Rethinking U-Net: Task-Adaptive Mixture of Skip Connections for Enhanced Medical Image Segmentation
Zichen Luo (Tianjin University), Biao Sun (Tianjin University)
CodeSegmentationConvolutional Neural NetworkMixture of ExpertsImageBiomedical DataMagnetic Resonance Imaging
π― What it does: A TA-MoSC module is proposed based on U-Net, treating skip connections as a task allocation problem, and dynamically assigning appropriate multi-scale features for different decoding stages through a sparse Mixture-of-Experts router, thus achieving adaptive skip connections.
π― What it does: A global enhancement and optimization network named GEONet is proposed for lane detection in complex scenarios, combining a Global Feature Extraction Module (GFEM) and a Top-level Supplement Module (TTSM), and incorporating Whitened Batch Normalization (WBN), Whitened Contrastive Learning (WCL), as well as novel Angle Loss and GRIoU Loss to improve detection accuracy and robustness.
Retrieval-Augmented Visual Question Answering via Built-in Autoregressive Search Engines
Xinwei Long (Tsinghua University), Bowen Zhou (Tsinghua University)
CodeGenerationRetrievalTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextMultimodalityRetrieval-Augmented Generation
π― What it does: This paper proposes a unified retrieval-generation framework named ReAuSE for knowledge-driven visual question answering tasks, integrating generative retrieval and answer generation within the same large-scale multimodal language model.
Revelations: A Decidable Class of POMDPs with Omega-Regular Objectives
Marius Belly (CNRS, LaBRI, Universite de Bordeaux), Pierre Vandenhove (University of Antwerp)
CodeReinforcement Learning
π― What it does: Weak and strong reveal constraints are defined in partially observable Markov decision processes (POMDPs), and it is proven that under these two types of constraints, almost-sure strategies that satisfy Ο-regular objectives (including Parity, Buchi, CoBuchi) can be decided, and can be solved by reducing the POMDP to a belief-supported MDP, with an algorithmic complexity of EXPTIME;
Reverse Distribution Based Video Moment Retrieval for Effective Bias Elimination
Lingdu Kong (Northeastern University), Xiangmin Zhou (RMIT University)
CodeRetrievalGaussian SplattingVideoText
π― What it does: A video moment retrieval method based on reverse distribution, ReDis-VMR, and a dynamically scalable adjustment DEA is proposed to eliminate data and model bias and improve retrieval performance.
Reverse Region-to-Entity Annotation for Pixel-Level Visual Entity Linking
Zhengfei Xu (Beijing Institute of Technology), Xin Xin (Tencent)
CodeObject DetectionSegmentationTransformerLarge Language ModelSupervised Fine-TuningImage
π― What it does: This paper studies the pixel-level visual entity linking (PL-VEL) task and constructs the MaskOVEN-Wiki large-scale pixel-level annotated dataset.
π― What it does: This paper proposes an anomaly detection framework based on graph contrastive learning (AD-GCL), specifically addressing the poor performance of tail anomaly detection caused by structural imbalance (between head nodes and tail nodes).
π― What it does: A label correction method based on gradient scaling theory, DULC, is designed to dynamically determine the interpolation weight of each sample using normalized Jensen-Shannon divergence, and to enhance the model's memory of true labels through weak/strong data augmentation and prediction sharpening.
π― What it does: Proposes the GS-MCC framework, which constructs a multimodal interaction graph through a sliding window, uses Fourier graph operators to extract high and low-frequency semantics, and employs contrastive learning to achieve semantic consistency and complementarity, ultimately performing sentiment prediction in a single MLP.
π― What it does: This paper systematically analyzes the impact of multimodal fusion architecture on 3D anomaly detection performance, proposing a two-layer fusion search space and implementing 3D-ADNAS.
π― What it does: This paper proposes an open-set text detection task for forged scenes, constructs a high-quality benchmark dataset covering eight generative text editing models, and introduces a texture jitter pre-training and difference-aware forensics framework to enhance the model's detection capability for unknown forgeries.
Revolutionizing Encrypted Traffic Classification with MH-Net: A Multi-View Heterogeneous Graph Model
Haozhen Zhang (Tsinghua University), Ye Zhang (Nanjing University of Posts and Telecommunications)
CodeClassificationSafty and PrivacyGraph Neural NetworkContrastive LearningGraph
π― What it does: Proposes MH-Net, which utilizes multi-view heterogeneous flow graphs for packet-level and flow-level classification of encrypted traffic.
RFL: Simplifying Chemical Structure Recognition with Ring-Free Language
Qikai Chang (University of Science and Technology of China), Jinshui Hu (iFLYTEK Research)
CodeRecognitionRecurrent Neural NetworkImage
π― What it does: This paper proposes a Ring-Free Language (RFL) that decouples chemical structure diagrams into molecular skeletons, individual ring, and branch information, and based on this, designs a unified Molecular Skeleton Decoder (MSD) for end-to-end recognition.
π― What it does: A RHanDS framework based on diffusion models is proposed to correct deformed hand structures in generated images while maintaining the original style.
RhythmMamba: Fast, Lightweight, and Accurate Remote Physiological Measurement
Bochao Zou (University of Science and Technology Beijing), Huimin Ma (China Academy of Electronics and Information Technology)
CodeComputational EfficiencyConvolutional Neural NetworkTransformerVideoTime Series
π― What it does: Designed and implemented RhythmMamba, an efficient remote photoplethysmography (rPPG) signal extraction framework based on state space models.
Riemann-based Multi-scale Attention Reasoning Network for Text-3D Retrieval
Wenrui Li (Harbin Institute of Technology), Xiaopeng Fan (Peng Cheng Laboratory)
CodeRetrievalContrastive LearningTextPoint Cloud
π― What it does: The RMARN model is proposed, utilizing Riemannian geometry and multi-scale attention to achieve retrieval alignment between text and 3D point clouds.
RingFormer: A Ring-Enhanced Graph Transformer for Organic Solar Cell Property Prediction
Zhihao Ding (Hong Kong Polytechnic University), Chen Jason Zhang (Hong Kong Polytechnic University)
CodeGraph Neural NetworkTransformerGraph
π― What it does: Developed RingFormer, a graph Transformer framework that combines atomic layers, ring layers, and interlayers for predicting the properties of organic solar cell (OSC) molecules.
CodeOptimizationReinforcement LearningTabularFinance Related
π― What it does: This paper studies the introduction of Entropy Risk Measure (ERM) and Entropy Value at Risk (EVaR) as risk measures within the framework of Total Return Criterion (TRC) in finite Markov Decision Processes (MDPs), and proves the existence of an optimal deterministic stationary policy in transient MDPs.
RMath: A Logic Reasoning-Focused Datasets Toward Mathematical Multistep Reasoning Tasks
Ziyi Hu (National University of Defense Technology), Yiping Song (National University of Defense Technology)
CodeTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A mathematical reasoning dataset RMath for multi-step logical reasoning has been constructed, and a standardized reasoning framework and annotation scheme based on propositional logic have been proposed for training and evaluating the logical reasoning capabilities of LLMs.
π― What it does: A robust image hashing framework called CMAA based on end-to-end contrastive masked autoencoders is proposed, which enhances robustness against large-scale and mixed attacks by combining weak-strong augmentation alignment, ViT masks, and non-parametric quantization.
Robust Tracking via Mamba-based Context-aware Token Learning
Jinxia Xie (Guangxi Normal University), Shuxiang Song (Wuzhou University)
CodeObject TrackingVideo
π― What it does: We propose TemTrack, a visual tracker that separates temporal information learning from appearance modeling, capturing target appearance changes and motion trends through trajectory tokens in a sliding window and a mamba-cross attention module.
RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data
Chenglong Wang (Northeastern University), Jingbo Zhu (Northeastern University)
CodeRecommendation SystemReinforcement LearningVision Language ModelImageText
π― What it does: Through a three-stage progressive training and optimal transport preference data selection, the visual reward model is enhanced using text preference data, allowing large visual language models to better align with human preferences.
Rule-Guided Graph Neural Networks for Explainable Knowledge Graph Reasoning
Zhe Wang (Griffith University), Zhiqiang Zhuang (Tianjin University)
CodeExplainability and InterpretabilityGraph Neural NetworkGraphBenchmark
π― What it does: This paper proposes a rule-guided graph neural network for interpretable knowledge graph reasoning, which can utilize existing rules to guide learning during the training process, and directly extract high-confidence rules from the network parameters after the model training is completed.
π― What it does: A method for indoor scene synthesis called S-INF (Scene Implicit Neural Field) is proposed, which utilizes implicit neural fields to simultaneously learn the relationships of scene layout and detailed objects, generating more realistic and stylistically consistent 3D indoor scenes.
π― What it does: A Semantic Structure-aware Denoising Network (SΒ²DN) is proposed for Inductive Knowledge Graph Completion (Inductive KGC), which enhances inference on new entities by smoothing semantic similarity relationships in subgraphs and adaptively filtering structures.
π― What it does: A semantic stacking (S2S2) training strategy is proposed as an add-on to existing segmentation models to enhance the robustness of the models.
Myung-Joon Kwon (Korea Advanced Institute of Science and Technology), Changick Kim (NAVER WEBTOON AI)
CodeSegmentationContrastive LearningImage
π― What it does: We propose SAFIRE, a point-guided multi-source image segmentation method designed to achieve both counterfeit localization and multi-source differentiation simultaneously.
SalMΒ²: An Extremely Lightweight Saliency Mamba Model for Real-Time Cognitive Awareness of Driver Attention
Chunyu Zhao (Southwest Jiaotong University), Tao Deng (Southwest Jiaotong University)
CodeAutonomous DrivingContrastive LearningImage
π― What it does: A lightweight SalM2 driver attention prediction model has been developed, capable of capturing attention distribution in driving scenes in real-time.
SAM-Aware Graph Prompt Reasoning Network for Cross-Domain Few-Shot Segmentation
Shi-Feng Peng (Nanjing University of Science and Technology), Guo-Sen Xie (Southeast University)
CodeSegmentationGraph Neural NetworkLarge Language ModelImageBenchmark
π― What it does: Proposes the SAM-aware Graph Prompt Reasoning Network (GPRN), which generates visual prompts through SAM, utilizes a graph attention network to achieve global semantic consistency among prompts, and employs adaptive point selection for refined segmentation during testing.
Sample-aware Adaptive Structured Pruning for Large Language Models
Jun Kong (Yunnan University), Xuejie Zhang (Yunnan University)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a sample-aware adaptive structural pruning framework for large language models called AdaPruner, which achieves more efficient structural pruning by adaptively searching for the best calibration data and importance estimation metrics.
π― What it does: This paper proposes a single-step correction machine learning algorithm based on Singular Value Decomposition (SVD) called SAP, which uses a small number of low cross-entropy loss samples to estimate reliable samples and projects the model weights into a reliable activation space, thereby eliminating the impact of label noise.
π― What it does: This paper proposes SatCLIP, a globally universal location encoder trained through contrastive learning by aligning satellite images with latitude and longitude coordinates, and fine-tuned for various geographic prediction tasks.
Scalable Acceleration for Classification-Based Derivative-Free Optimization
Tianyi Han (Beijing Supreium Technology), Yuan Jin (Beijing Supreium Technology)
CodeOptimizationHyperparameter SearchTabular
π― What it does: This paper studies a classification-based gradient-free optimization algorithm framework and proposes a new RACE-CARS algorithm, which accelerates the search through random coordinate shrinking and region reduction.
π― What it does: Two scalable decentralized online mean estimation algorithms (BβColME and CβColME) are proposed, which estimate the mean with communication limited to a finite number of neighbors under multi-dimensional distributions.
π― What it does: This paper proposes a knowledge reconstruction method based on constraint optimization, aimed at compressing the size of logic programs.
π― What it does: This paper proposes ScaleTUL, a scalable trajectory-user linking framework that utilizes a dual-stream encoder and spatiotemporal enhancement to address the large-scale trajectory matching problem.
π― What it does: The ScaleMatch framework is proposed to enhance the scale invariance and pseudo-label quality of semi-supervised semantic segmentation models under limited labeled data through Cross-Scale Interaction Fusion (CIF) and Intra-Scale Consistency Learning (ISVC, FSVC).
π― What it does: An improved heuristic based on Limited Rollback Beam Search (LRBS) is proposed to enhance the reasoning performance of deep reinforcement learning models in combinatorial optimization.
SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering
Zouying Cao (Shanghai Jiao Tong University), Hai Zhao (Shanghai Jiao Tong University)
CodeSafty and PrivacyLarge Language ModelPrompt EngineeringText
π― What it does: A security awareness control method based on activation layer driving (SCANS) is proposed, which mitigates the misrejection of legitimate queries by extracting rejection behavior vectors and performing activation shifts in security-critical layers.
π― What it does: A scene graph-based image generation framework SGG-IG is designed and implemented, directly aligning scene graphs with generated images without an intermediate layout, integrating modules such as mask self-supervision, spatial contrastive loss, and image-scene alignment.
ScholarGEC: Enhancing Controllability of Large Language Model for Chinese Academic Grammatical Error Correction
Zixiao Kong (University of Science and Technology of China), Yu Su (Hefei Normal University)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: This study proposes the ScholarGEC framework, which enhances the controllability and accuracy of LLM in grammatical error correction of Chinese academic papers by utilizing knowledge prefixes, a detect-correct structure, and self-reflection techniques.
SCOPE: Sign Language Contextual Processing with Embedding from LLMs
Yuqi Liu (ShanghaiTech University), Lan Xu (ShanghaiTech University)
CodeRecognitionImage TranslationTransformerLarge Language ModelSupervised Fine-TuningVideoText
π― What it does: This paper proposes a context-aware sign language recognition and translation framework called SCOPE based on LLM embeddings, and contributes the first large-scale Chinese sign language dialogue dataset.
ScoreNet: Consistency-driven Framework with Multi-side Information Fusion for Session-based Recommendation
Piao Tong (University of Electronic Science and Technology of China), Tian Lan (Byte Dance)
CodeRecommendation SystemTransformerSequential
π― What it does: Proposes the ScoreNet framework, which combines the MPRE and PECS modules for multi-side information fusion, improving interest modeling and score consistency in session-based recommendations.
π― What it does: This paper proposes a fully unsupervised camouflage object detection method called SdalsNet, which utilizes self-distillation attention localization and transfer to learn target features and directly obtain segmentation results from the attention map.
π― What it does: This paper proposes a speech discretization encoder SECodec based on Structural Entropy (SE) and a speech language model SESLM driven by Structural Entropy, which can automatically determine the codebook size and reduce Euclidean distance distortion through a 2D structural entropy quantization method, thereby enhancing speech reconstruction and zero-shot TTS performance.
Security Attacks on LLM-based Code Completion Tools
Wen Cheng (Nanjing University), Wei Wang (Nanjing University)
CodeSafty and PrivacyAdversarial AttackAI Code AssistantTransformerLarge Language ModelPrompt EngineeringText
π― What it does: This paper studies the security vulnerabilities of large language model-driven code completion tools (LCCT) and proposes and validates two types of attacks against LCCT: jailbreak attacks based on context information aggregation and hierarchical code exploitation, as well as code-based training data extraction attacks.
π― What it does: This study investigates cross-subject fMRI brain decoding models, using shallow adapters to map fMRI data from different subjects into a unified representational space, and then performing high-level semantic recognition and low-level pixel-level reconstruction through a shared deep decoding module.
π― What it does: This paper proposes a training-free and prompt-free offline framework called SeeDiff, which utilizes the attention maps from Stable Diffusion to generate high-quality pixel-level segmentation masks and can directly synthesize semantic segmentation datasets.
CodeRecognitionTransformerLarge Language ModelVideoMultimodality
π― What it does: The SeFAR framework is proposed, which enhances performance in semi-supervised fine-grained action recognition through dual-layer temporal element modeling, moderate temporal perturbation enhancement, and adaptive regulation.
π― What it does: This paper proposes SegFace, a face segmentation method that utilizes a lightweight transformer decoder and learnable class-specific tokens, specifically aimed at improving long-tail categories.
Selective Forgetting: Advancing Machine Unlearning Techniques and Evaluation in Language Models
Lingzhi Wang (Harbin Institute of Technology), Georg Gottlob (University of Calabria)
CodeData-Centric LearningTransformerLarge Language ModelText
π― What it does: The SEUL method is proposed to achieve selective and fine-grained forgetting in language models, and sensitive information forgetting evaluation metrics S-EL, S-MA, as well as an online/offline sensitive span automatic labeling process are designed.
π― What it does: A Selective Visual Prompting (SVP) method specifically designed for Vision Mamba is proposed and implemented for efficient fine-tuning.
Self-attention-based Diffusion Model for Time-series Imputation in Partial Blackout Scenarios
Mohammad Rafid Ul Islam (Oregon State University), Alan Fern (Oregon State University)
CodeDiffusion modelScore-based ModelTime Series
π― What it does: This paper proposes a self-attention-based diffusion model SADI for high-quality imputation of missing values in multivariate time series under the 'partial blackout' missing pattern.
π― What it does: Distill the pre-trained flow matching teacher model into a model capable of achieving single-step and few-step consistent image generation.
Self-Explainable Graph Transformer for Link Sign Prediction
Lu Li (Huazhong Agricultural University), Zeyu Zhang (Huazhong Agricultural University)
CodeRecommendation SystemExplainability and InterpretabilityGraph Neural NetworkTransformerGraph
π― What it does: Developed a self-explanatory signature graph Transformer (SE-SGformer) that can predict link symbols and provide interpretable positive/negative neighbor information.
π― What it does: This paper studies the semantic ambiguity of semi-positive sample generation in visual cross-view geographic localization, proposing an automatic balance between retrieval and offset regression joint training by modeling the uncertainty of offset regression and propagating it to the retrieval task.
π― What it does: Proposes the Semi-IIN network, which combines self-training with unsupervised data to dynamically separate and select multimodal interaction information for sentiment analysis.
π― What it does: A semi-implicit neural ODE (SINODE) framework is proposed, utilizing linear-nonlinear partitioning and IMEX Runge-Kutta integration, and achieving reversible exact gradients through discrete adjoint differentiation.
π― What it does: This paper studies image-text semi-supervised multimodal classification in scenarios with scarce labels and proposes the MSC framework.
SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning
Xinyang Liu (Shenzhen Institute of Artificial Intelligence and Robotics for Society), Bo Liu
CodeFederated LearningDiffusion modelImage
π― What it does: This paper proposes SemiDFL, a new framework for semi-supervised federated learning that operates without a central server, enabling collaborative training among clients with label scarcity and highly non-homogeneous distributions through model and data space consensus.
Sequential Conditional Transport on Probabilistic Graphs for Interpretable Counterfactual Fairness
Agathe Fernandes Machado (Universite du Quebec a Montreal), Ewen Gallic (Aix Marseille University)
CodeExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTabular
π― What it does: A sequential transport method based on conditional univariate optimal transport is proposed to generate counterfactual samples that conform to causal graphs, thereby assessing individual-level fairness.
π― What it does: This paper proposes a single-camera 2D-to-3D human pose estimation model that incorporates sequential joint dependency perception, utilizing a state space model (SSM) to integrate the sequential relationships of joints into feature learning.
π― What it does: This paper proposes a semantic-guided triple co-training framework SGTC, which achieves semi-supervised learning for medical image segmentation by labeling three orthogonal slices with a small number of volume samples.
π― What it does: Utilizing frequency-integrated priors to transform low-frequency SDF observations into high-frequency complete SDF, thereby enhancing the detail and integrity of neural implicit models.
Sharper Error Bounds in Late Fusion Multi-view Clustering with Eigenvalue Proportion Optimization
Liang Du (Shanxi University), Yuhua Qian (Shanxi University)
CodeOptimizationImageBiomedical Data
π― What it does: This paper proposes a stricter analysis of the generalization error of multiple kernel K-means using the eigenvalue ratio and local Rademacher complexity, and based on this, designs a low-pass graph filter to enhance the eigenvalue ratio, thereby improving the performance of multi-view late fusion clustering.
CodeClassificationRetrievalTransformerSupervised Fine-TuningVision Language ModelVideoTextBenchmark
π― What it does: The BestShot task is proposed, which aims to accurately locate highlight frames in human videos through language queries, and a corresponding benchmark has been constructed.
π― What it does: This paper proposes the SIGMA framework, which utilizes Partial Flip Mamba (PF-Mamba), Dense Selective Gate (DS Gate), and Feature Extraction GRU (FE-GRU) to improve sequence recommendation systems.
π― What it does: This study proposes a sign language gesture generation framework called Sign-IDD based on diffusion models, which achieves high-quality generation of sign language gestures from text gloss by transforming 3D joint representations into 4D skeletal representations and incorporating an attribute-controllable diffusion module.
Similar Modality Enhancement and Action Consistency Learning for Weakly Supervised Temporal Action Localization
Maodong Li (Wuhan University), Bing Li (Hubei Luojia Laboratory)
CodeRecognitionObject DetectionOptical FlowVideo
π― What it does: Proposes the SEAL method, which combines the SME module and the ACL module to achieve temporal action localization using weakly supervised video-level labels.
CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: By constructing prompts that include syntactic and semantic similar examples, the generation performance of Aspect Sentiment Quad Prediction (ASQP) is improved.
Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models
Haoran Li (Hong Kong University of Science and Technology), Yangqiu Song (Hong Kong University of Science and Technology)
CodeGenerationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: A backdoor removal method named SANDE is proposed for generative large language models, which can directly repair LLMs that have been implanted with backdoors without relying on clean models.
Single Image Rolling Shutter Removal with Diffusion Models
Zhanglei Yang (University of Electronic Science and Technology of China), Shuaicheng Liu (Megvii Technology)
CodeRestorationDiffusion modelOptical FlowImage
π― What it does: This paper proposes RS-Diffusion, which utilizes a diffusion model to achieve single-frame rolling shutter correction and constructs a real annotated RS-Real dataset.
π― What it does: This paper proposes a single-view graph contrastive learning framework called SIGNA, which learns node representations using 'soft neighborhood awareness' without relying on cross-view augmentation.
SKI Models: Skeleton Induced Vision-Language Embeddings for Understanding Activities of Daily Living
Arkaprava Sinha (University of North Carolina at Charlotte), Srijan Das (University of North Carolina at Charlotte)
CodeRecognitionKnowledge DistillationRepresentation LearningVision Language ModelVideoText
π― What it does: The SKI model is proposed, which achieves zero-shot recognition and video description of daily life actions by injecting 3D skeletal information into the visual-language embedding space.
SkillTree: Explainable Skill-Based Deep Reinforcement Learning for Long-Horizon Control Tasks
Yongyan Wen (Harbin Institute of Technology), Peng Liu (Kuaishou Technology)
CodeExplainability and InterpretabilityKnowledge DistillationRobotic IntelligenceReinforcement LearningTabular
π― What it does: A hierarchical interpretable deep reinforcement learning framework called SkillTree is constructed, which maps continuous action spaces to discrete skill spaces and implements high-level policies using differentiable decision trees, making the decision-making process at the skill level interpretable.
π― What it does: This paper proposes a monocular 3D semantic scene completion method based on a variational autoencoder conditional latent space and Skimba diffusion network.
SkipPool: Improved Sparse Hierarchical Graph Pooling with Differentiable Exploration
Sarith Imaduwage (Independent Researcher)
CodeGraph Neural NetworkPoint CloudGraph
π― What it does: An improved sparse hierarchical graph pooling method called SKIPPOOL is proposed, which skips over-represented areas through differentiable exploration to better preserve graph-level information.
CodeClassificationRecognitionTransformerPrompt EngineeringVision Language ModelImageMultimodality
π― What it does: This paper proposes the SLIP framework through a CLIP-based visual-language pre-training model and prompt learning, utilizing language-guided pseudo-spoof prompts to generate spoof cue maps, achieve feature decoupling, and integrate pseudo-spoof image features, addressing the performance decline caused by the lack of spoof data in single-class FAS.
π― What it does: This paper proposes a semi-supervised local community detection method based on reinforcement learning, called SLRL, which utilizes known communities to guide local structure extraction and community expansion.
Small Language Model Makes an Effective Long Text Extractor
Yelin Chen (Xinjiang University), Jie Tang (Tsinghua University)
CodeRecognitionTransformerLarge Language ModelText
π― What it does: A lightweight span-based method SeNER is proposed for entity recognition in long texts, capable of identifying long entity blocks under GPU memory-friendly conditions.
π― What it does: The paper presents SMMF, an adaptive learning rate optimizer that compresses arbitrary-order momentum tensors through square matricization and one-time matrix decomposition, significantly reducing memory usage.
π― What it does: This paper proposes a UGDA method for direct structural smoothing on the target graphβTDSS, which combines neighborhood sampling and Laplacian smoothing to enhance the robustness of node representations.
SoLA: Leveraging Soft Activation Sparsity and Low-Rank Decomposition for Large Language Model Compression
Xinhao Huang (Hong Kong University of Science and Technology), Zeyi Wen (Hong Kong University of Science and Technology)
CodeCompressionTransformerLarge Language ModelText
π― What it does: A training-free LLM compression method called SoLA is proposed, which utilizes soft activation sparsity and low-rank decomposition for fine-grained model compression, retaining a small number of highly activated neurons and performing SVD decomposition on the remaining parts, achieving significant compression rates without the need for post-training.
Solving Higher-Order Quantified Boolean Satisfiability via Higher-Order Model Checking
Hiroshi Unno (Tohoku University), Jie-Hong Roland Jiang (National Taiwan University)
CodeBenchmark
π― What it does: The first higher-order quantified Boolean satisfiability (HOQBF) solver HOMCSAT is proposed and implemented by transforming HOQBF into a higher-order model checking problem and utilizing the HORSAT2 solver to complete the solution.
SongEditor: Adapting Zero-Shot Song Generation Language Model as a Multi-Task Editor
Chenyu Yang (Chinese University of Hong Kong), Haizhou Li
CodeGenerationTransformerLarge Language ModelDiffusion modelAuto EncoderTextAudio
π― What it does: This paper presents SongEditor, a multi-task editor that extends a zero-shot song generation language model, supporting segment-wise and track-wise editing as well as generating songs from scratch.