π― What it does: The ISPDiffuser framework is proposed, which splits the RAW-to-sRGB task into two steps: grayscale detail reconstruction and color consistency mapping, achieving high-quality imaging through a texture-aware diffusion model and a histogram-guided color consistency module.
IsUMap: Manifold Learning and Data Visualization Leveraging Vietoris-Rips Filtrations
Parvaneh Joharinad (Max Planck Institute for Mathematics in the Sciences), JΓΌrgen Jost (Max Planck Institute for Mathematics in the Sciences)
CodeBiomedical Data
π― What it does: A hybrid manifold learning algorithm named IsUMap is proposed, which integrates UMAP and Isomap, and utilizes Vietoris-Rips persistence and fuzzy simplicial metrics to achieve low-dimensional embedding.
Mitchell Plyler (North Carolina State University), Min Chi (North Carolina State University)
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelText
π― What it does: A method for Iterative Counterfactual Data Augmentation (ICDA) has been designed and implemented to improve interpretability and reduce bias in text classification tasks.
Iterative Sparse Attention for Long-sequence Recommendation
Guanyu Lin (Carnegie Mellon University), Depeng Jin (Tsinghua University)
CodeRecommendation SystemTransformerSequential
π― What it does: This paper proposes an ISA model for long sequence recommendation, which combines sparse attention layers and iterative attention layers to achieve efficient modeling of long behavior sequences and capture interest drift.
IteRPrimE: Zero-shot Referring Image Segmentation with Iterative Grad-CAM Refinement and Primary Word Emphasis
Yuji Wang (Shenzhen International Graduate School Tsinghua University), Yansong Tang (Shenzhen International Graduate School Tsinghua University)
CodeObject DetectionSegmentationTransformerVision Language ModelImage
π― What it does: A zero-shot referential image segmentation framework called IteRPrimE is proposed, which completes segmentation by aligning the target regions in the original image using Grad-CAM.
π― What it does: Proposed and validated the Instance-aware Test Pruning (ITP) method for post-processing on pre-trained models to enhance OOD detection performance.
IWRN:A Robust Blind Watermarking Method for Artwork Image Copyright Protection Against Noise Attack
Feifei Kou (Beijing University of Posts and Telecommunications), Xuejing Kang (Beijing University of Posts and Telecommunications)
CodeData SynthesisSafty and PrivacyConvolutional Neural NetworkContrastive LearningImage
π― What it does: A robust blind watermarking method for artistic images, IWRN, is proposed, which can resist various noise attacks while maintaining invisibility.
π― What it does: Jointly designing neural network structures, ultra-low bit mixed precision quantization, and dedicated accelerator architectures to construct an end-to-end differentiable JAQ framework;
Joint Knowledge Editing for Information Enrichment and Probability Promotion
Wenhang Shi (Renmin University of China), Xiaoyong Du
CodeTransformerContrastive LearningText
π― What it does: A knowledge editing method called JEEP is proposed, which simultaneously edits low-level information richness and high-level probability enhancement, utilizing contrastive detection to identify key stages of editing.
JoVALE: Detecting Human Actions in Video Using Audiovisual and Language Contexts
Taein Son (Hanyang University), Jun Won Choi (Seoul National University)
CodeRecognitionTransformerVision Language ModelVideoMultimodalityAudio
π― What it does: This paper proposes a multi-modal video action detection framework called JoVALE, which combines audio, visual, and language contextual information to achieve action recognition through actor-centric feature aggregation.
Just What You Desire: Constrained Timeline Summarization with Self-Reflection for Enhanced Relevance
Muhammad Reza Qorib (National University of Singapore), Hwee Tou Ng (National University of Singapore)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper proposes the Constrained Timeline Summarization (CTLS) task and constructs a new dataset called CREST. It also designs a training-free method based on large language models (LLM) called REACTS, which generates constraint-compliant timelines through steps such as self-reflection, clustering, and summarization.
K-hop Hypergraph Neural Network: A Comprehensive Aggregation Approach
Linhuang Xie (Xiamen University), Taisong Jin (South China Normal University)
CodeGraph Neural NetworkGraph
π― What it does: A K-hop hypergraph neural network (KHGNN) based on path features is proposed to address the issues of over-compression and over-smoothing in traditional hypergraph GNNs.
Feijiang Li (Institute of Big Data Science and Industry Shanxi University), Liang Du (Institute of Big Data Science and Industry Shanxi University)
CodeOptimizationTabular
π― What it does: A clustering ensemble method based on k-HyperEdge Medoids, CEHM, is proposed. It first constructs the base clustering results into a hypergraph, and then iteratively obtains the final k non-overlapping hyperedges through three steps: initialization, diffusion, and tuning, directly providing the clustering results.
KAES: Multi-aspect Shared Knowledge Finding and Aligning for Cross-prompt Automated Scoring of Essay Traits
Xia Li (Guangdong University of Foreign Studies), Wenjing Pan (Guangdong University of Foreign Studies)
CodeOptimizationConvolutional Neural NetworkRecurrent Neural NetworkTransformerLarge Language ModelText
π― What it does: An optimized framework for multi-dimensional shared knowledge discovery and alignment, KAES, is proposed for cross-prompt automated essay scoring.
KALAHash: Knowledge-Anchored Low-Resource Adaptation for Deep Hashing
Shu Zhao (Pennsylvania State University), Vijaykrishnan Narayanan (Pennsylvania State University)
CodeRetrievalDomain AdaptationVision Language ModelImage
π― What it does: Under extremely low resources (such as 1-shot, 2-shot), adaptive learning is applied to pre-trained deep hashing models to address the decline in retrieval performance caused by distribution shift.
π― What it does: To address the robustness of target detection models against adversarial patch attacks, a KDAT mechanism is proposed, which enhances the detection performance of the model on both adversarial and normal images through knowledge distillation and adversarial fine-tuning while maintaining the original inference time.
KernelMatmul: Scaling Gaussian Processes to Large Time Series
Tilman Hoffbauer (RWTH Aachen University), Jakob Bossek (Paderborn University)
CodeTime Series
π― What it does: Proposes KernelMatmul, which accelerates Gaussian process regression and reduces memory usage for large-scale irregular time series.
Key-Point-Driven Data Synthesis with Its Enhancement on Mathematical Reasoning
Yiming Huang (Microsoft), Weizhu Chen (Microsoft)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A keypoint-driven data synthesis method (KPDDS) is proposed, generating large-scale mathematical reasoning datasets KPMath and KPMath-Plus. Subsequently, various LLMs are fine-tuned under supervision, improving performance on multiple mathematical benchmarks.
KGCRR: An Effective Metric-Driven Knowledge Graph Completion Framework by Designing a Novel Upper Bound Function with Adaptive Approximation to Reciprocal Rank
π― What it does: A framework named KGCRR is proposed, which effectively optimizes the reciprocal ranking (RR) metric in knowledge graph completion by designing a new upper bound function CRR, addressing the inconsistency between existing loss functions and the RR objective.
KITS: Inductive Spatio-Temporal Kriging with Increment Training Strategy
Qianxiong Xu (Nanyang Technological University), Zhishuai Li (SenseTime Research)
CodeGraph Neural NetworkGraphTime Series
π― What it does: An incremental training strategy named KITS is proposed, which uses virtual nodes to simulate unobserved nodes during training. This approach combines reference node feature fusion and node-aware cyclic adjustment to perform semi-supervised learning on spatial-temporal Kriging, addressing the 'graph gap' issue between training and inference graphs.
π― What it does: A semantic segmentation framework KWYAF based on event cameras is proposed, which enhances segmentation accuracy by utilizing historical motion information through spatiotemporal propagation.
Knowledge-Guided Domain Adaptation Model for Transferring Drug Response Prediction from Cell Lines to Patients
Xuan Liu (Xinyang Normal University), Menglu Li (Huazhong Agricultural University)
CodeDomain AdaptationDrug DiscoveryGraph Neural NetworkAuto EncoderBiomedical Data
π― What it does: A knowledge-driven domain adaptation model, TransDRP, is proposed to transfer drug response prediction from cell lines to patients, supporting one-time predictions for multiple drugs.
π― What it does: This paper proposes the KnowPO method, which enhances the knowledge selection ability of retrieval-augmented language models in the face of internal and external knowledge conflicts by constructing a knowledge conflict dataset and preference optimization.
Kolmogorov-Arnold Networks Still Catastrophically Forget but Differently from MLP
Anton Lee (Victoria University of Wellington), W. Bastiaan Kleijn (University of Waikato)
CodeTime SeriesPhysics Related
π― What it does: This study verifies that Kolmogorov-Arnold Networks (KAN) still experience catastrophic forgetting in more complex tasks with overlapping inputs, and proposes a parameter isolation and hierarchical pruning method called WiseKAN for KAN.
KPL: Training-Free Medical Knowledge Mining of Vision-Language Models
Jiaxiang Liu (Zhejiang University), Zuozhu Liu (Zhejiang University)
CodeClassificationRetrievalTransformerLarge Language ModelVision Language ModelContrastive LearningImageMultimodalityBiomedical Data
π― What it does: This study proposes the Knowledge Proxy Learning (KPL) method, which constructs a knowledge-enhanced foundation, utilizes visual retrieval to filter relevant descriptions to generate richer text proxies, and combines the Stable Green Horn algorithm for multimodal proxy learning, significantly improving performance in zero-shot medical image classification tasks.
L3TC: Leveraging RWKV for Learned Lossless Low-Complexity Text Compression
Junxuan Zhang (Ant Group), Li Song (Shanghai Jiao Tong University)
CodeCompressionRecurrent Neural NetworkLarge Language ModelText
π― What it does: This paper proposes a low-complexity lossless text compression method called L3TC, which utilizes the RWKV model to predict probabilities and implements compression using arithmetic coding.
L4DR: LiDAR-4DRadar Fusion for Weather-Robust 3D Object Detection
Xun Huang (Xiamen University), Cheng Wang (Xiamen University)
CodeObject DetectionAutonomous DrivingPoint Cloud
π― What it does: This paper proposes a 3D object detection framework L4DR that integrates LiDAR and 4D radar, aiming to enhance detection robustness in adverse weather conditions.
Label Aggregation for Composite Crowd Tasks by Worker Ability Constraint Satisfaction
Jiyi Li (University of Yamanashi)
CodeText
π― What it does: This paper proposes a label aggregation method for Composite Crowd Tasks (CCT), which bridges worker information across different tasks by satisfying worker capability constraints, enabling multi-task joint aggregation.
π― What it does: A backdoor attack method without label information is proposed in vertical federated learning, utilizing embedded gradients to construct poisoned samples of the target class, and enhancing the attack effect through selective sample exchange.
LAMA-UT: Language Agnostic Multilingual ASR Through Orthography Unification and Language-Specific Transliteration
Sangmin Lee (Yonsei University), Hong-Goo Kang (Yonsei University)
CodeRecognitionTransformerLarge Language ModelPrompt EngineeringAudio
π― What it does: This paper proposes a multilingual speech recognition pipeline LAMA-UT without language-specific modules: it first generates a unified romanized transcription using wav2vec2.0-XLSR, and then converts this transcription into the original text of various languages using a frozen large language model (LLM).
π― What it does: This paper proposes the Inexact Proximal Langevin Algorithm (IPLA), which extends the sampling capability of Langevin Monte Carlo (LMC) to non L-smooth, super-quadratic potential by using approximate proximal operations in Wasserstein space, and provides corresponding error analysis.
Large Language Models Are Read/Write Policy-Makers for Simultaneous Generation
Shoutao Guo (Chinese Academy of Sciences), Yang Feng (Chinese Academy of Sciences)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringTextMultimodalityAudio
π― What it does: A synchronous generation framework LSG based on large language models is proposed, allowing the LLM to read input in real-time while deciding the timing of generation and outputting text.
Large Language Models Enhanced Personalized Graph Neural Architecture Search in Federated Learning
Hui Fang (Zhejiang University), Haishuai Wang (Zhejiang University)
CodeFederated LearningNeural Architecture SearchGraph Neural NetworkLarge Language ModelPrompt EngineeringGraph
π― What it does: In a federated learning environment, a personalized graph neural network architecture search is achieved by combining large language models and hypernetworks, automatically generating GNN architectures that adapt to the heterogeneous graph data of each client.
Latent Reward: LLM-Empowered Credit Assignment in Episodic Reinforcement Learning
Yun Qu (Tsinghua University), Xiangyang Ji (Tsinghua University)
CodeTransformerLarge Language ModelReinforcement LearningPrompt EngineeringSequential
π― What it does: This paper proposes a Latent Reward framework (LaRe) based on large language models (LLMs), which generates multi-dimensional and interpretable task performance metrics and achieves finer credit allocation through return decomposition.
LATTE: Improving Latex Recognition for Tables and Formulae with Iterative Refinement
Nan Jiang (Purdue University), Lin Tan (Purdue University)
CodeRecognitionImage TranslationTransformerSupervised Fine-TuningVision Language ModelImageTabular
π― What it does: A LaTeX recognition framework called LATTE is proposed, which can generate renderable LaTeX source code from images of tables and formulas and continuously improve.
LayerAct: Advanced Activation Mechanism for Robust Inference of CNNs
Kihyuk Yoon (Ulsan National Institute of Science and Technology), Chiehyeon Lim (Ulsan National Institute of Science and Technology)
CodeClassificationSegmentationAdversarial AttackConvolutional Neural NetworkImageBiomedical Data
π― What it does: This paper proposes LayerActβa mechanism for activation in CNNs achieved through layer normalization to enhance the network's robustness to noise;
π― What it does: A multi-agent reinforcement learning method for querying unlabeled streaming data in federated learning (LeaDQ) is proposed, enabling clients to make sample labeling decisions based on the global model state without sharing data.
Learn2Aggregate: Supervised Generation of Chvatal-Gomory Cuts Using Graph Neural Networks
Arnaud Deza (University of Toronto), Yong Zhang (Huawei Technologies Canada)
CodeOptimizationGraph Neural NetworkGraph
π― What it does: By training a graph neural network for binary classification of MILP constraints, the focus is placed only on important constraints during CG cutting generation, significantly reducing the size of the separation problem.
π― What it does: A joint source-channel coding (HJSCC) framework based on Hierarchical Variational Autoencoder (VAE) is proposed, supporting adaptive rate wireless image transmission in the presence of feedback links.
π― What it does: This paper proposes the MIFAG framework, which extracts invariant affordance knowledge from multiple human-machine interaction images and integrates it into 3D point clouds to achieve precise localization of functional areas of 3D objects.
π― What it does: This paper proposes a complex heterogeneous multimodal fake news detection framework HML based on social latent network inference and self-supervised multimodal learning. It can estimate the influence of news through an event-level Hawkes process and construct a potential heterogeneous graph in the absence of real propagation chains, and then classify fake news by combining self-supervised enhanced multimodal features.
Learning Concept Prerequisite Relation via Global Knowledge Relation Optimization
Miao Zhang (Hubei University), Zhifei Li (Hubei University)
CodeGraph Neural NetworkGraphTabular
π― What it does: This paper proposes a model called GKROM based on global knowledge relationship optimization for learning prerequisite relationships of concepts.
Learning Cross-Domain Representations for Transferable Drug Perturbations on Single-Cell Transcriptional Responses
Hui Liu (Nanjing Tech University), Shikai Jin (Nanjing Tech University)
CodeRepresentation LearningDrug DiscoveryAuto EncoderBiomedical Data
π― What it does: Proposes the XTransferCDR framework to predict transcriptional responses at the single-cell level caused by drug and genetic perturbations through cross-domain representation learning.
π― What it does: Learn the dynamics of time-series data through neural networks, and use the nonlinear KYP lemma to construct a differentiable projection that maps any dynamics to a dissipative subspace, thereby ensuring the system's dissipation (internal stability, input-output stability, and energy conservation) during the training process.
π― What it does: This paper proposes a drone visual dialogue navigation method for fine-grained entity-landmark alignment, enhancing navigation effectiveness through semantic grid representation and three auxiliary tasks.
Learning Fine-grained Domain Generalization via Hyperbolic State Space Hallucination
Qi Bi (Wuhan University), Gui-Song Xia (Wuhan University)
CodeDomain AdaptationImage
π― What it does: This paper proposes Hyperbolic State Space Hallucination (HSSH), which enhances fine-grained domain generalization by performing style hallucination in the state space and applying consistency constraints in hyperbolic space.
Learning from Mistakes: Self-correct Adversarial Training for Chinese Unnatural Text Correction
Xuan Feng (Jinan University), Liang Chang (Guangxi University of Electronic Technology)
CodeGenerationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningGenerative Adversarial NetworkContrastive LearningText
π― What it does: A self-correcting adversarial training framework called LIMIT is proposed, which enhances the robustness of error correction and adversarial resistance for non-natural text in both Chinese and English through a generative correction mechanism, adversarial sample generation based on self-prediction, and decoding intervention strategies.
Learning Local Neighborhoods of Non-Gaussian Graphical Models
Sarah Liaw (California Institute of Technology), Ricardo Baptista (California Institute of Technology)
CodeGraph Neural NetworkReinforcement LearningBiomedical Data
π― What it does: An expandable algorithm L-SING is proposed to learn the local neighborhood of continuous non-Gaussian distribution graphical models using transport mapping;
π― What it does: A recommendation model based on guided conditional diffusion, GCDR, is proposed, which learns user uncertainty through multi-distribution.
Learning Regularization for Graph Inverse Problems
Moshe Eliasof (University of Cambridge), Eldad Haber (University of British Columbia)
CodeGraph Neural NetworkGraph
π― What it does: A general framework based on graph neural networks is proposed to solve Graph Inverse Problems (GRIP), and it is validated on various tasks.
Learning Robust and Privacy-Preserving Representations via Information Theory
Binghui Zhang (Illinois Institute of Technology), Binghui Wang (Illinois Institute of Technology)
CodeSafty and PrivacyRepresentation LearningAdversarial AttackTabular
π― What it does: This paper proposes an information-theoretic framework called ARPRL, which learns representations that are both adversarially robust and privacy-preserving, and are generalizable to any downstream task.
π― What it does: This paper studies a framework for learning set functions under the supervision of an optimal subset oracle and improves the gradient computation method based on variational inference.
Learning Theorem Rationale for Improving the Mathematical Reasoning Capability of Large Language Models
Yu Sheng (Chinese Academy of Sciences), Daniel Dajun Zeng (Chinese Academy of Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
π― What it does: By constructing a dataset of question-theorem-answer triples and performing instruction fine-tuning, we propose Explicit Learning Theorem Reasoning (Theorem Rationale, TR) to enhance the mathematical reasoning capabilities of large language models.
Learning to Generate Gradients for Test-Time Adaptation via Test-Time Training Layers
Qi Deng (South China University of Technology), Xiping Hu (Artificial Intelligence Research Institute Shenzhen MSU-BIT University)
CodeDomain AdaptationOptimizationImage
π― What it does: A Meta Gradient Generator (MGG) based on learning optimization is proposed, which generates reliable gradients during the adaptive process at testing time, thereby replacing traditional manual optimizers;
Learning to Prompt with Text Only Supervision for Vision-Language Models
Muhammad Uzair Khattak (MBZ University of AI), Federico Tombari (Khalifa University)
CodeClassificationObject DetectionSegmentationDomain AdaptationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodality
π― What it does: ProText is proposed, which utilizes text descriptions generated by large language models for text-supervised prompt learning to enhance the zero-shot generalization ability of CLIP in visual tasks.
π― What it does: A machine forgetting method based on inverse weight prediction, InvWNN, is proposed, which can eliminate model memory with only the data to be forgotten.
Learning Together Securely: Prototype-Based Federated Multi-Modal Hashing for Safe and Efficient Multi-Modal Retrieval
Ruifan Zuo (Qilu University of Technology), Xiaofeng Qu (University of Jinan)
CodeRetrievalFederated LearningSafty and PrivacyMultimodality
π― What it does: Proposes Prototype-based Federated Multi-modal Hashing (PFMH), which combines federated learning with multi-modal hashing to achieve secure and efficient multi-modal retrieval.
π― What it does: Utilizing self-supervised learning, the graph neural network directly predicts Lagrange multipliers, thereby generating effective and compact dual lower bounds in constraint programming, replacing or accelerating traditional subgradient iterations;
Learning Visually Grounded Domain Ontologies via Embodied Conversation and Explanation
Jonghyuk Park (University of Edinburgh), Subramanian Ramamoorthy (University of Edinburgh)
CodeObject DetectionSegmentationExplainability and InterpretabilityRobotic IntelligenceAgentic AIImage
π― What it does: Proposes an interpretable interactive learning framework that allows agents to gradually build a visual domain ontology and improve visual recognition through corrections from teachers based on their explanations.
π― What it does: To address the problem of noisy labels in an open world, a dual representation space joint learning framework is proposed, which learns features of the prototype space and class-independent space through a projection network and a one-to-many network.
π― What it does: A learnware normalization method named LANE is proposed, which utilizes label information and random neural networks to map training data into a neural embedding space, thereby generating more accurate model specifications.
Less Is More: Adaptive Program Repair with Bug Localization and Preference Learning
Zhenlong Dai (Zhejiang University), Jingyuan Chen (Zhejiang University)
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes the AdaPR task and designs the AdaPatcher two-stage framework (Bug Locator + Program Modifier) to achieve adaptive program repair, generating patches that meet specifications and have minimal modifications.
π― What it does: To address the problem of multi-object (organ) detection under source-free unsupervised domain adaptation (SF-UDA), an AATS framework utilizing anatomical consistency (topological and morphological) is proposed, which can achieve ultrasound image detection transfer across hospitals and devices using only source domain pre-trained weights.
π― What it does: A flow model based on constraint violation signals, CV-Flow, has been developed to directly map sampled potential actions to the actionable action space in reinforcement learning with action constraints, and it is combined with Soft Actor-Critic (SAC) for safe control.
Leveraging First and Zeroth-Order Gradient to Address Imbalanced Black-Box Prompt Tuning via Minimax Optimization
Haozhen Zhang (Jilin University), Yi Chang (Jilin University)
CodeOptimizationTransformerPrompt EngineeringText
π― What it does: A black-box prompt tuning framework (BPT-FZG) is designed to handle imbalanced data in downstream tasks by maximizing AUC and transforming it into a non-convex-concave extremum problem.
Leveraging Group Classification with Descending Soft Labeling for Deep Imbalanced Regression
Ruizhi Pu (Western University), Boyu Wang (Western University)
CodeClassificationOptimizationMixture of ExpertsContrastive LearningText
π― What it does: This paper proposes a group classification and multi-expert regression framework based on symmetric decreasing soft labels, utilizing group contrastive learning and soft labels to enhance deep imbalance regression performance.
Leveraging Large Language Models for Node Generation in Few-Shot Learning on Text-Attributed Graphs
Jianxiang Yu (East China Normal University), Xuecang Zhang (Huawei Technologies)
CodeClassificationGenerationData SynthesisMeta LearningGraph Neural NetworkLarge Language ModelPrompt EngineeringTextGraph
π― What it does: A lightweight framework LLM4NG is proposed, which generates node text samples using large language models in few-shot scenarios, and integrates these generated nodes into the original graph through an edge predictor, achieving node classification without modifying the original graph data.
Leveraging the Dual Capabilities of LLM: LLM-Enhanced Text Mapping Model for Personality Detection
Weihong Bi (Beijing University of Posts and Telecommunications), Mingying Xu (North China University of Technology)
CodeClassificationTransformerLarge Language ModelContrastive LearningText
π― What it does: This paper proposes an LLM-enhanced Text Mapping Model (ETM) that encodes posts using a small model, extracts long text embeddings with a lightweight LLM, fuses user vectors through cross-attention, generates multi-dimensional MBTI label explanations with a powerful LLM, and optimizes the mapping of user vectors to labels using contrastive learning, thereby achieving personality detection.
π― What it does: A list-decodable federated learning framework LiD-FL is proposed, where the server maintains a set of models, randomly selects client updates, and ensures that at least one model is valid through a voting mechanism.
π― What it does: In an unsupervised, task-agnostic continual learning scenario, a Dynamic Expansion Memory Unit (DEMU) is proposed, which determines whether to add new memory buffers to preserve key data through Maximum Mean Discrepancy (MMD);
π― What it does: The IFDD framework is proposed, utilizing a learnable wavelet lifting scheme for two-stage implicit decomposition of video expression dynamics, separating emotion-related dynamic features from global context through the Inter-frame Static-dynamic Splitting Module (ISSM) and the Lifting-based Aggregation-disentanglement Module (LADM);
π― What it does: A lightweight text-to-motion generation model, Light-T2M, is proposed, significantly reducing the number of parameters and inference latency while maintaining or improving generation quality.
LightPROF: A Lightweight Reasoning Framework for Large Language Model on Knowledge Graph
Tu Ao (Beijing University of Posts and Telecommunications), Zhen Cai (National University of Singapore)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringTextGraphRetrieval-Augmented Generation
π― What it does: The LightPROF framework is proposed, which enhances KGQA reasoning capabilities by generating soft prompts for small-scale LLMs through a three-step process of retrieval, embedding, and inference using the structural information of knowledge graphs.
π― What it does: Proposes LightGCN2N, a lightweight graph capsule convolutional network with subspace alignment, to address the shared account sequential recommendation problem.
Linking Industry Sectors and Financial Statements: A Hybrid Approach for Company Classification
Guy Stephane Waffo Dzuyo (Forvis Mazars), Luis Belmar-Letelier (Forvis Mazars)
CodeClassificationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTextTabularFinance Related
π― What it does: This paper develops various financial statement representation methods and constructs a company industry classification model based on machine learning and language models, exploring the enhancement of classification performance through textual information.
π― What it does: This paper proposes the LiON method, which combines selective classification and point-level abandonment penalty to address the issue of LiDAR point cloud anomaly detection, and generates diverse realistic anomalies through ShapeNet.
LLaVA Needs More Knowledge: Retrieval Augmented Natural Language Generation with Knowledge Graph for Explaining Thoracic Pathologies
Ameer Hamza (Kyung Hee University), Seong Tae Kim (Kyung Hee University)
CodeGenerationRetrievalTransformerLarge Language ModelVision Language ModelImageTextMultimodalityComputed TomographyElectronic Health RecordsRetrieval-Augmented Generation
π― What it does: This paper proposes a knowledge graph retrieval-augmented visual-language framework (KG-RAG) for generating natural language explanations (NLE) of chest X-ray images.
CodeRecommendation SystemTransformerLarge Language ModelReinforcement LearningTextSequential
π― What it does: A user simulator based on large language models has been constructed, combining logical reasoning and statistical learning to explicitly simulate the 'like/dislike' interactions in recommendation systems.
LLM-RG4: Flexible and Factual Radiology Report Generation Across Diverse Input Contexts
Zhuhao Wang (Tsinghua University), Hongen Liao (Tsinghua University)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataElectronic Health Records
π― What it does: A radiology report generation framework LLM-RG4 is proposed, which can flexibly adapt to various clinical input scenarios, and a corresponding dataset MIMIC-RG4 is constructed.
LLM+AL: Bridging Large Language Models and Action Languages for Complex Reasoning About Actions
Adam Ishay (Arizona State University), Joohyung Lee (Samsung Research)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: A framework that combines large language models (LLM) with action language (AL) is proposed (LLM+AL) for complex action reasoning tasks.
LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation
Yatong Sun (Northeastern University), Xinghua Qu (Bytedance)
CodeRecommendation SystemKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringSequential
π― What it does: By using large language models (LLM) to semantically correct unreliable input-target pairs in sequence recommendation training samples, the accuracy of robust sequence recommendations is improved.
LLMEmb: Large Language Model Can Be a Good Embedding Generator for Sequential Recommendation
Qidong Liu (Xi'an Jiaotong University), Yefeng Zheng (Tencent)
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringContrastive LearningSequential
π― What it does: This paper proposes LLMEmb, a framework that utilizes large language models to generate project embeddings for sequential recommendation systems.
π― What it does: In high-density multi-agent path planning tasks, the authors propose the LNS2+RL algorithm, which adaptively switches between low-speed high-quality planning based on MARL and high-speed low-quality planning based on PP+SIPPS during the iterative process, significantly reducing the number of conflicts while maintaining scalability.
π― What it does: This paper proposes the open vocabulary object detection task 'Locate Anything on Earth (LAE)' in the field of remote sensing, and constructs the LAE-1M dataset containing approximately one million instances through the LAE-Label Engine. It also designs and trains the LAE-DINO model, enhancing the open object detection capability of remote sensing images.
π― What it does: This paper proposes an OOD detection framework called PATT under long-tail distribution, focusing on balancing features and classifiers through attention enhancement for tail classes and implicit semantic enhancement, as well as feature calibration during inference.
Look Before You Leap: Enhance Attention and Vigilance Regarding Harmful Content with GuidelineLLM
Shaoqing Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: A defense framework named GuidelineLLM is proposed, which utilizes an auxiliary LLM to identify risks in queries before answering and generate safety guidelines. These guidelines are then provided along with the original query to the target LLM, thereby reducing the probability of generating harmful content.
Loosely Synchronized Rule-Based Planning for Multi-Agent Path Finding with Asynchronous Actions
Shuai Zhou (South China University of Technology), Zhongqiang Ren (Shanghai Jiao Tong University)
CodeOptimizationGraphBenchmark
π― What it does: This paper proposes a rule-based loosely coupled planning algorithm (LSRP and its SWAP variant) to address the issue of asynchronous actions in multi-agent path planning, capable of quickly obtaining acceptable approximate solutions under a large number of agents (up to 1000).
Low-Light Image Enhancement via Generative Perceptual Priors
Han Zhou (McMaster University), Jun Chen (McMaster University)
CodeRestorationTransformerVision Language ModelDiffusion modelImage
π― What it does: This paper proposes a low-light image enhancement framework based on Generative Perceptual Priors (GPP), called GPP-LLIE. It first utilizes a pre-trained vision-language model (LLaVA) to extract global and local perceptual priors from the image, and then guides the enhancement through GPP-LN and LPP-Attn in a diffusion Transformer, achieving more realistic and detail-rich enhancements.
Christian Hagemeier (University of Oxford), Moshe Y. Vardi (Rice University)
CodeReinforcement Learning
π― What it does: A method for LTLf synthesis that simultaneously satisfies the main objective and backup objective when the input variables are unreliable is proposed, addressing the need to consider both fully observable and partially observable scenarios during synthesis.
M^3EL: A Multi-task Multi-topic Dataset for Multi-modal Entity Linking
Fang Wang (Peking University), Yi Liang (Xinjiang University)
CodeTransformerLarge Language ModelContrastive LearningImageTextMultimodality
π― What it does: A multi-task, multi-topic, multi-modal entity linking dataset Mβ―ELβ―3 has been constructed, consisting of 79K instances and 318.5K images, and a training strategy that incorporates entity descriptions into the text for modal enhancement has been proposed.
M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images
Hongyi Wang (Zhejiang University), Lanfen Lin (Zhejiang University)
CodeTransformerImageBiomedical Data
π― What it does: A multi-scale to one regression Transformer (M2OST) is proposed, which can utilize pathological images of different resolutions to predict spatial transcriptomics (ST) gene expression.
π― What it does: A multi-modal multi-task network M3Net is proposed, capable of simultaneously performing 3D detection, BEV semantic segmentation, and 3D occupancy prediction.
π― What it does: Proposes the MABR framework, which dynamically identifies biased samples at multiple levels of the Transformer encoder through an auxiliary bias detector, and eliminates these biases using adversarial training without any prior bias knowledge or protected attribute labels.