AAAI Conference on Artificial Intelligence Β· 1442 papers
DreamPhysics: Learning Physics-Based 3D Dynamics with Video Diffusion Priors
Tianyu Huang (Harbin Institute of Technology), Rynson W. H. Lau (City University of Hong Kong)
CodeGenerationData SynthesisKnowledge DistillationDiffusion modelScore-based ModelVideoPhysics Related
π― What it does: Utilizing the physical priors of video diffusion models to learn the material properties of 3D Gaussian point fields, and then simulating dynamics through Material Point Method (MPM) to ultimately generate 4D content that complies with physical laws.
DreamUHD: Frequency Enhanced Variational Autoencoder for Ultra-High-Definition Image Restoration
Yidi Liu (University of Science and Technology of China), Xueyang Fu (University of Science and Technology of China)
CodeRestorationSuper ResolutionAuto EncoderImage
π― What it does: A super high-resolution image restoration framework based on frequency domain priors, called Frequency Domain Variational Autoencoder (FE-VAE) with a wavelet adapter, is designed to perform various UHD image restoration tasks in a low-parameter and efficient latent space.
π― What it does: We propose DriveEditor, a unified framework based on diffusion models that enables object relocation, insertion, replacement, and deletion in driving scene videos.
π― What it does: We propose DrivingForward, a feedforward model based on 3D Gaussian splatting that can reconstruct driving scenes in real-time and supports multi-frame input with any number of surrounding views.
Drop the Beat! Freestyler for Accompaniment Conditioned Rapping Voice Generation
Ziqian Ning (Northwestern Polytechnical University), Lei Xie (Microsoft)
CodeGenerationTransformerLarge Language ModelAudio
π― What it does: Freestyler has been developed - the first model capable of directly generating rap vocals based on lyrics and accompaniment, and a large-scale rap dataset called RapBank has been constructed.
DSGram: Dynamic Weighting Sub-Metrics for Grammatical Error Correction in the Era of Large Language Models
Jinxiang Xie (Beijing Jiaotong University), Xiaojun Wan (Peking University)
CodeLarge Language ModelPrompt EngineeringText
π― What it does: A framework named DSGram is proposed for the dynamic weighted sub-indicator evaluation of grammar error correction (GEC) models, implementing sub-indicator scoring and weight generation based on LLM.
π― What it does: A DSRC framework is proposed, which enhances the robustness of multi-agent collaborative perception under natural disturbances through sparse-to-dense knowledge distillation and semantically guided point cloud re-rendering.
π― What it does: A dual conditional motion diffusion framework (DCMD) that integrates reconstruction and prediction is proposed for video anomaly detection based on skeletal poses.
Dual-calibrated Co-training Framework for Personalized Federated Semi-Supervised Medical Image Segmentation
Delin Pan (Jiangnan University), Xiang Pan (Jiangnan University)
CodeSegmentationFederated LearningImageBiomedical Data
π― What it does: A personalized federated semi-supervised learning framework is proposed for medical image segmentation tasks, utilizing dual calibration co-training to enhance model performance.
π― What it does: A dual-channel interactive graph Transformer (DigTraffic) is proposed for traffic classification, utilizing a message-level interactive graph (MTIG) to construct packet-level nodes and designing three types of heterogeneous edges. It combines dual-channel encoding of packet length and timing, and incorporates centrality, spatial, and edge encoding in the Transformer to capture global structural information.
π― What it does: The DualDynamics framework is proposed, which integrates implicit models based on NDE with explicit models of reversible neural flows to handle irregular and missing time series.
π― What it does: In the text-to-image diffusion model, a module is proposed that only modifies the skip connection features, achieving precise elimination of target concepts while maintaining the generation quality of non-target concepts.
DuSSS: Dual Semantic Similarity-Supervised Vision-Language Model for Semi-Supervised Medical Image Segmentation
Qingtao Pan (Shandong University), Shuo Li (Case Western Reserve University)
CodeSegmentationTransformerVision Language ModelContrastive LearningImageBiomedical DataComputed Tomography
π― What it does: Combining visual-language models with semi-supervised medical image segmentation, the DuSSS method is proposed, which utilizes dual contrastive learning and semantic similarity supervision to enhance cross-modal consistency. It also improves the quality of pseudo-labels through text-guided pseudo-label generation and a teacher-student framework, ultimately achieving more accurate semi-supervised segmentation.
Jun Zhu (Tsinghua University), Cheng Chang (National Center for Protein Sciences)
CodeRestorationGraph Neural NetworkAuto EncoderBiomedical Data
π― What it does: A denoising method for spatial transcriptomics (SRT) called DUSTED is proposed, which can restore high-quality spatial transcriptomic data without using external images.
DutyTTE: Deciphering Uncertainty in Origin-Destination Travel Time Estimation
Xiaowei Mao (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
CodeOptimizationTransformerReinforcement LearningMixture of ExpertsTime SeriesSequential
π― What it does: Developed the DutyTTE method, which first optimizes OD path prediction using deep reinforcement learning, and then quantifies segment-level travel time uncertainty through Mixture of Experts.
π― What it does: This paper proposes the dyAb framework, which utilizes the pre-bound antigen structures predicted by AlphaFold2 and completes the antibody design and binding process modeling through coarse-grained interface alignment and fine-grained flow matching.
π― What it does: This paper proposes a dynamic adapter and semantic decoupling method for cross-language and cross-modal retrieval, aiming to achieve alignment between visual data and low-resource languages without the need for labeled data in the target language.
π― What it does: This paper proposes a convolutional network DCCNeXt based on global clustering, which utilizes dynamic clustering convolution (DCConv) to group image patches into semantically similar clusters and extracts features using shared convolutional kernels.
Dynamic Entity-Masked Graph Diffusion Model for Histopathology Image Representation Learning
Zhenfeng Zhuang (Xiamen University), Liansheng Wang (Xiamen University)
CodeClassificationRepresentation LearningGraph Neural NetworkDiffusion modelAuto EncoderImageBiomedical Data
π― What it does: A self-supervised pathological image representation learning method based on a dynamic entity masking graph diffusion model (H-MGDM) is proposed, which captures the topological relationships of tissue entities using graph structures and reconstructs masked subgraphs through graph diffusion.
π― What it does: A Dynamic Expansion Diffusion Model (DEDM) has been designed and implemented, capable of continuously adding new diffusion components in an online continual learning scenario without task boundaries, while retaining memory of learned knowledge without catastrophic forgetting.
π― What it does: A Dynamic Multi-Interest Graph Neural Network (DMI-GNN) is proposed for conversational recommendation, modeling conversational data as a graph and extracting multi-interest representations.
CodeAnomaly DetectionGraph Neural NetworkGraphFinance Related
π― What it does: This paper proposes a dynamic spectral graph anomaly detection framework called DSGAD, which improves upon traditional manually designed wave functions and feature concatenation methods by using learnable Beta-mixed wave functions and channel convolution fusion.
Dynamic Syntactic Feature Filtering and Injecting Networks for Cross-lingual Dependency Parsing
Jianjian Liu (Kunming University of Science and Technology), Shengxiang Gao (Kunming University of Science and Technology)
CodeRecurrent Neural NetworkTransformerLarge Language ModelText
π― What it does: This paper proposes a dynamic syntactic feature filtering and injection network based on a shared-private framework for cross-lingual dependency parsing.
Dynamic-Width Speculative Beam Decoding for LLM Inference
Zongyue Qin (University of California Los Angeles), Yizhou Sun (University of California Los Angeles)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A dynamic width speculative beam search (DSBD) algorithm is proposed, which combines speculative decoding with beam search to improve the inference efficiency and output quality of large language models.
Each Fake News Is Fake in Its Own Way: An Attribution Multi-Granularity Benchmark for Multimodal Fake News Detection
Hao Guo (Nation University of Defense Technology), Xiang Zhao (Nation University of Defense Technology)
CodeClassificationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
π― What it does: This paper proposes the task of attribution and detection of multimodal fake news, constructs the first fine-grained attribution dataset AMG, and designs a multi-granularity clue alignment model MGCA.
π― What it does: This paper proposes the Prediction Uncertainty Index (PU-index) and its concept drift detector based on Chi-square test, PUDD, and implements online drift detection through an adaptive binning algorithm.
π― What it does: The EBBS (Ensemble with Bi-Level Beam Search) method is proposed to improve translation quality in multilingual zero-shot machine translation by combining direct translation with various pivot paths, and further utilizing the high-quality translations generated by EBBS for knowledge distillation to enhance model inference efficiency.
EBS-CFL: Efficient and Byzantine-robust Secure Clustered Federated Learning
Zhiqiang Li (East China Normal University), Hong-Ning Dai
CodeFederated LearningSafty and PrivacyImage
π― What it does: Designed and implemented an efficient, Byzantine robust secure aggregation framework EBS-CFL to protect user clustering identity and gradient privacy in cluster federated learning.
π― What it does: The EchoDiffusion framework is proposed, which encodes the acoustic fingerprint spectrum into a latent space and uses sound waveforms to guide the diffusion process to generate depth maps.
π― What it does: A new optimal priority push search algorithm called ECO SEARCH is proposed, which can achieve constant delay enumeration in program synthesis.
EDENet: Echo Direction Encoding Network for Place Recognition Based on Ground Penetrating Radar
Pengyu Zhang (National University of Defense Technology), Liang Shen (National University of Defense Technology)
CodeRecognitionRetrievalConvolutional Neural NetworkTime Series
π― What it does: Proposes the EDENet network, which utilizes learnable Gabor filters and direction-aware attention to geometrically encode ground-penetrating radar (GPR) echo sequences, generating compact local recognition descriptors.
π― What it does: Without using any graph data augmentation, the AFECL (Augmentation-Free Edge Contrastive Learning) model is proposed, which generates edge features using node embeddings and performs contrastive learning at the edge level.
π― What it does: EditBoard is proposed, providing a comprehensive evaluation benchmark that includes nine automatic metrics and four assessment dimensions (fidelity, execution, consistency, style) for the systematic evaluation of text-driven video editing models.
Wei Zhou (Huazhong University of Science and Technology), Fei Wang (Ping An Property and Casualty Insurance Company of China)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper achieves local editing of factual knowledge in large language models by identifying only about 1% of high-contribution neurons in the model, fine-tuning them, and performing data augmentation with samples that share the same knowledge relationships.
Effective and Efficient Representation Learning for Flight Trajectories
Shuo Liu (University of Chinese Academy of Sciences), Jingping Bi (Chinese Academy of Sciences)
CodeAnomaly DetectionRepresentation LearningTransformerTime Series
π― What it does: The FLIGHT2VEC framework is proposed for unified representation learning of flight trajectories, addressing issues of uneven behavior density and 3D spatial continuity.
π― What it does: This paper proposes a diffusion Transformer architecture called DiT-SR, which is trained from scratch. It utilizes a U-shaped full transformer to achieve multi-scale feature extraction and enhances super-resolution quality through frequency-adaptive time step conditioning.
π― What it does: This paper proposes an event-driven Sparse Spiking Convolution (SSC) and Spike Voxel Coding (SVC) aimed at sparse 3D point clouds, integrating them into the E-3DSNN backbone network to efficiently handle 3D classification, detection, and segmentation tasks.
π― What it does: A lightweight Transformer framework ASID is proposed, achieving single image super-resolution through information distillation structure, attention sharing, and channel splitting.
Efficient Connectivity-Preserving Instance Segmentation with Supervoxel-Based Loss Function
Anna Grim (Allen Institute), Uygar SΓΌmbΓΌl
CodeSegmentationComputational EfficiencyImageBiomedical Data
π― What it does: This paper proposes a topology-aware loss function based on supervoxels to efficiently maintain the connectivity of neuron instance segmentation.
π― What it does: A hybrid ANN-SNN framework is proposed to fuse the sparse asynchronous data from event cameras with the spatial information of frame images, achieving efficient semantic segmentation.
π― What it does: This paper proposes an indoor depth completion network based on Mask-adaptive Gated Convolution (MagaConv) and Bidirectional Aligned Projection (Bid-AP) to fill in the missing areas of depth maps generated by sensors such as TOF and structured light.
π― What it does: A lightweight residual network combined with reversible flow is used to compress hundreds of 3D color lookup tables into a unified model of less than 0.25 MB, while keeping color distortion during reconstruction to no more than ΞE 2.
Efficient Reinforcement Learning in Probabilistic Reward Machines
Xiaofeng Lin (Boston University), Xuezhou Zhang (Boston University)
CodeReinforcement Learning
π― What it does: This paper studies reinforcement learning in Markov decision processes with probabilistic reward machines (PRM), proposing the UCBVI-PRM algorithm and providing an approximate optimal regret upper bound.
π― What it does: Proposes an attack method to accelerate model robustness evaluation through dynamic relaxation of perturbation constraints (Constraint Relaxation Attack, CR Attack).
π― What it does: A self-supervised video hashing method S5VH based on the Mamba state space model is proposed, utilizing bidirectional Mamba layers and a self-local-global learning strategy to achieve efficient temporal modeling and hash code generation.
π― What it does: An adaptive backpropagation method based on reverse solving enhanced fractional differential equations (FDE) is proposed, significantly reducing the memory consumption during the training process of neural FDE while maintaining the same performance as traditional forward automatic differentiation methods.
Efficiently Enhancing Long-term Series Forecasting via Ultra-long Lookback Windows
Suxin Tong (Wuhan University of Technology), Jingling Yuan (Wuhan University of Technology)
CodeOptimizationComputational EfficiencyTransformerTime Series
π― What it does: The IRPA framework is proposed and implemented, which extracts key information from an ultra-long lookback window through the Input Refinement Module (IRM) and the Prediction Assistance Module (PAM), and uses this information to enhance the accuracy of long-period time series forecasting.
π― What it does: A lightweight visual model called EfficientVMamba is proposed, which combines state space models (SSM) with convolution, and achieves efficient extraction of global and local features through sparse scanning (ES2D) and dual-channel fusion.
π― What it does: This study addresses the single hyperspectral image super-resolution (single-HSI-SR) problem and proposes a new framework called EigenSR. This framework first transfers a pre-trained RGB Transformer (IPT) to the feature map (eigenimage) domain for single-channel detail learning, and then utilizes Iterative Spectral Regularization (ISR) to restore spectral consistency during inference, thereby enhancing SR performance in both spatial and spectral dimensions simultaneously.
π― What it does: A video restoration framework RGVI is proposed, which combines optical flow-guided dual pixel propagation and large-scale diffusion models, enabling high-quality object removal and content generation.
Eliciting Causal Abilities in Large Language Models for Reasoning Tasks
Yajing Wang (BNU-HKBU United International College), Bo Han (Hong Kong Baptist University)
CodeOptimizationExplainability and InterpretabilityMeta LearningTransformerLarge Language ModelPrompt EngineeringText
π― What it does: By treating prompts as processing variables, this study uses causal inference to estimate their impact on the accuracy of reasoning results from large language models, and based on this, generates better prompts to enhance the model's performance on reasoning tasks.
π― What it does: A zero-shot text-to-speech generation framework based on a language model, ELLA-V, is proposed, utilizing an inserted phoneme-acoustic interleaved sequence for fine-grained phoneme-level control.
Empowering Self-Learning of LLMs: Inner Knowledge Explicitation as a Catalyst
Shijue Huang (Harbin Institute of Technology), Ruifeng Xu (Harbin Institute of Technology)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: The SKE-Learn framework is proposed, which enables LLM to explicitly extract intrinsic knowledge, verify it, and utilize this knowledge for reasoning, thereby achieving reliable self-learning data filtering.
End-to-End Autonomous Driving Through V2X Cooperation
Haibao Yu (University of Hong Kong), Zaiqing Nie (AIR Tsinghua University)
CodeAutonomous DrivingTransformerPoint Cloud
π― What it does: Designed and implemented an end-to-end V2X collaborative autonomous driving framework called UniV2X, which integrates perception, online mapping, occupancy prediction, and planning modules into a single network, and achieves cross-view fusion through sparse-dense mixed data transmission.
Energy vs. Noise: Towards Robust Temporal Action Localization in Open-World
Chenyu Mu (Xidian University), Cheng Deng (Xidian University)
CodeOptimizationMeta LearningVideo
π― What it does: This paper proposes the Energy-Driven Meta Purifier (EDMP), which utilizes an energy-driven meta-learning framework to remove boundary and category noise in Temporal Action Localization (TAL), enhancing the model's robustness against open-world noise.
π― What it does: This paper proposes a training-free and inversion-free energy-guided optimization framework for personalized image editing, which gradually optimizes the latent code of the target image under the guidance of text and image energy, achieving cross-category object replacement.
π― What it does: This paper proposes a CLIP fine-tuning method based on positive incentive noise (PiNI), which enhances visual-language alignment by injecting learned noise into the visual and text encoders, thereby achieving better performance on downstream tasks.
Enhanced Denesity Peak Clustering for High-Dimensional Data
Zhongli Wang (Hangzhou Normal University), Weiguo Sheng (University of Technology Sydney)
CodeOptimizationSupervised Fine-TuningTabularBiomedical Data
π― What it does: An Enhanced Density Peak Clustering (EDPC) method is proposed, which combines dimensionality reduction using a multilayer perceptron and hierarchical label assignment, significantly improving clustering performance on high-dimensional data.
π― What it does: This study proposes performing importance sampling in the normalized flow latent space to improve the efficiency of rare event simulation.
π― What it does: A sample selection method based on confidence tracking is proposed, which can identify correct but hard-to-learn samples in image classification tasks with noisy labels.
Enhancing Chain of Thought Prompting in Large Language Models via Reasoning Patterns
Yufeng Zhang (Institute of Automation, Chinese Academy of Sciences), Jinqiao Wang (Institute of Automation, Chinese Academy of Sciences)
CodeExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: A chain-of-thought demonstration selection method based on reasoning patterns, Pattern-CoT, is proposed to enhance the reasoning performance of large language models.
π― What it does: This paper improves the quality of near-field synthesis from viewpoints far from the training perspective through a pseudo-labeling learning strategy, particularly focusing on detail reconstruction at close distances.
π― What it does: Two methods, JointCrop and JointBlur, are proposed to control the augmentation parameters of positive sample pairs (the area ratio of Crop and the degree of Gaussian Blur) using joint distribution, generating more challenging positive sample pairs, and unifying them into a JointAugmentation framework.
Enhancing Elusive Clues in Knowledge Learning by Contrasting Attention of Language Models
Jian Gao (Tsinghua University), Ji Wu (Tsinghua University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: By comparing the attention weights of large models and small models, important clues that are difficult to capture in the text are identified, and data augmentation based on token-dropout is conducted to enhance knowledge learning efficiency.
π― What it does: Using distance geometry constraints (METRIZATION) for pre-training diffusion generative models and fine-tuning on real data, we propose the Metrization-Informed Geometric Diffusion (MIGDIFF) framework to enhance the generalization performance of molecular conformation generation.
Enhancing Generalized Few-Shot Semantic Segmentation via Effective Knowledge Transfer
Xinyue Chen (King's College London), Sophia Tsoka (King's College London)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: A generalized few-shot semantic segmentation method GFSS-EKT based on effective knowledge transfer is proposed, which achieves the transfer of base class knowledge to new classes through three main modules: prototype modulation, classifier calibration, and context consistency learning.
Enhancing Healthcare Recommendations: A Privacy-Protective and Interpretable Cross-Domain Framework
Xun Liang (Renmin University of China), Hongxun Jiang (Renmin University of China)
CodeRecommendation SystemSafty and PrivacyExplainability and InterpretabilityGraph Neural NetworkLarge Language ModelVideoTextMultimodality
π― What it does: A cross-domain recommendation framework (HCR) for healthcare services that is interpretable and privacy-preserving has been constructed.
π― What it does: This paper proposes a symmetric power transformation for the input data of implicit neural representations (INR) to enhance their expressive capability and accelerate training;
Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions
Yu Long (Beijing Institute of Technology), Yuming Fang (Jiangxi University of Finance and Economics)
CodeRestorationData SynthesisImage
π― What it does: A low-light image synthesis pipeline from RAW inverse ISP to sRGB is proposed, which can automatically generate an unlimited amount of aligned low-light-normal light paired data.
Enhancing Multi-Hop Fact Verification with Structured Knowledge-Augmented Large Language Models
Han Cao (Institute of Information Engineering, Chinese Academy of Sciences), Songlin Hu (Institute of Information Engineering, Chinese Academy of Sciences)
CodeGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextGraph
π― What it does: A structured knowledge-enhanced LLM network (LLM-SKAN) is proposed for multi-hop fact verification, where LLM first extracts fine-grained entity relationships and then a graph neural network integrates reasoning.
π― What it does: A large-scale multilingual multimodal live comment dataset LCAffect was constructed, and a video encoder was trained using contrastive learning to generate synthetic live comment features for videos without comments, aimed at enhancing multimodal sentiment analysis tasks.
Enhancing Multimodal Large Language Models Complex Reason via Similarity Computation
Xiaofeng Zhang (Shanghai Jiaotong University), Jiawei Yao (University of Washington)
CodeTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: By calculating the similarity between image and text embeddings, relevant image tokens are filtered to enhance the complex reasoning performance of multimodal large language models.
π― What it does: A GILD module is proposed, which adaptively learns a general imitation learning objective through meta-learning using offline demonstration data, thereby improving the online training effectiveness of offline RL in sparse reward environments.
π― What it does: A cross-domain Portuguese variant identification dataset PtBrVarId was constructed, and based on this, the BERTimbau fine-tuning model was used to distinguish between European Portuguese and Brazilian Portuguese.
Enhancing Question Generation through Diversity-Seeking Reinforcement Learning with Bilevel Policy Decomposition
Tianyu Ren (Queen's University Belfast), Karen Rafferty (Queen's University Belfast)
CodeGenerationTransformerLarge Language ModelReinforcement LearningText
π― What it does: The BPD-DSRL framework is proposed, utilizing a dual-layer strategy decomposition and a reinforcement learning objective for diversity seeking to enhance the sample efficiency and diversity of question generation.
Enhancing Relation Extraction via Supervised Rationale Verification and Feedback
Yongqi Li (Wuhan University), Tieyun Qian (Wuhan University)
CodeData-Centric LearningTransformerLarge Language ModelContrastive LearningText
π― What it does: An automatic feedback framework SRVF based on large language models (LLM) is proposed, which can correct biased predictions in relation extraction (RE) by validating the reasoning process and providing examples for re-selection.
π― What it does: This paper proposes GlobalDiff, a Plug-and-Play framework that overlays diffusion models on existing sequential recommendation models. It compensates for local sequential information by restoring global non-sequential data structures, thereby improving the prediction accuracy of the next item.
Enhancing SQL Query Generation with Neurosymbolic Reasoning
Henrijs Princis (University of Cambridge), Alan Mycroft (University of Cambridge)
CodeGenerationAI Code AssistantTransformerLarge Language ModelText
π― What it does: A neurosymbolic architecture called Xander is proposed, which combines symbolic reasoning and pre-trained language models to achieve SQL generation, supporting multi-path exploration, backtracking, and query repair.
π― What it does: This paper proposes the incorporation of an autoencoder for manifold projection in adversarial training and adversarial distillation to enhance model robustness.
π― What it does: A trainable ranking-based adaptive conformal prediction framework RCP-GNN is proposed for the graph node classification task to achieve controllable improvements in boundary coverage and prediction set efficiency.
π― What it does: A framework for variational information utilization based on the entire space (EVI) is proposed to enhance CVR prediction accuracy through unbiased pseudo-labels and variational information maximization.
π― What it does: This paper proposes an entropy regularization task representation learning method called ER-TRL, which utilizes GAN to approximate the entropy of the meta-behavior policy to reduce the context distribution shift in offline meta reinforcement learning, thereby enhancing adaptability and generalization ability on new tasks.
CodeSegmentationComputational EfficiencyTransformerVision Language ModelImage
π― What it does: Proposes a single-stage shared efficient spatial awareness framework EOV-Seg to address the computational overhead and speed bottleneck of open vocabulary panoptic segmentation.
Pascal R. van der Vaart (Delft University of Technology), Neil Yorke-Smith (Delft University of Technology)
CodeReinforcement Learning
π― What it does: This paper proposes Epistemic Bellman Operators (EBO) to unify and theorize various Bayesian-based uncertainty RL algorithms, proving that they are contraction mappings and converge; it uses this framework to analyze and improve Bayesian Q-learning and designs an uncertainty-aware variant of PPO (ECPPO).
π― What it does: This paper proposes EquiDetect, a framework for 3D object detection using LiDAR-camera multimodal fusion under adverse weather conditions.
π― What it does: A two-stage graph unlearning method called ETR is proposed, which first eliminates the target samples and their propagation effects through parameter editing, and then corrects the model performance using subgraph gradient approximation.
Error Diversity Matters: An Error-Resistant Ensemble Method for Unsupervised Dependency Parsing
Behzad Shayegh (University of Alberta), Lili Mou (University of Alberta)
CodeText
π― What it does: A multi-model ensemble for unsupervised dependency parsing has been constructed, and a social entropy-based error diversity-driven ensemble selection method has been proposed.
π― What it does: This paper proposes the ESEG framework, which utilizes the motion edge characteristics of event cameras, introduces explicit semantic edge supervision, and enhances event-driven semantic segmentation performance through a multi-layer fusion module.
π― What it does: This paper presents ESPRESSO, a retrieval module specifically designed for conversational recommendation systems (CRS), aimed at enhancing the authenticity and information richness of recommendation responses by retrieving paragraphs that match user preferences.
π― What it does: This paper proposes an event camera-assisted blurry video super-resolution (BVSR) method called Ev-DeblurVSR, which can recover high-resolution, clear videos from low-resolution and blurry inputs.
EventSum: A Large-Scale Event-Centric Summarization Dataset for Chinese Multi-News Documents
Mengna Zhu (National University of Defense Technology), Juanzi Li (Tsinghua University)
CodeGenerationTransformerLarge Language ModelText
π― What it does: This paper proposes and implements the Event-Centric Multi-Document Summarization (ECS) task, which aims to automatically generate concise summaries that cover the core sub-events, time, location, people, and causal relationships of multiple related news articles.
π― What it does: An evaluation protocol based on connected components (CC-Metrics) is proposed, which calculates traditional semantic segmentation metrics (Dice, Hausdorff, Surface Dice, etc.) locally for each lesion, eliminating bias towards lesion size.
Everywhere Attack: Attacking Locally and Globally to Boost Targeted Transferability
Hui Zeng (Southwest University of Science and Technology), Anjie Peng (Southwest University of Science and Technology)
CodeAdversarial AttackImage
π― What it does: Proposes and implements the 'everywhere attack', which significantly enhances the cross-model transferability of targeted attacks by simultaneously optimizing attack targets in both global and multiple local regions of the image.
EvoChart: A Benchmark and a Self-Training Approach Towards Real-World Chart Understanding
Muye Huang (Xi'an Jiaotong University), Jun Liu (Xi'an Jiaotong University)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodalityBenchmark
π― What it does: Developed the EvoChart method and the EvoChart-QA benchmark, generating high-quality synthetic chart data through multi-stage self-training and training chart understanding models;
Evolutionary Large Language Model for Automated Feature Transformation
Nanxu Gong (Arizona State University), Yanjie Fu (Arizona State University)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTabular
π― What it does: A novel evolutionary large language model framework (ELLM-FT) is proposed for automatic feature transformation, combining reinforcement learning to generate various population data and a small number of examples from LLM to create better feature combinations.
π― What it does: This paper proposes the EvSTVSR method, which utilizes the high temporal resolution of event cameras and a small number of RGB frames to achieve spatial-temporal video super-resolution.
EWMoE: An Effective Model for Global Weather Forecasting with Mixture-of-Experts
Lihao Gan (University of Electronic Science and Technology of China), Jie Shao (University of Electronic Science and Technology of China)
CodeTransformerMixture of ExpertsTime Series
π― What it does: The EWMoE model is proposed for global weather forecasting, achieving high-accuracy predictions using only two years of ERA5 data while significantly reducing training resources.
EXCGEC: A Benchmark for Edit-Wise Explainable Chinese Grammatical Error Correction
Jingheng Ye (Tsinghua University), Wenhao Jiang (Sun Yat-Sen University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: The EXGEC task is proposed, and the EXCGEC Chinese interpretable grammar error correction benchmark is constructed, providing editing-style explanations and training and evaluation of multi-task models.