the fusion of visual features, ultimately completing emotion recognition.
π― What it does: A deep reinforcement learning scheduling framework named GATES, which integrates graph attention networks (GAT) with evolutionary strategies (ES), is proposed to address the cost-aware dynamic workflow scheduling (CADWS) problem in cloud computing environments.
General Incomplete Time Series Analysis via Patch Dropping Without Imputation
Yangyang Wu (Zhejiang University), Meng Xi (Zhejiang University)
CodeClassificationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningTime Series
π― What it does: Propose an end-to-end framework called INTER that directly analyzes incomplete multivariate time series, completely bypassing traditional missing value imputation steps.
π― What it does: Proposed a Target-oriented Graph Neural Network (TGNN), which enhances the representation capability of multi-view graph data by separating the determinative and incidental features of each view and using a class-level dual-objective loss to achieve task-oriented semantic learning.
Good Advisor for Source Localization: Using Large Language Model to Guide the Source Inference Process
Dongpeng Hou (Northwestern Polytechnical University), Zhen Wang (Northwestern Polytechnical University)
CodeAnomaly DetectionGraph Neural NetworkTransformerLarge Language ModelContrastive LearningText
π― What it does: Proposed the CRSLL framework, which uses LLM as an 'advisor' to generate comment source analysis, combining contrastive learning, differentiable feature masking, and cross-modal attention to achieve the localization of rumor sources
GPI-Net: Gestalt-Guided Parallel Interaction Network via Orthogonal Geometric Consistency for Robust Point Cloud Registration
Weikang Gu (Fujian Agriculture and Forestry University), Lifang Wei (Fujian Agriculture and Forestry University)
CodePose EstimationTransformerPoint Cloud
π― What it does: This paper proposes a point cloud registration network called GPI-Net, which effectively removes outliers and accurately estimates rigid transformations in the initial correspondence point set.
Zhaolong Ling (Anhui University), Zhangling Duan (Institute of Artificial Intelligence Hefei Comprehensive National Science Center)
CodeOptimizationAuto EncoderGraphBiomedical Data
π― What it does: GCFS employs gradient optimization combined with autoencoders, acyclic constraints, and mask methods to select the Markov Blanket (causal features) for the target variable.
π― What it does: Proposes the GRAML (Goal Recognition As Metric Learning) framework, which uses deep metric learning (Siamese LSTM) to distinguish trajectories of different goals in the embedding space, achieving online dynamic goal recognition (ODGR) and supporting offline self-supervised training, one-time adaptation to new goals, and unified processing of continuous and discrete environments.
GRAPE: Heterogeneous Graph Representation Learning for Genetic Perturbation with Coding and Non-Coding Biotype
Changxi Chi (Zhejiang University), Stan Z. Li (Westlake University)
CodeRepresentation LearningGraph Neural NetworkLarge Language ModelContrastive LearningMultimodalityGraphBiomedical Data
π― What it does: This paper proposes the GRAPE model, which initializes gene representations using multimodal features from gene descriptions and DNA sequences, constructs a heterogeneous graph to learn gene regulatory networks, and achieves accurate predictions in single-cell gene perturbation tasks.
π― What it does: Proposes the GMVC framework, unifying multi-view clustering (MVC) and multi-graph clustering (MVGC) under a single model, and achieving representation learning and clustering through contrastive learning on graph embeddings;
π― What it does: Proposed a parameterizable logical reasoning forward-chaining grounding method that controls reasoning depth and width to balance scalability and expressiveness in NeSy models.
Handling Infinite Domain Parameters in Planning Through Best-First Search with Delayed Partial Expansions
Γngel Aso-Mollar (Valencian Research Institute for Artificial Intelligence, Universitat PolitΓ¨cnica de ValΓ¨ncia), Eva Onaindia (Valencian Research Institute for Artificial Intelligence, Universitat PolitΓ¨cnica de ValΓ¨ncia)
CodeOptimizationTabularBenchmark
π― What it does: This paper proposes a method for automated planning problems involving infinite-domain control parameters, treating control parameters as decision points rather than constraints. It implements systematic search using a Sampling Best-First Search (SBFS) algorithm based on best-first search with delayed partial expansion, and proves its probabilistic completeness under certain conditions;
π― What it does: Designed and implemented a fundamental attack model HeTa based on relational units, which can migrate across different heterogeneous graph neural networks (HGNN) and quickly adapt to new graphs, achieving low-budget node injection attacks.
π― What it does: Node classification on heterogeneous graphs, proposing the HAPPY framework: combining Heterogeneity-Aware Personalized PageRank (H-PPR) with adaptive subgraph extraction to capture both homogeneous and heterogeneous neighbors.
CodeClassificationGraph Neural NetworkMixture of ExpertsGraph
π― What it does: Propose HGEN, an ensemble learning framework for heterogeneous graphs, which generates diverse homogeneous subgraphs through meta-paths and trains multiple base learners (GNNs). Subsequently, residual attention fusion and correlation regularization are applied for integration, significantly improving node classification accuracy.
π― What it does: This paper proposes a high-confidence local structure-guided consensus graph learning method, HLSCG IMC, to address the problems of missing views and information imbalance in incomplete multi-view clustering.
π― What it does: Propose the GraphWalker framework, which can directly generate high-fidelity road network maps from noisy trajectories, achieving end-to-end road network reconstruction.
HIPP: Protecting Image Privacy via High-Quality Reversible Protected Version
Xi Ye (Wuhan University), Geying Yang (Wuhan University)
CodeSafty and PrivacyFlow-based ModelGenerative Adversarial NetworkImage
π― What it does: Propose a reversible thumbnail-based privacy protection scheme HIPP, which utilizes latent space to separate detail and contour information and generates natural protected images
HiTuner: Hierarchical Semantic Fusion Model Fine-Tuning on Text-Attributed Graphs
Zihan Fang (Fuzhou University), Shiping Wang (Fuzhou University)
CodeClassificationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraph
π― What it does: Propose the HiTuner framework, which enhances the node classification performance of text attribute graphs (TAG) by combining the multi-layer hidden states of LLM with fine-tuned PLM and utilizing a confidence network to adaptively fuse semantic information from different layers.
How to Mitigate Information Loss in Knowledge Graphs for GraphRAG: Leveraging Triple Context Restoration and Query-Driven Feedback
Manzong Huang (Hefei University of Technology), Xindong Wu (Hefei University of Technology)
CodeRetrievalRepresentation LearningTransformerLarge Language ModelTextGraphRetrieval-Augmented Generation
π― What it does: Designed and implemented the Triple Context Restoration and Query-Driven Feedback (TCR-QF) framework, which first restores the original text context of triplets to compensate for information loss, and then dynamically completes missing knowledge graphs through a query-driven feedback loop during inference.
HPDM: A Hierarchical Popularity-aware Debiased Modeling Approach for Personalized News Recommender
Xiangfu He (Tianjin University), Hongtao Liu (Du Xiaoman Financial Technology)
CodeRecommendation SystemTransformerText
π― What it does: Propose a hierarchical news recommendation model HPDM that considers news popularity, aiming to correct popularity bias in user click data.
Fusheng Hao (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences), Jun Cheng (Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences)
CodeClassificationObject DetectionSafty and PrivacyRepresentation LearningTransformerImage
π― What it does: Propose an image encryption and learnable privacy protection framework, utilizing two strategies: random shuffling (RS) and subblock mixing (MI), to encrypt images into human-unidentifiable but machine-recognizable forms, with minimal modifications to ViT and YOLOS to support classification and detection on encrypted images.
Human-Readable Neuro-Fuzzy Networks from Frequent Yet Discernible Patterns in Reward-Based Environments
John Wesley Hostetter (North Carolina State University), Min Chi (North Carolina State University)
CodeExplainability and InterpretabilityReinforcement LearningAuto EncoderTabular
π― What it does: This paper proposes a self-organizing and simplified neural fuzzy network (NFN) method, leveraging fuzzy information granulation and graph theory techniques to retain only frequent but discriminable patterns, thereby generating interpretable strategies.
π― What it does: This paper proposes a multimodal emotion recognition framework based on hybrid relational graphs and emotional semantic alignment (HRG-SSA).
π― What it does: Proposes the HyperTrans method to achieve cross-domain few-shot industrial defect detection, leveraging hypergraphs and a perturbation correction framework to transfer and fuse features from the source domain RGB images with the target domain 3D depth maps.
Identifying Drivers of Predictive Aleatoric Uncertainty
Pascal Iversen (University of Potsdam), Bernhard Y. Renard (University of Potsdam)
CodeExplainability and InterpretabilityImageTabular
π― What it does: This paper proposes a simple method that modifies neural network outputs to follow a Gaussian distribution and directly interprets the variance output using existing interpreters to explain the model's predictive uncertainty.
Image-Enhanced Hybrid Encoding with Reinforced Contrastive Learning for Spatial Domain Identification in Spatial Transcriptomics
Daoyuan Wang (Central South University), Fei Guo (Central South University)
CodeGraph Neural NetworkTransformerReinforcement LearningAuto EncoderContrastive LearningImageMultimodalityBiomedical Data
π― What it does: Propose the IE-HERCL framework, which integrates gene expression, spatial coordinates, and tissue images for spatial domain identification in spatial transcriptomics through hybrid encoding, cross-attention, and enhanced contrastive learning.
Imagination-Limited Q-Learning for Offline Reinforcement Learning
Wenhui Liu (East China Normal University), Shuigeng Zhou (Fudan University)
CodeReinforcement LearningDiffusion modelWorld Model
π― What it does: Proposed a new offline reinforcement learning method called Imagination-Limited Q-learning (ILQ), which corrects value estimation by generating out-of-distribution (OOD) action values through a model and capping them with the maximum value of the behavior policy.
CodeExplainability and InterpretabilityReinforcement Learning from Human FeedbackReinforcement LearningAgentic AI
π― What it does: Propose a hierarchical agent architecture named HA2, leveraging hierarchical reinforcement learning to achieve zero-shot collaboration, and conduct cooperative experiments with humans and unknown agents in the Overcooked game.
Improving Consistency Identification in Task-oriented Dialogue Through Multi-Agent Collaboration
Peng Wang (Central South University), Libo Qin (Central South University)
CodeRecognitionLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
π― What it does: This paper proposes a multi-agent collaborative zero-shot task-oriented dialogue consistency identification framework called MAC-CIToD.
π― What it does: Proposes an early exit method based on CAP (Certainty-Aware Probability) scores, integrating class-related logits with class-agnostic information (NSP scores) to improve confidence estimation in predictions.
In-context Learning Demonstration Generation with Text Distillation
Wuyuqing Wang (Xidian University), Cheng Deng (Xidian University)
CodeKnowledge DistillationData-Centric LearningTransformerLarge Language ModelText
π― What it does: Proposed a demonstration generation framework DDG based on data distillation, which automatically synthesizes representative demonstration examples that capture information from the original dataset by matching gradients between the generator and computational model, as well as employing a teacher-student EMA mechanism, to enhance the In-Context Learning (ICL) performance of large language models.
π― What it does: In a multi-task environment, the In-Context Meta LoRA (ICM-LoRA) framework is proposed, which utilizes a conditional variational autoencoder (CVAE) to generate LoRA parameters based on task vectors, enabling task-specific customization of large language/visual models without requiring additional fine-tuning.
Incentivizing Safer Actions in Policy Optimization for Constrained Reinforcement Learning
Somnath Hazra (Indian Institute of Technology Kharagpur), Soumyajit Dey (Indian Institute of Technology Kharagpur)
CodeReinforcement LearningBenchmark
π― What it does: Propose a safe reinforcement learning algorithm named Incrementally Penalized Proximal Policy Optimization (IP3O), which balances return maximization with constraint satisfaction by incorporating an adaptive incentive-penalty mechanism into the reward function.
Inference of Human-derived Specifications of Object Placement via Demonstration
Alex Cuellar (Massachusetts Institute of Technology), Julie A Shah (Massachusetts Institute of Technology)
Code
π― What it does: Proposed a position-enhanced RCC (PARCC) spatial specification language and presented an inference framework based on demonstration learning; validated through human-machine experiments that the framework can better capture human spatial arrangement preferences compared to manually provided specifications.
InfVC: An Inference-Enhanced Local Search Algorithm for the Minimum Vertex Cover Problem in Massive Graphs
Rui Sun (Northeast Normal University), Jian Gao (Northeast Normal University)
CodeOptimizationGraph
π― What it does: Propose a reasoning-enhanced local search algorithm called InfVC for solving the minimum vertex cover problem on large-scale graphs.
Injecting Imbalance Sensitivity for Multi-Task Learning
Zhipeng Zhou (University of Science and Technology of China), Wei Gong (University of Science and Technology of China)
CodeOptimizationImageTextBenchmark
π― What it does: Propose IMGrad, a gradient descent method that injects imbalance sensitivity in multi-task learning, capable of simultaneously avoiding Pareto failure and individual progress imbalance.
Instantiation-based Formalization of Logical Reasoning Tasks Using Language Models and Logical Solvers
Mohammad Raza (Qatar Computing Research Institute), Natasa Milic-Frayling (Qatar Computing Research Institute)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: Propose the Semantic Self-Verification (SSV) method, which combines large language models with logic solvers, using concrete instantiations to verify and correct abstract formalized logical reasoning tasks.
InstGAN: Instant Actor-Critic-Driven GAN for De Novo Molecule Generation and Property Optimization
Huidong Tang (Shandong Institute of Commerce and Technology), Yasuhiko Morimoto (Hiroshima University)
CodeOptimizationDrug DiscoveryRecurrent Neural NetworkReinforcement LearningGenerative Adversarial NetworkBiomedical Data
π― What it does: This paper proposes InstGAN, a GAN based on actor-critic reinforcement learning, which generates multi-attribute optimized molecules at the SMILES level by utilizing immediate rewards and global rewards.
π― What it does: Semantic embeddings obtained by optimizing the gradient of the attribute classifier guide text-to-image diffusion models to achieve high-quality, decoupled image editing.
Integrating Answer Set Programming and Large Language Models for Enhanced Structured Representation of Complex Knowledge in Natural Language
Mario Alviano (University of Calabria), Luis Angel Rodriguez Reiners
CodeRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmark
π― What it does: This paper proposes integrating Answer Set Programming (ASP) with Large Language Models (LLM) by automatically generating prompts via YAML configuration files, enabling the extraction of structured facts from natural language. It leverages ASP for logical reasoning, ultimately achieving a structured representation of complex knowledge.
Integration of Old and New Knowledge for Generalized Intent Discovery: A Consistency-driven Prototype-Prompting Framework
Xiao Wei (Tianjin University), Jianwu Dang (Shenzhen Institute of Advanced Technology)
CodeClassificationDomain AdaptationTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
π― What it does: Propose a consistency-driven prototype-prompt framework (CPP) that integrates old and new knowledge in the general intent discovery task, enhancing the recognition of unknown intents;
Lvye Cui (Beijing Institute of Technology), Dario Paccagnan (Imperial College London)
CodeOptimization
π― What it does: Propose a general inverse game framework that estimates players' utility parameters from observed Nash equilibria through an incenter selection method.
KnowMDD: Knowledge-guided Cross Contrastive Learning for Major Depressive Disorder Diagnosis
Anchen Lin (Hangzhou City University), Binbin Zhou (Hangzhou City University)
CodeRepresentation LearningGraph Neural NetworkContrastive LearningGraphBiomedical Data
π― What it does: Proposes a knowledge-guided cross-contrastive learning framework called KnowMDD, which constructs multi-view brain networks using multi-template brain atlases, enhances feature learning by integrating the default mode network (DMN) subgraph with attention mechanisms, and obtains robust MDD diagnostic representations through cross-contrastive learning.
KnowRA: Knowledge Retrieval Augmented Method for Document-level Relation Extraction with Comprehensive Reasoning Abilities
Chengcheng Mai (Nanjing Normal University), Yihua Huang (Nanjing University)
CodeRetrievalRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraphRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Designed and implemented the KnowRA model for document-level relation extraction, integrating multi-layer heterogeneous graph encoding, coreference resolution, external knowledge retrieval and filtering, and axial attention mechanisms to achieve comprehensive reasoning capabilities.
Siyuan Yang (Hong Kong University of Science and Technology (Guangzhou)), Wenjia Wang (Hong Kong University of Science and Technology (Guangzhou))
CodeComputational EfficiencyPhysics Related
π― What it does: This paper proposes a novel physics-informed neural network framework called KP-PINNs, which solves various partial differential equations in both forward and inverse problems by using the RKHS norm instead of the traditional L2 error and accelerating the kernel matrix inversion using the Kernel Packet (KP) method.
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringAgriculture RelatedPhysics Related
π― What it does: Developed a hybrid Bayesian optimization framework named BORA, integrating large language models (LLMs) with traditional Bayesian optimization (BO) to dynamically adjust LLM interventions and enhance the efficiency of experimental design search.
Language-Conditioned Open-Vocabulary Mobile Manipulation with Pretrained Models
Shen Tan (Harbin Institute of Technology), Guanghui Sun (Harbin Institute of Technology)
CodePose EstimationRobotic IntelligenceConvolutional Neural NetworkLarge Language ModelVision Language ModelVision-Language-Action ModelImageTextMultimodality
π― What it does: This paper proposes the LOVMM framework, which utilizes LLM and VLM to achieve open-vocabulary mobile manipulation driven by free-text instructions, covering navigation and 6-DoF grasping and placement.
Latte: Transfering LLMs' Latent-level Knowledge for Few-shot Tabular Learning
Ruxue Shi (Jilin University), Xin Wang (Jilin University)
CodeClassificationKnowledge DistillationRepresentation LearningData-Centric LearningMeta LearningTransformerLarge Language ModelTabular
π― What it does: Propose the Latte framework, which extracts latent knowledge through LLM during the training phase to guide few-shot table learning.
Anastasia Isychev (TU Wien), Maria Christakis (TU Wien)
CodeComputational EfficiencyImageTextTabularAudio
π― What it does: Implemented a lazy testing framework called LAZ, which dynamically skips redundant model calls and reorders the calling sequence when performing hyperproperty testing on machine learning models, thereby improving test throughput.
π― What it does: Propose a Joint View Imputation and Label Generation (JVILG) method, which utilizes fine-grained anchors to simultaneously recover missing views and directly generate soft labels, thereby achieving incomplete multi-view clustering.
Learn to Think: Bootstrapping LLM Logic Through Graph Representation Learning
Hang Gao (Institute Of Software Chinese Academy Of Sciences), Huaping Liu (Tsinghua University)
CodeOptimizationRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextGraphChain-of-Thought
π― What it does: Propose the L2T framework, modeling the LLM reasoning process as a graph, and jointly learning reasoning strategies through LLM and GNN to accomplish multi-task reasoning without task-specific prompts.
π― What it does: This paper proposes a new linear Transformer self-attention mechanism called Attention Graph Filter (AGF), which enhances the expressiveness of self-attention by directly learning graph filters in the singular value domain.
π― What it does: Propose a new neuro-symbolic learning framework called LUBAC, which approximates gradients by compiling lower and upper bound arithmetic circuits under CNF constraints that cannot be fully compiled.
π― What it does: Proposes the HetPFL framework, which simultaneously learns the local Pareto frontier and global Pareto frontier for each client in federated learning.
π― What it does: Propose a model based on multi-dimensional neural jump stochastic differential equations (SDE) and an encoder-free latent graph structure for modeling multivariate point processes;
π― What it does: Propose a frequency-domain based neural vocoder RNDVoC, achieving linear inversion of Mel spectra and residual detail reconstruction through range-nullspace decomposition;
Learning Probabilistic Temporal Logic Specifications for Stochastic Systems
Rajarshi Roy (University of Oxford), Marta Kwiatkowska (University of Oxford)
CodeOptimizationExplainability and InterpretabilityReinforcement LearningSequential
π― What it does: This paper proposes a passive learning algorithm for learning concise probabilistic linear temporal logic (PLTL) specifications from positive and negative sample Markov chains, capable of distinguishing systems with random behavior.
Learning to Explain: Towards Human-Aligned Explainability in Deep Reinforcement Learning via Attention Guidance
Bokai Ji (Xidian University), Gang Xiao (National Key Laboratory for Complex Systems Simulation)
CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerReinforcement LearningImageBenchmark
π― What it does: Developed an attention-guided interpretable deep reinforcement learning framework, Concept-PPO, which can generate concept-level explanations highly consistent with human cognition during the decision-making process.
LensNet: An End-to-End Learning Framework for Empirical Point Spread Function Modeling and Lensless Imaging Reconstruction
Jiesong Bai (University of Macau), Xuhang Chen (University of Macau)
CodeRestorationSuper ResolutionConvolutional Neural NetworkImagePhysics Related
π― What it does: This paper proposes LensNet, an end-to-end deep learning framework for point spread function (PSF) modeling and image reconstruction in optical encoded lensless imaging systems.