π― What it does: A dual-branch arithmetic regression reasoning framework (DARR) is proposed to address the machine number reasoning (MNR) task, which involves identifying candidate answer images that follow the same arithmetic rules among three contextual images.
π― What it does: The DASK method is proposed, utilizing distribution replay and adaptive style kernel learning to achieve zero-shot lifelong person re-identification, avoiding the storage of historical data.
Data Augmentation for Instruction Following Policies via Trajectory Segmentation
Niklas Hoepner (University of Amsterdam), Herke van Hoof (Vrije Universiteit Amsterdam)
CodeSegmentationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningVideo
π― What it does: Using a limited number of labeled sub-trajectories to segment unlabeled long trajectories and assign instructions to each segment, thereby providing more training samples for instruction-following strategies.
π― What it does: This paper proposes FedZGE, a data-independent, black-box federated learning framework that trains a generator on the server side to generate synthetic data for knowledge distillation, without the need to share model parameters or auxiliary data throughout the process.
π― What it does: A data-free universal adversarial perturbation generation method called IntriUAP is proposed, which utilizes the inherent vulnerability of linear layers in deep networks to generate attack perturbations.
π― What it does: A Dual-Codebook Point Completion Network (DCβPCN) is proposed, achieving complete reconstruction of point clouds through an encoder-decoder architecture.
DCA: Dividing and Conquering Amnesia in Incremental Object Detection
Aoting Zhang (Institute of Information Engineering Chinese Academy of Sciences), Yu Zhou (Nankai University)
CodeObject DetectionKnowledge DistillationTransformerLarge Language ModelImage
π― What it does: This paper proposes an incremental object detection method based on Transformer, called DCA, which separates localization from recognition and utilizes the semantic features of a pre-trained language model to guide recognition, while employing a dual classifier fusion to alleviate recognition forgetting.
DCC: Differentiable Cardinality Constraints for Partial Index Tracking
Wooyeon Jo (Ajou University), Hyunsouk Cho (Ajou University)
CodeOptimizationComputational EfficiencyTime SeriesFinance Related
π― What it does: This study investigates how to solve the partial replication problem in index tracking using differentiable cardinality constraints.
DCILP: A Distributed Approach for Large-Scale Causal Structure Learning
Shuyu Dong (INRIA), Koji Maruhashi (Fujitsu Limited)
CodeGraph Neural NetworkGraphTabularBiomedical Data
π― What it does: A divide-and-conquer causal structure learning framework DCILP is proposed, which first estimates the Markov blanket of each variable and solves local subproblems in parallel, and then unifies these local results using integer linear programming (ILP) to obtain a global causal graph.
DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection
Tao Dai (Shenzhen University), Zexuan Zhu (Shenzhen University)
CodeObject DetectionKnowledge DistillationImage
π― What it does: A dynamic channel-space feature distillation framework DCSF-KD is proposed for knowledge distillation in object detection, which utilizes channel weights to dynamically extract spatial features for distillation.
π― What it does: In the JPEG image restoration task, the DCTMamba framework is proposed, which combines discrete cosine transform with Mamba to achieve causal scanning from low frequency to high frequency.
De-singularity Subgradient for the q-th-Powered lβ-Norm Weber Location Problem
Zhao-Rong Lai (Jinan University), Cheng Li (Jinan University)
CodeOptimizationTime SeriesFinance Related
π― What it does: A de-singular subgradient method is proposed for the q-th powered βp-norm Weber location problem for 1 β€ p < 2 and 1 β€ q β€ p, along with the corresponding q-P-NWAWS algorithm.
CodeClassificationRecommendation SystemGraph Neural NetworkTransformerLarge Language ModelTabularBiomedical DataElectronic Health Records
π― What it does: This paper proposes a framework named DearLLM, which utilizes large language models (LLM) to derive quantitative associations between diagnostic codes in the context of personalized medicine, constructs a feature prior graph, and enhances EHR prediction performance through graph convolutional networks (GCN) and frequency-aware graph pooling.
Debate on Graph: A Flexible and Reliable Reasoning Framework for Large Language Models
Jie Ma (Xi'an Jiaotong University), Lizhen Cui (Northwestern Polytechnical University)
CodeTransformerLarge Language ModelPrompt EngineeringTextGraph
π― What it does: An iterative interactive framework named DoG is proposed, which utilizes large language models for subgraph focusing and multi-role debate on knowledge graphs, gradually reasoning and generating answers.
Debiased All-in-one Image Restoration with Task Uncertainty Regularization
Gang Wu (Harbin Institute of Technology), Xianming Liu (Harbin Institute of Technology)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: An adaptive multi-task regularization method based on task uncertainty (Task Uncertainty Regularization, TUR) is proposed, which automatically learns task weights during joint training of all tasks and improves recovery quality.
Debiased Distillation for Consistency Regularization
Lu Wang (Northeastern University), Jun Cheng (Shenzhen Institute of Advanced Technology)
CodeClassificationKnowledge DistillationImage
π― What it does: This paper proposes a new knowledge distillation strategyβIntra-Class Knowledge Distillation (IKD), which achieves prediction consistency by sharing the average logits within the same class, reducing noise and bias, and enhancing the generalization ability of the student model.
Decentralized Federated Learning with Model Caching on Mobile Agents
Xiaoyu Wang (New York University), Yong Liu (New York University)
CodeFederated LearningImage
π― What it does: A decentralized federated learning framework called Cached-DFL is proposed, which utilizes D2D communication and caching mechanisms between mobile agents to achieve delay-tolerant model propagation and aggregation.
π― What it does: By constructing a difference-aware contrastive diffusion model and incorporating adversarial perturbations, natural language descriptions of subtle changes between two similar images are achieved.
Decomposing and Fusing Intra- and Inter-Sensor Spatio-Temporal Signal for Multi-Sensor Wearable Human Activity Recognition
Haoyu Xie (Chongqing University), Li Liu (Chongqing University)
CodeRecognitionConvolutional Neural NetworkSupervised Fine-TuningMultimodalityTime Series
π― What it does: A decomposition and fusion framework called DecomposeWHAR is proposed to better capture the internal and external spatiotemporal relationships in multi-sensor wearable human action recognition.
π― What it does: A frame-free annotation self-supervised visual tracking framework called SSTrack is developed, which learns cross-frame object representations through separated spatiotemporal consistency training and instance contrastive loss.
Decoupling Metacognition from Cognition: A Framework for Quantifying Metacognitive Ability in LLMs
Guoqing Wang (East China Normal University), Hong Zheng (East China Normal University)
CodeTransformerLarge Language ModelText
π― What it does: Proposed and validated the DMC framework for quantifying the metacognitive abilities of large language models (LLMs) and decoupling them from cognitive abilities.
π― What it does: In scattering medium (fog/underwater) scenarios, a pseudo-label guided NeRF model is proposed, combined with a cyclic progressive dimension optimization strategy (CPDOS) to achieve the separation and reconstruction of object and medium density.
Deep Evidential Hashing for Trustworthy Cross-Modal Retrieval
Yuan Li (Sichuan University), Peng Hu
CodeRetrievalMultimodality
π― What it does: This paper proposes a Deep Evidential Cross-Modal Hashing (DECH) method that utilizes evidence learning to directly generate binary codes and provides reliability assessment for retrieval results.
π― What it does: This paper proposes a Transformer-based deep generative model called GearFormer, which is used to generate configurations of gear transmission systems that meet multiple design constraints in a single pass, and further combines it with traditional search methods (EDA, MCTS) to form a hybrid design scheme.
Deep Graph Online Hashing for Multi-Label Image Retrieval
Yuan Cao (Ocean University of China), Jie Gui (Southeast University)
CodeRetrievalGraph Neural NetworkImage
π― What it does: This paper proposes and implements Deep Graph Online Hashing (DGOH), achieving online deep hashing in multi-label image retrieval scenarios while balancing retrieval accuracy and training efficiency.
CodeRecurrent Neural NetworkReinforcement LearningTime Series
π― What it does: This paper embeds a scale-invariant memory model based on computational neuroscience into deep reinforcement learning agents and validates its effectiveness on various interval timing and memory tasks.
Deep Submodular Optimization and LLM for Multimodal Content Extraction and Automatic Poster Generation from Long Document
Vijay Jaisankar (International Institute of Information Technology), Shwetha Somasundaram (Adobe Research)
CodeGenerationOptimizationTransformerLarge Language ModelImageTextMultimodality
π― What it does: This paper presents an end-to-end system called PostDoc, which automatically converts long multimodal documents (text + images) into single-page posters, covering content extraction, LLM paraphrasing, and template generation.
Siddharth Baskar (Wisconsin Institute for Discovery), Daniel L. Pimentel-AlarcΓ³n
CodeRestorationAuto EncoderTabular
π― What it does: This paper proposes a Deep Union Completion (DUC) framework that utilizes a pseudo-completion layer to reconstruct data with a high proportion of missing values, and simultaneously performs missing value imputation and subspace clustering through an autoencoder and a closed clustering layer.
DeepSN: A Sheaf Neural Framework for Influence Maximization
Asela Hevapathige (Australian National University), Ahad N. Zehmakan (Australian National University)
CodeOptimizationGraph Neural NetworkGraph
π― What it does: A deep neural framework called DeepSN based on sheaf theory is proposed for learning influence diffusion models and optimizing seed selection to achieve maximum influence.
Defending Against Sophisticated Poisoning Attacks with RL-based Aggregation in Federated Learning
Yujing Wang (Beihang University), Binghui Guo (Beihang University)
CodeFederated LearningReinforcement LearningImage
π― What it does: Proposes an adaptive aggregation method AdaAggRL based on reinforcement learning to defend against advanced poisoning attacks targeting servers.
Defense Against Model Stealing Based on Account-Aware Distribution Discrepancy
Jian-Ping Mei (Zhejiang University of Technology), Tiantian Zhu (Macquarie University)
CodeClassificationAnomaly DetectionImage
π― What it does: A non-parametric detector based on Account-Aware Distribution Difference (ADD) is proposed, which is combined with random label poisoning to form an end-to-end real-time defense module D-ADD, aimed at preventing model stealing attacks on image classification models.
DELTA: Pre-Train a Discriminative Encoder for Legal Case Retrieval via Structural Word Alignment
Haitao Li (Tsinghua University), Yiqun Liu (Tsinghua University)
CodeRetrievalTransformerContrastive LearningText
π― What it does: A pre-training framework called DELTA based on structural word alignment is proposed to enhance the discriminative ability of legal case retrieval.
DeMo: Decoupled Feature-Based Mixture of Experts for Multi-Modal Object Re-Identification
Yuhao Wang (Dalian University of Technology), Pingping Zhang (Dalian University of Technology)
CodeRecognitionRetrievalTransformerMixture of ExpertsMultimodality
π― What it does: The DeMo framework is proposed, which first uses the Patch-Integrated Feature Extractor (PIFE) to extract multi-scale features, then decouples the multi-modal features hierarchically through the Hierarchical Decoupling Module (HDM), and finally applies the Attention-Triggered Mixture of Experts (ATMoE) to dynamically weight the decoupled features for multi-modal object ReID.
DeMo: Deep Motion Field Consensus with Learnable Kernels for Two-view Correspondence Learning
Yifan Lu (Wuhan University), Jiayi Ma (Chinese University of Hong Kong)
CodePose EstimationOptimizationSimultaneous Localization and MappingOptical FlowImage
π― What it does: The DeMo network is proposed for outlier removal in two-view correspondence learning, utilizing global motion consensus to enhance matching quality.
DeNC: Unleash Neural Codecs in Video Streaming with Diffusion Enhancement
Qihua Zhou (Shenzhen University), Song Guo (Hong Kong University of Science and Technology)
CodeGenerationCompressionDiffusion modelVideo
π― What it does: This paper proposes the Diffusion-enhanced Neural Codec (DeNC), which simultaneously reduces frame resolution and color bit depth through the encoder, while the decoder employs a diffusion model for perceptually guided restoration, achieving a balance among rate, distortion, and perception at extremely low bit rates for video streaming.
π― What it does: The DETRIS framework is proposed, utilizing a Dense Aligner and text adapter to achieve parameter-efficient fine-tuning for the Referring Image Segmentation task.
Densest k-Subgraph Mining via a Provably Tight Relaxation
Qiheng Lu (University of Virginia), Aritra Konar (KU Leuven)
CodeOptimizationGraph Neural NetworkGraph
π― What it does: The paper addresses the Densest k-Subgraph problem and proposes a continuous relaxation obtained by diagonal loading (Ξ»=1) of the adjacency matrix, proving that it converges to an integer solution when the size of the subgraph does not exceed the size of the maximum clique.
π― What it does: This paper proposes a Matrix Graph Neural Network (MGNN) that directly constructs a matrix graph using the 2D feature maps of each frame in variable-length videos, enabling end-to-end learning of depression-related spatiotemporal features in full-length videos.
π― What it does: This paper proposes a deep center learning framework that jointly utilizes atmospheric scattering models and brightness consistency constraints to achieve dehazing and depth estimation of monocular hazy videos.
π― What it does: This paper proposes DepthFM, which utilizes flow matching technology to achieve direct generation from images to depth maps, significantly reducing sampling steps.
DEQA: Descriptions Enhanced Question-Answering Framework for Multimodal Aspect-Based Sentiment Analysis
Zhixin Han (Nankai University), Bitong Luo (JD AI Research)
CodeClassificationTransformerLarge Language ModelPrompt EngineeringMixture of ExpertsImageTextMultimodality
π― What it does: This study proposes a multimodal emotion analysis framework named DEQA, which utilizes GPT-4 to generate image descriptions and inputs text, images, and descriptions into different experts separately, completing multimodal aspect-based sentiment analysis (MABSA) through a multi-round question-and-answer approach.
π― What it does: Achieve untrained background inpainting through key masking self-attention, and implement multi-object spatial editing within a multi-layer latent decomposition and fusion framework;
Designing Ambiguity Sets for Distributionally Robust Optimization Using Structural Causal Optimal Transport
Ahmad Reza Ehyaei, Samira Samadi (Max Planck Institute for Intelligent Systems)
CodeOptimizationTabular
π― What it does: This paper proposes Structural Causal Optimal Transport (SCOT) and its relaxed version to construct a distribution uncertainty set that better aligns with real causal structures, and provides efficient solving algorithms.
Designing Specialized Two-Dimensional Graph Spectral Filters for Spatial-Temporal Graph Modeling
Yuxin Chen (Nanjing University of Science and Technology), Hui Yan (Nanjing University of Science and Technology)
CodeGraph Neural NetworkGraphTime Series
π― What it does: A new spatio-temporal graph neural network (STSGNN) is proposed, which jointly models the interaction between spatial and temporal dimensions through a two-dimensional graph frequency domain filter;
π― What it does: This paper proposes a destruction-repair framework based on hypergraphs (DRHG) for searching large neighborhoods and learning the repair process in large routing problems (TSP, CVRP).
π― What it does: This paper proposes a complete framework for identifying and defending against convolutional-based unlearned examples (UEs). It first implements detection through edge pixel statistics (EPD) and then disrupts the convolutional noise distribution using random matrix multiplication (COIN), while also extending two new types of convolutional UEs (VUDA, HUDA).
CodeGenerationOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelVision Language ModelTextMultimodality
π― What it does: This paper proposes a framework for hallucination detection and mitigation based on fine-grained AI feedback for LVLM, which includes automatically generating sentence-level hallucination annotation data, a detect-then-rewrite preference data construction pipeline, and the HSA-DPO method that incorporates hallucination severity into Direct Preference Optimization (DPO);
π― What it does: Polytune is proposed, an end-to-end model based on Transformer that can directly align performance audio with score audio and label each note as Correct, Missed, or Extra, supporting multiple instruments without the need for traditional manual alignment methods such as Dynamic Time Warping (DTW).
π― What it does: A deterministic policy gradient primal-dual algorithm for constrained Markov decision processes with continuous state-action spaces is proposed, along with non-asymptotic convergence guarantees.
π― What it does: A video prediction network named DFDNet is proposed, which can decompose dynamics in the spatial direction and remove instantaneous high-frequency noise through a learnable threshold filter.
π― What it does: A cross-modal retrieval method named DiCA is proposed, specifically addressing the retrieval task when only a partial set of labels is given (i.e., the candidate label set contains the true label).
π― What it does: A global-guided diffusion model called Diff-Shadow is proposed for shadow removal without obvious boundaries and consistent lighting.
π― What it does: This paper proposes DiffCLIP, a high-dimensional multimodal remote sensing image classification framework that combines unsupervised mask diffusion pre-training with language-based few-shot classification.
π― What it does: A differentiable adversarial attack framework for Marked Time Point Processes (MTPP) called PERMTPP has been designed and implemented, which significantly reduces the performance of MTPP in time prediction and label prediction while keeping the perturbation low perceptible.
π― What it does: A differentiated alignment domain adaptation framework for object detection is proposed, designing a predicted difference feedback instance alignment module and an uncertainty foreground-guided image alignment module to achieve weighted alignment for different regions.
DiffIM: Differentiable Influence Minimization with Surrogate Modeling and Continuous Relaxation
Junghun Lee (Korea Advanced Institute of Science and Technology), Kijung Shin (Korea Advanced Institute of Science and Technology)
CodeOptimizationGraph Neural NetworkGraph
π― What it does: A method for impact minimization based on differentiable learning, DIFFIM, is proposed, which combines surrogate models and continuous relaxation to achieve efficient edge deletion.
π― What it does: Using a pre-trained text-to-image diffusion model, we propose a training-free progressive high-resolution image generation method called DiffuseHigh, which generates higher resolution images guided by low-resolution images.
Yingfan Wang (Duke University), Cynthia Rudin (Duke University)
CodeGraphTabular
π― What it does: A dynamic local adjustment graph dimensionality reduction algorithm called LocalMAP is proposed and implemented, which can identify and remove misclassified neighboring edges during the dimensionality reduction process, dynamically resampling distant edges to achieve a clearer cluster distribution.
π― What it does: This paper presents DisCo-DSO, a deep generative model that combines discrete and continuous variables to efficiently solve black-box mixed space optimization problems.
π― What it does: This study addresses the cold start cross-domain recommendation problem and proposes the DisCo framework to learn users' fine-grained intentions and filter out irrelevant collaborative information.
π― What it does: This paper proposes a parameterizable and differentiable Analytic Ontology Template (AOT) and the AOTNet based on it, aimed at recognizing and utilizing joint structure and operability information at the conceptual level to achieve perception and interaction with novel joint objects.
Xingcheng Fu (Guangxi Normal University), Xianxian Li (Beihang University)
CodeClassificationOptimizationExplainability and InterpretabilityGraph Neural NetworkGraph
π― What it does: This paper proposes a graph neural network framework called CurvGIB, which combines discrete Ricci curvature with the information bottleneck to learn the optimal information transmission structure relevant to tasks while simultaneously optimizing node representations.
π― What it does: A bundling recommendation framework DCBR based on conditional diffusion models and dual-layer decoupled contrastive learning has been designed and implemented to denoise the user-bundle interaction graph and learn more robust user and bundle representations.
π― What it does: A model called DisRec is proposed for group recommendation that simultaneously decouples user preferences and social influences, enhancing group representation through social contrastive learning.
Disentangling Long-Short Term State Under Unknown Interventions for Online Time Series Forecasting
Ruichu Cai (Guangdong University of Technology), Zhifeng Hao (Guangdong University of Technology)
CodeAnomaly DetectionRecurrent Neural NetworkAuto EncoderTime Series
π― What it does: Designed and implemented a Long-Short Term State Decoupling Model (LSTD) for online time series forecasting, capable of decoupling and predicting long-term and short-term latent states in non-stationary environments caused by unknown interventions.
Disentangling Tabular Data Towards Better One-Class Anomaly Detection
Jianan Ye (Xi'an Jiaotong-Liverpool University), Kaizhu Huang (Duke Kunshan University)
CodeAnomaly DetectionTabularFinance Related
π― What it does: For a type of classification scenario in table anomaly detection, this paper proposes to implicitly split two non-overlapping and related attribute subsets from normal samples through a bidirectional self-attention module, and to reconstruct the original samples using these subsets, thereby learning the internal attribute associations of normal samples, with the reconstruction error serving as the anomaly discrimination metric.
Disentangling, Amplifying, and Debiasing: Learning Disentangled Representations for Fair Graph Neural Networks
Yeon-Chang Lee (Ulsan National Institute of Science and Technology), Sang-Wook Kim (Hanyang University)
CodeGraph Neural NetworkGraph
π― What it does: A graph neural network framework named DAB-GNN is designed to separate, amplify, and remove attribute bias, structural bias, and the potential bias generated by the interaction of both in node embeddings, thereby enhancing the fairness of the model.
Distances Between Top-Truncated Elections of Different Sizes
Piotr Faliszewski (AGH University), Tomasz WΔ s (University of Oxford)
CodeTabular
π― What it does: Expanded the voting map framework to handle elections with different scales of candidates/voters and top-truncated voting, and proposed two new distance metrics - positional extension and DAP distance based on diversity, agreement, and polarization characteristics;
Distilling Structured Rationale from Large Language Models to Small Language Models for Abstractive Summarization
Linyong Wang (Northwestern Polytechnical University), Kang Wang
CodeGenerationKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A structured reasoning mining and weak gating fusion framework based on large language models (LLM) is proposed (LSR-MWF), which distills three perspectives of structured reasoning (Essential Aspects, Associated Sentences, Triple Entity Relations) generated by a 70B-level LLM into a summary generation model with β€1B parameters.
π― What it does: A multi-modal hashing method based on distribution consistency (DCGMH) is proposed to enhance retrieval performance by filtering and reconstructing noisy labels.
π― What it does: Represent queries as vectors and model documents using multivariate Gaussian distributions, calculating the likelihood of the query vector under the document distribution as a relevance score.
π― What it does: This paper presents DiT4Edit, an image editing framework based on Diffusion Transformer, which supports high-resolution and arbitrary size editing.
π― What it does: A UAV-based hyperspectral-LiDAR dataset ATR-HUDE is proposed, and the first publicly available real-world spectral water depth estimation dataset is constructed.
DiverSAT: A Novel and Effective Local Search Algorithm for Diverse SAT Problem
Jiaxin Liang (Northeast Normal University), Minghao Yin (Northeast Normal University)
CodeOptimization
π― What it does: A local search algorithm called Diver SAT has been developed to find a diverse set of k satisfying solutions in a given CNF formula.
π― What it does: This paper proposes an algorithm that utilizes a pre-trained GAN and optimizes a multi-objective framework (rarity, diversity, similarity) to generate diverse rare samples without the need for retraining the GAN.
π― What it does: This paper proposes the Region-Guided Transformer (RGTR), which constructs queries through explicitly initialized static and dynamic anchor points, providing regional priors to address the query overlap issue in traditional DETR and achieving diverse, non-overlapping temporal sentence localization.
Divide-and-Conquer: Tree-structured Strategy with Answer Distribution Estimator for Goal-Oriented Visual Dialogue
Shuo Cai (University of Chinese Academy of Sciences), Shuhui Wang (Key Lab of Intelligent Information Processing)
CodeTransformerReinforcement LearningVision Language ModelImageText
π― What it does: Proposes the Tree-Structured Strategy with Answer Distribution Estimator (TSADE), which guides question generation through a binary search method, allowing the target visual dialogue to converge to a unique candidate object in fewer rounds.
Divide, Conquer and Combine: A Training-Free Framework for High-Resolution Image Perception in Multimodal Large Language Models
Wenbin Wang (Wuhan University), Dacheng Tao (Nanyang Technological University)
CodeRecognitionGenerationRetrievalTransformerLarge Language ModelImageMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: A training-free framework called Divide-Conquer-Combine (DC2) is proposed and implemented, which enhances the perception and understanding of multimodal large language models for 4K/8K images through recursive chunking, merging similar chunks, generating text descriptions, and utilizing visual memory.
Pan Wang (University of Pittsburgh), Jingtong Hu (University of Pittsburgh)
CodeClassificationTransformerVideoMultimodality
π― What it does: Proposes the Disentangled-Language-Focused (DLF) framework, which first uses a feature disentanglement module to separate multimodal features into shared and modality-specific spaces, then focuses complementary information from other modalities onto the language modality through the Language-Focused Attractor (LFA), and finally performs hierarchical prediction to enhance the accuracy of multimodal sentiment analysis.
DM-Adapter: Domain-Aware Mixture-of-Adapters for Text-Based Person Retrieval
Yating Liu (Shenzhen International Graduate School Tsinghua University), Qingmin Liao (Shenzhen International Graduate School Tsinghua University)
CodeRetrievalTransformerMixture of ExpertsContrastive LearningText
π― What it does: This paper proposes a parameter-efficient transfer learning framework based on CLIP, called DM-Adapter, which uses Sparse Mixture-of-Adapters and Domain-Aware Router to perform fine-grained modeling of portrait features in text retrieval while training only about 16M parameters.
DMT-RoleBench: A Dynamic Multi-Turn Dialogue Based Benchmark for Role-Playing Evaluation of Large Language Model and Agent
Dingbo Yuan (Ant Group), Song Liu (Ant Group)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: Designed and implemented the DMT-RoleBench benchmark, which evaluates the role-playing capabilities of large language models and agents using dynamic multi-turn dialogues.
DocMamba: Efficient Document Pre-training with State Space Model
Pengfei Hu (University of Science and Technology of China), Jianshu Zhang (iFLYTEK Research)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A document pre-training framework called DocMamba based on a state space model is proposed, achieving linear time complexity for visual text understanding.
π― What it does: A new negative sample sampling strategy for graph contrastive learning, called E2Neg, is proposed. Theoretical analysis shows that a large number of negative samples can weaken semantic differentiation, so it is changed to only use a small number of high-quality and non-topologically coupled representative negative samples for contrastive learning.
Does GPT Really Get It? A Hierarchical Scale to Quantify Human and AIβs Understanding of Algorithms
Mirabel Reid (Georgia Institute of Technology), Santosh S. Vempala (Georgia Institute of Technology)
CodeTransformerLarge Language ModelTextMultimodality
π― What it does: This paper proposes and validates a hierarchical algorithm understanding scale, exploring the understanding of algorithms by large language models and humans.
Does VLM Classification Benefit from LLM Description Semantics?
Pingchuan Ma (CompVis at LMU Munich), BjΓΆrn Ommer (CompVis at LMU Munich)
CodeClassificationRecognitionVision Language ModelContrastive LearningImageText
π― What it does: This paper proposes a training-free method to select descriptors that can significantly improve the image classification accuracy of Vision-Language models (such as CLIP);
Does Your AI Agent Get You? A Personalizable Framework for Approximating Human Models from Argumentation-based Dialogue Traces
Yinxu Tang (Washington University in St. Louis), William Yeoh (Washington University in St. Louis)
CodeTextSequential
π― What it does: This paper proposes the Persona framework, which utilizes probabilistic Bayesian updating and a prospect theory weighting function to dynamically learn and update the psychological model of human users in argumentative dialogues.
DoGA: Enhancing Grounded Object Detection via Grouped Pre-Training with Attributes
Yang Liu (Institute of Computing Technology Chinese Academy of Sciences), Zhiqiang He (Lenovo Ltd)
CodeObject DetectionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage
π― What it does: Proposes the DoGA scheme, which enhances visual-language object detectors through attribute-driven group pre-training, mainly including attribute prompts (context definition and hard negatives), parallel text encoding, entity extraction fusion, and polysemous phrase-level training.
Zepeng Zhang (Ecole Polytechnique Federale de Lausanne), Olga Fink (Ecole Polytechnique Federale de Lausanne)
CodeDomain AdaptationGraph Neural NetworkGraph
π― What it does: This paper proposes the Cascaded Propagation (CP) strategy, which makes architectural improvements to the Unfolded Graph Neural Network (UGNN) in the unsupervised target domain graph domain adaptation task to enhance cross-domain transfer performance.
π― What it does: A domain generalization framework for medical landmark detection based on single-center data was constructed, utilizing publicly available segmentation data for boundary-aware pre-training, and introducing a mixed loss of log-cosh and MSE during the pre-training phase, followed by fine-tuning in the landmark detection task;
DOMBA: Double Model Balancing for Access-Controlled Language Models via Minimum-Bounded Aggregation
Tom Segal (Ben-Gurion University of the Negev), Yuval Elovici (Ben-Gurion University of the Negev)
CodeSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The DOMBA method is proposed, which utilizes two sub-models trained on two different access level datasets, and implements access control during inference using minimum-bounded relative probabilities, balancing security and efficiency.
π― What it does: Proposes the DomCLP method, which utilizes Domain-wise Contrastive Learning (DCon) and Prototype Mixup (PMix) for unsupervised domain generalization.
π― What it does: This paper proposes an online streaming inference framework named StreamiNNC, which leverages the translational invariance of CNN convolutions to skip unnecessary computations under overlapping windows, significantly reducing the inference cost of temporal CNNs.
DPCL-DiffοΌTemporal Knowledge Graph Reasoning Based on Graph Node Diffusion Model with Dual-Domain Periodic Contrastive Learning
Yukun Cao (Shanghai University of Electric Power), Luobin Huang (Shanghai University of Electric Power)
CodeGraph Neural NetworkTransformerDiffusion modelContrastive LearningGraphTime Series
π― What it does: In the task of temporal knowledge graph reasoning, the DPCL-Diff model is proposed, which combines a graph node diffusion model and dual-domain periodic contrastive learning to achieve high-quality predictions for new and periodic events.
π― What it does: The DR-VAE framework is proposed, combining a debiasing estimator and a continuous representation enhancer to address exposure bias and posterior collapse issues in collaborative filtering.
π― What it does: This paper proposes a multi-label attribute detection framework for remote tongue diagnosis called SignNet, and constructs a novel tongue image dataset.