π― What it does: By training a 3D Gaussian scene representation with only three perspective images, a depth-constrained neural language field is constructed, enabling open vocabulary 3D scene understanding and new perspective language querying and synthesis under sparse viewpoints.
π― What it does: This paper proposes a sparse cross-layer connection mechanism called SparX, and constructs two visual backbone networks, SparX-Mamba and SparX-Swin, based on this mechanism to achieve cross-layer feature fusion and reuse.
Spatial-Temporal Heterogenous Graph Contrastive Learning for Microservice Workload Prediction
Mohan Gao (Shanghai Jiao Tong University), Haoyuan Ge (Ant Group)
CodeGraph Neural NetworkContrastive LearningGraphTime Series
π― What it does: A model for microservice workload prediction called STEAM is proposed, which combines spatiotemporal heterogeneous graph contrastive learning and multi-scale learning;
π― What it does: This paper proposes a delivery recommendation framework based on spatial-temporal knowledge distillation, STKDRec, which first extracts high-order dependencies using a spatial-temporal knowledge graph (STKG), then models user dynamic preferences with a spatial-temporal Transformer, and integrates knowledge from both perspectives through knowledge distillation.
π― What it does: This paper proposes a self-supervised video denoising spatiotemporal blind spot network (STBN), which combines bidirectional optical flow alignment, spatial receptive field expansion, and unsupervised optical flow distillation to achieve global spatiotemporal information aggregation.
Spatiotemporal-aware Trend-Seasonality Decomposition Network for Traffic Flow Forecasting
Lingxiao Cao (Ocean University of China), Junyu Dong (Ocean University of China)
CodeGraph Neural NetworkTransformerTime Series
π― What it does: This paper proposes a Spatio-Temporal Aware Trend-Season Decomposition Network (STDN) that achieves traffic flow prediction through dynamic relationship graph learning, spatio-temporal embedding, and a trend-season decomposition module.
π― What it does: This paper studies multi-label incremental learning and proposes the HCP framework, which addresses the contradiction of learning objectives through dynamic feature purification, memory enhancement, and unknown knowledge mining, significantly alleviating catastrophic forgetting under no-replay conditions.
CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningBenchmarkAudio
π― What it does: This work systematically explores the key design and optimal configuration of combining large language models (LLMs) with speech encoders for automatic speech recognition (ASR), and provides a complete set of benchmark experiments.
Speedup Techniques for Switchable Temporal Plan Graph Optimization
He Jiang (Carnegie Mellon University), Jiaoyang Li (Carnegie Mellon University)
CodeOptimizationGraph
π― What it does: To address the delay issues that arise during the execution of Multi-Agent Path Finding (MAPF), this paper proposes an optimal optimization method based on the Switchable Temporal Plan Graph (STPG) and achieves significant acceleration on the original GSES search algorithm through four acceleration techniques.
SpeHeaTal: A Cluster-Enhanced Segmentation Method for Sperm Morphology Analysis
Yi Shi (Nanjing University), Rong Zeng (Nanjing University)
CodeSegmentationImage
π― What it does: This paper addresses the issues of overlapping tails and staining impurities in clinical sperm images by proposing an unsupervised sperm complete segmentation method called SPEHEATAL, which achieves complete segmentation of sperm heads and tails.
π― What it does: This paper proposes a Transformer-based spiking neural network frameworkβSpiking Point Transformer (SPT)βfor 3D point cloud classification.
π― What it does: This paper proposes SpikingSSMs, which combines sparse parallel temporal neural networks with state space models for learning long sequence tasks.
SPU-IMR: Self-supervised Arbitrary-scale Point Cloud Upsampling via Iterative Mask-recovery Network
Ziming Nie (Northwestern Polytechnical University), Jiaqi Yang (Northwestern Polytechnical University)
CodeRestorationGenerationTransformerPoint Cloud
π― What it does: This paper proposes a self-supervised iterative mask recovery network (SPU-IMR), treating point cloud upsampling as a global shape completion task, achieving arbitrary point cloud densification through mask recovery.
Spurious Feature Eraser: Stabilizing Test-Time Adaptation for Vision-Language Foundation Model
Huan Ma (Tianjin University), Bingzhe Wu (Tencent)
CodeDomain AdaptationRepresentation LearningTransformerPrompt EngineeringVision Language ModelImageMultimodality
π― What it does: This paper proposes a testing-time prompt tuning method called Spurious Feature Eraser (SEraser), which utilizes auxiliary images to eliminate spurious features (decision shortcuts) in visual-language foundation models (such as CLIP), forcing the model to focus on causally invariant features during inference.
SR-FoT: A Syllogistic-Reasoning Framework of Thought for Large Language Models Tackling Knowledge-based Reasoning Tasks
Wentao Wan (Sun Yat-sen University), Keze Wang (Sun Yat-sen University)
CodeTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: A multi-stage reasoning framework SR-FoT is proposed, enabling large language models to solve knowledge-based question-answering tasks through deductive reasoning.
π― What it does: An auxiliary task is proposed - predicting the spatial distribution map of handwritten mathematical expressions, and a Symbol Space Awareness Network (SSAN) is designed to be jointly trained with the HMER model;
π― What it does: This paper proposes a Structured Sparse Coding Variational Autoencoder (SSC-VAE), which enhances image reconstruction and denoising detail preservation by introducing a learnable threshold inference and refinement module into sparse coding.
SSE-SAM: Balancing Head and Tail Classes Gradually Through Stage-Wise SAM
Xingyu Lyu (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
CodeClassificationOptimizationImage
π― What it does: A two-stage variant of Sharpness-Aware Minimization (SAM) called SSE-SAM is proposed, which first uses SAM to help the main class escape saddle points, and then switches to ImbSAM to focus on the tail class for balanced training under long-tail distributions.
SSL-STMFormer Self-Supervised Learning Spatio-Temporal Entanglement Transformer for Traffic Flow Prediction
Zetao Li (University of Electronic Science and Technology of China), Shimin Cai (University of Electronic Science and Technology of China)
CodeAutonomous DrivingOptimizationTransformerGraphTime Series
π― What it does: A model called SSL-STMFormer based on self-supervised learning and spatiotemporal attention is proposed for traffic flow prediction.
π― What it does: A multi-modal 3D object detection framework called SSLFusion is designed and implemented, which enhances detection accuracy by aligning and fusing features from LiDAR point clouds and camera images at different scale levels.
SSUN-Net: Spatial-Spectral Prior-Aware Unfolding Network for Pan-Sharpening
Shijie Fang (Northwestern Polytechnical University), Hongping Gan (Northwestern Polytechnical University)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: This paper proposes a deep unfolding network based on spatial-spectral priors, SSUN-Net, for image pyramid fusion (pan-sharpening). By explicitly introducing physical prior constraints of PAN and MS and constructing a multi-scale prior structure, it significantly enhances the recovery of spatial and spectral details in images.
ST-FiT: Inductive Spatial-Temporal Forecasting with Limited Training Data
Zhenyu Lei (University of Virginia), Chen Chen (University of Virginia)
CodeGraph Neural NetworkAuto EncoderGraphTime Series
π― What it does: Proposes the ST-FiT framework to address the inductive prediction problem in spatial-temporal graphs when training nodes lack time series.
CodeComputational EfficiencyRepresentation LearningTransformerContrastive LearningTime Series
π― What it does: Proposes ST-ReP, a lightweight self-supervised spatio-temporal representation learning model that integrates current value reconstruction with future value prediction and incorporates multi-scale temporal supervision;
π― What it does: The Stable Mean Teacher framework is proposed, which combines student-teacher consistency, the EoR error recovery module, and DoP temporal consistency constraints in semi-supervised learning for video action detection.
π― What it does: The research explores how to simultaneously utilize collaborative information and multimodal information in e-commerce scenarios, proposing the STAIR model and implementing multimodal initialization and step graph convolution fusion.
STAMPsy: Towards SpatioTemporal-Aware Mixed-Type Dialogues for Psychological Counseling
Jieyi Wang (Peking University), Yu Huang (Peking University)
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
π― What it does: This paper proposes and implements a hybrid psychological counseling dialogue system, which first helps users clarify their counseling goals and then interacts based on different dialogue types (diagnosis, knowledge Q&A, recommendations, empathy, and Q&A).
π― What it does: A systematic comparison of state encoding schemes in heuristic learning of graph neural networks (GNN) for classical planning tasks is conducted, evaluating their expressiveness and computational efficiency in offline search, and it is found that the minimal encoding based on Gaifman graphs performs best in the IPC 2023 learning track.
STD-PLM: Understanding Both Spatial and Temporal Properties of Spatial-Temporal Data with PLM
Yiheng Huang (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
CodeData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTime Series
π― What it does: This paper proposes STD-PLM, a unified pre-trained language model framework for simultaneously performing spatial-temporal prediction and imputation tasks.
STEM-LTS: Integrating Semantic-Temporal Dynamics in LLM-driven Time Series Analysis
Zhe Zhao (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
CodeTransformerLarge Language ModelPrompt EngineeringTime SeriesFinance Related
π― What it does: The STEM-LTS framework is proposed, which combines temporal decomposition, LLM semantic-temporal alignment, and multi-task learning to achieve high-precision forecasting of high-dimensional multi-scale time series.
Step-Calibrated Diffusion for Biomedical Optical Image Restoration
Yiwei Lyu (University of Michigan), Todd C. Hollon (University of Michigan)
CodeRestorationDiffusion modelImageBiomedical Data
π― What it does: A method for unpaired optical image restoration called RSCD is proposed, utilizing the reverse diffusion process of diffusion models and dynamic step calibration to achieve denoising and reconstruction of Raman-based optical images.
π― What it does: Proposes STGC-NeRF, which enhances the NeRF reconstruction of dynamic scenes from LiDAR through spatial-temporal geometric consistency.
STLC-KG:A Social Text Steganalysis Method Combining Large-Scale Language Models and Common-Sense Knowledge Graphs
Zhuang Wang (Beijing University of Posts and Telecommunications), Zhongliang Yang (Guangzhou University)
CodeClassificationAnomaly DetectionGraph Neural NetworkTransformerLarge Language ModelContrastive LearningText
π― What it does: A social text steganography detection framework STLC-KG is proposed and implemented, which combines large language models with common sense knowledge graphs. After expanding the text semantics using the knowledge graph, semantic features are extracted using ChatGLM2-6B, and knowledge graph features are encoded through GAT, ultimately completing the detection of steganographic text.
StoryWeaver: A Unified World Model for Knowledge-Enhanced Story Character Customization
Jinlu Zhang (Xiamen University), Xiaoshuai Sun (Xiamen University)
CodeGenerationGraph Neural NetworkTransformerVision Language ModelDiffusion modelImageGraphBenchmark
π― What it does: A unified story visualization framework called StoryWeaver is proposed, which utilizes a Character-Graph to encode characters, attributes, and relationships in the story world, generating knowledge-enhanced scene descriptions, and refines cross-attention through Knowledge Enhanced Spatial Guidance (KE-SG) to avoid identity entanglement of multiple characters.
π― What it does: This paper proposes a plug-in method called StructSR that does not require fine-tuning, external priors, or semantic knowledge to improve the structural fidelity of diffusion models in real image super-resolution and suppress pseudo-details.
π― What it does: This paper proposes a structure entropy-guided probabilistic encoding model (SEPC), which combines structure entropy regularization with a probabilistic encoding tree for classification and regression tasks in natural language understanding.
π― What it does: We propose SEGO, an unsupervised graph structure OOD detection framework that uses a coding tree obtained through structural entropy minimization as an anchor view, and distinguishes between ID and OOD graphs through multi-granularity (node, graph, tree) contrastive learning.
π― What it does: This paper proposes a filter pruning method based on spatially aware information redundancy (SIRFP), specifically addressing the pruning needs of semantic segmentation models.
π― What it does: Designed and validated the Structured Information Bottleneck (SIB) framework, which enhances the representation capability of traditional IB methods using auxiliary encoders.
Structured Packing in LLM Training Improves Long Context Utilization
Konrad Staniszewski (University of Warsaw), Piotr MiΕoΕ (University of Warsaw)
CodeRetrievalOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityRetrieval-Augmented Generation
π― What it does: This paper proposes and implements SPLICE, a structured packaging method that stitches semantically related documents into long training samples through retrieval, aimed at enhancing the long context utilization capability of large language models.
π― What it does: The SIURec model is proposed, utilizing sub-interest-aware uniformization for hierarchical representation learning in recommendation systems.
Subgraph Invariant Learning Towards Large-Scale Graph Node Classification
Leilei Wang (Shenzhen University), Ying Tiffany He (Hong Kong University of Science and Technology Guangzhou)
CodeClassificationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerContrastive LearningGraph
π― What it does: Proposes a Subgraph Invariant Learning framework that utilizes subgraph training with an Invariance Encoder (IRE) and Node Encoder (NRE) to achieve large-scale graph node classification, significantly reducing computational costs and enhancing out-of-distribution (OOD) generalization.
Lorenzo Loconte (University of Edinburgh), Antonio Vergari (University of Artois)
CodeImageTabular
π― What it does: This paper studies the expressive power of differentiable probabilistic circuits (PCs), proposes a new model called 'sum of squares (SOS) PCs', and constructs an expression hierarchy, proving that it significantly enhances expressiveness compared to single square PCs and monotonic PCs.
SUMO: Search-Based Uncertainty Estimation for Model-Based Offline Reinforcement Learning
Zhongjian Qiao (Tsinghua University), Xiu Li (Tsinghua University)
CodeReinforcement LearningTabular
π― What it does: A search-based uncertainty estimation method called SUMO is proposed, which is integrated into model-based offline reinforcement learning (such as MOPO, AMOReL). It estimates the cross-entropy between model-generated dynamics and real dynamics through k-NN search, thereby enhancing policy performance.
π― What it does: This paper proposes a unified framework for online workflow recognition and effectiveness detection of the Pringle maneuver in laparoscopic liver resection, and constructs the corresponding PmLR50 dataset.
π― What it does: This paper proposes an adaptive diversity adversarial style perturbation method (SVasP), which stabilizes the global style adversarial updates by utilizing locally cropped style gradients in single-source cross-domain few-shot learning, thereby enhancing the model's domain transfer performance.
SVTformer: Spatial-View-Temporal Transformer for Multi-View 3D Human Pose Estimation
Wanruo Zhang (Peking University), Wenhao Li (Nanyang Technological University)
CodePose EstimationTransformerImage
π― What it does: This paper proposes SVTformer, which utilizes a spatial-view-time three-dimensional transformer for 3D pose recovery from multi-view 2D skeletons.
π― What it does: This paper proposes the SymmCompletion framework, which achieves high-fidelity and high-consistency point cloud completion through the Local Symmetric Transformation Network (LSTNet) and the Symmetric-guided Transformer (SGFormer).
CodeGenerationRetrievalExplainability and InterpretabilityTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: The SMART multi-agent framework is proposed, where four specialized agents collaborate to complete knowledge-intensive tasks through intent reconstruction, knowledge retrieval, fact localization, and response generation, significantly enhancing the factual consistency and interpretability of the generated results.
CodeGenerationDrug DiscoveryGraph Neural NetworkFlow-based ModelBiomedical Data
π― What it does: A framework for antibody design called PG-AbD is proposed, which combines generative flow networks with pre-trained protein language models to generate highly developable, diverse, and novel antibody CDRs.
π― What it does: This study focuses on synthetic data augmentation in extremely imbalanced tabular data, proposing the Overlap Region Detection (ORD) method. It changes the binary classification labels to three classes by introducing an overlapping class, using a conditional generator to produce higher quality minority class samples, and training the classifier after removing overlapping majority class samples.
π― What it does: This paper proposes a method that simultaneously converts each record into a fully connected graph and serialized text, and then performs joint table representation learning using graph neural networks and a pre-trained text encoder.
CodeTransformerLarge Language ModelSupervised Fine-TuningTabularBenchmarkChain-of-Thought
π― What it does: A new complex table question-answering benchmark named TableBench has been constructed, and the TableInstruct instruction corpus has been created to train the TABLELLM model, which is used to evaluate the table reasoning capabilities of LLMs in industrial scenarios.
Tackling Intertwined Data and Device Heterogeneities in Federated Learning with Unlimited Staleness
Haoming Wang (University of Pittsburgh), Wei Gao (University of Pittsburgh)
CodeFederated LearningImage
π― What it does: A gradient inversion-based federated learning framework is proposed to convert delayed model updates into non-delayed updates, thereby addressing the issue of intertwined data and device heterogeneity.
π― What it does: The TAMER model is proposed, which learns sequences and tree structures in parallel through a Tree-Aware Module and Transformer to recognize handwritten mathematical expressions.
Target Semantics Clustering via Text Representations for Robust Universal Domain Adaptation
Weinan He (University of Science and Technology of China), Yixin Zhang (University of Science and Technology of China)
CodeDomain AdaptationContrastive LearningText
π― What it does: This paper proposes a Text-based Target Semantic Clustering (TASC) framework for achieving semantic clustering and domain alignment in Unified Domain Adaptation (UniDA).
Task-level Distributionally Robust Optimization for Large Language Model-based Dense Retrieval
Guangyuan Ma (Institute of Information Engineering Chinese Academy of Sciences), Songlin Hu (Langboat Technology)
CodeRetrievalOptimizationTransformerLarge Language ModelContrastive LearningText
π― What it does: In multi-task multi-source training data, an end-to-end data distribution optimization method called tDRO is proposed for large language model dense retrieval (LLM-DR) to enhance general domain generalization capabilities.
π― What it does: A pre-training-free TaylorSeg network is proposed, which uses Taylor series-inspired TaylorConv to extract local structures of point clouds, and aligns query and prototype features through the APP module to achieve few-shot point cloud semantic segmentation.
TC-Diffuser: Bi-Condition Multi-Modal Diffusion for Tropical Cyclone Forecasting
Shiqi Zhang (Zhejiang University of Technology), Cong Bai (Shandong University)
CodeGenerationData SynthesisRecurrent Neural NetworkTransformerDiffusion modelMultimodalityTime Series
π― What it does: A dual-condition multimodal diffusion model (TC-Diffuser) is proposed for predicting the trajectory, pressure, and wind speed of tropical cyclones.
π― What it does: A task-oriented Top-Down Attention Guided Mixup (TdAttenMix) data augmentation method is proposed, which utilizes human eye gaze to guide attention balance between top-down and bottom-up attention for selecting mixing regions and dynamically adjusting label ratios.
π― What it does: This paper proposes a teacher-guided edge discriminator and a personalized graph mask autoencoder (TEDMAE) for self-supervised learning of node representations in graphs with a mix of homogeneous and heterogeneous edges.
Liat Bezalel (Tel Aviv University), Amir Globerson (Tel Aviv University)
CodeGenerationMeta LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper designs and implements CORGI, a framework for LLM trained through interactive critical feedback and reinforcement learning (PPO) to improve the accuracy of constrained text generation.
π― What it does: A controllable singing synthesis system called TechSinger is proposed, which is based on flow matching and allows for fine control of singing techniques through natural language prompts.
Template-Driven LLM-Paraphrased Framework for Tabular Math Word Problem Generation
Xiaoqiang Kang (Xi'an Jiaotong-Liverpool University), Qiufeng Wang (Xi'an Jiaotong-Liverpool University)
CodeGenerationData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningTabular
π― What it does: A framework based on template-driven and LLM rephrasing (TeLL) is proposed for the automatic generation of table mathematics problems with high correctness and diversity, and based on this, the TabMWP-TeLL dataset is constructed.
π― What it does: A temporal action localization framework based on Cross-Layer Task Decoupling and Refinement (CLTDR) and Gated Multi-Granularity Features (GMG) is proposed, which can decouple action classification and localization tasks at the feature level and further align the predictions of both through a refinement head.
Matteo Cardellini (University of Genoa), Enrico Giunchiglia (University of Genoa)
CodeOptimization
π― What it does: This paper extends Symbolic Pattern Planning (SPP) to temporal numeric planning, providing a set of SMT-based encodings and proving their correctness and completeness.
Temporal Specification Optimisation for the Event Calculus
Periklis Mantenoglou (NCSR Demokritos), Alexander Artikis (University of Piraeus)
CodeOptimizationTabularTime Series
π― What it does: A compiler has been constructed to convert simple fluents in RTEC into statically determined fluents, thereby optimizing event descriptions.
Temporal Task and Motion Planning with Metric Time for Multiple Object Navigation
Elisa Tosello (Fondazione Bruno Kessler), Andrea Micheli (Fondazione Bruno Kessler)
CodeOptimizationRobotic IntelligenceSimultaneous Localization and MappingTabularBenchmark
π― What it does: This paper studies the time task and motion planning problem for multi-object simultaneous navigation (TTAMP) and proposes a novel interactive solver T-TAMPEST.
TEncDM: Understanding the Properties of the Diffusion Model in the Space of Language Model Encodings
Alexander Shabalin (HSE University), Dmitry Vetrov (Constructor University)
CodeGenerationTransformerDiffusion modelText
π― What it does: Train diffusion models in the encoding space of pre-trained language models and propose the TEncDM framework, which combines Transformer decoders with techniques such as self-conditioning and noise scheduling.
π― What it does: This paper proposes the TIMO method, which utilizes the mutual guidance of image and text features from CLIP (Image-Guided-Text and Text-Guided-Image) to enhance training-free few-shot classification performance.
Text Proxy: Decomposing Retrieval from a 1-to-N Relationship into N 1-to-1 Relationships for Text-Video Retrieval
Jian Xiao (Hefei University of Technology), Richang Hong (Hefei University of Technology)
CodeRetrievalTransformerVision Language ModelContrastive LearningVideoTextMultimodality
π― What it does: This paper proposes TV-ProxyNet, which decomposes the original 1-to-N text-video retrieval relationship into N 1-to-1 relationships by splitting the text query into a series of text proxies, thereby achieving finer semantic alignment.
π― What it does: This paper proposes a multi-layer negative contrastive learning framework that uses language as a filter rather than directly aligning text with point cloud features, significantly improving text-point cloud localization accuracy at the urban scale.
Text2Data: Low-Resource Data Generation with Textual Control
Shiyu Wang (Salesforce AI Research), Silvio Savarese (Salesforce AI Research)
CodeGenerationData SynthesisDiffusion modelTabularTime SeriesFinance Related
π― What it does: This paper proposes a framework called Text2Data, which utilizes unlabeled data to pre-train a diffusion model and then fine-tunes it with a small amount of labeled data for text-to-data generation in low-resource scenarios.
Text2midi: Generating Symbolic Music from Captions
Keshav Bhandari (Queen Mary University of London), Dorien Herremans (Singapore University of Technology and Design)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTextAudio
π― What it does: An end-to-end text2midi model has been developed that can directly generate high-quality MIDI files based on natural language descriptions.
π― What it does: A saliency prediction method for texture meshes is proposed, and the first texture mesh saliency dataset based on a 6DOF VR eye-tracking experiment is constructed.
The Distributional Reward Critic Framework for Reinforcement Learning Under Perturbed Rewards
Xi Chen (Ohio State University), Andrew Perrault (Ohio State University)
CodeReinforcement LearningSequential
π― What it does: This paper proposes the Distributional Reward Critic (DRC) framework and its general version (General Distributional Reward Critic, GDRC) for training reinforcement learning (RL) agents in environments where rewards are subject to unknown disturbances (Generalized Confusion Matrix, GCM) or continuous noise. By transforming the reward estimation into a classification task and using Ordinal Cross-Entropy (OCE), this method can estimate the reward distribution online and replace actual rewards with estimated rewards, enhancing learning robustness.
π― What it does: A general gradient definition βAMC based on Algebraic Model Counting (AMC) is proposed, and an optimized backpropagation algorithm that can be executed on various semirings is designed based on this.
The Illusion of Empathy: How AI Chatbots Shape Conversation Perception
Tingting Liu (National Institute on Drug Abuse), JoΓ£o Sedoc (New York University)
CodeTransformerLarge Language ModelText
π― What it does: This study compares the differences in dialogue quality and user emotional perception between humans and the GPT-4o chatbot, exploring the impact of emotional perception on dialogue quality.
The Impact of Literal Sorting on Cardinality Constraint Encodings
Joseph E. Reeves (Carnegie Mellon University), Marijn J. H. Heule (University of Amsterdam)
Code
π― What it does: This paper proposes various automated methods for sorting literals in cardinality constraint encoding to change the semantics of auxiliary variables and enhance the learning effectiveness of CDCL solvers.
The Indoor-Training Effect: Unexpected Gains from Distribution Shifts in the Transition Function
Serena Bono (Massachusetts Institute of Technology), Gabriel Kreiman (Harvard University)
CodeReinforcement LearningSequential
π― What it does: This study investigates the phenomenon of training environments that are not completely identical to testing environments in reinforcement learning, known as the 'indoor training effect.' By adding controllable Gaussian noise to the transition function of the original MDP to generate various Ξ΄-environments, the performance differences between training in a noise-free environment and testing in a noisy environment (Generalization Agent) versus training and testing in the same noisy environment (Learnability Agent) are compared.
The VOROS: Lifting ROC Curves to 3D to Summarize Unbalanced Classifier Performance
Christopher Ratigan (Tufts University), Lenore Cowen (Tufts University)
CodeClassificationAnomaly DetectionTabular
π― What it does: A new evaluation metric called VOROS (Volume over the ROC Surface) is proposed, which elevates the traditional ROC curve to three-dimensional space to simultaneously consider the performance of classifiers under different costs and class imbalances.
Thin-Plate Spline-based Interpolation for Animation Line Inbetweening
Tianyi Zhu (Harbin Institute of Technology), Dongwei Ren (Tianjin University)
CodeOptical FlowImageVideo
π― What it does: Proposes an animation line drawing intermediate frame interpolation method based on Thin Plate Spline (TPS) and a motion refinement module;
Thinking in Granularity: Dynamic Quantization for Image Super-Resolution by Intriguing Multi-Granularity Clues
Mingshen Wang (Hefei University of Technology), Meng Wang (Anhui University)
CodeRestorationSuper ResolutionImage
π― What it does: The Granular-DQ framework is proposed to achieve layer-invariant dynamic quantization based on multi-granularity features and entropy statistics for single image super-resolution networks.
Thought-Path Contrastive Learning via Premise-Oriented Data Augmentation for Logical Reading Comprehension
Chenxu Wang (Beijing Institute of Technology), Zhen Yang (Beijing Institute of Technology)
CodeLarge Language ModelSupervised Fine-TuningContrastive LearningTextChain-of-Thought
π― What it does: A premise-based Contrastive Generation and Thought Comparison Learning (PODA-TPCL) framework is proposed to enhance logical reading comprehension abilities.
π― What it does: This paper proposes a graph data augmentation method based on spectral theory called Dual-Prism (DP), which achieves global deformation and attribute preservation of graph structures by modifying only high-frequency features while keeping low-frequency spectra unchanged.
π― What it does: A deep neural network layer pruning method called TLC based on batch normalization parameters is proposed, which removes unimportant layers by evaluating the ON/OFF status of each layer, thereby reducing the network depth.
Tilted Quantile Gradient Updates for Quantile-Constrained Reinforcement Learning
Chenglin Li (Tsinghua University), Hua Geng (Tsinghua University)
CodeReinforcement Learning
π― What it does: This paper proposes a quantile-constrained reinforcement learning method TQPO based on tilted quantile gradient updates, which directly samples to estimate the quantile gradient and addresses the overly conservative problem through tilted updates.
Time Series Supplier Allocation via Deep Black-Litterman Model
Xinke Jiang (University of Electronic Science and Technology of China), Jiayuan Luo (Zhongnan University of Economics and Law)
CodeRecommendation SystemOptimizationGraph Neural NetworkTime SeriesFinance Related
π― What it does: This paper proposes a Deep Black-Litterman Model (DBLM) for Time Series Supplier Allocation (TSSA), which achieves automatic learning and allocation decisions for the future supplier perspective matrix by combining the traditional Black-Litterman portfolio model with spatiotemporal graph neural networks.
TIME-FS: Joint Learning of Tensorial Incomplete Multi-View Unsupervised Feature Selection and Missing-View Imputation
Yanyong Huang (Southwestern University of Finance and Economics), Tianrui Li (Southwest Jiaotong University)
CodeOptimizationRepresentation LearningTabular
π― What it does: A joint learning framework is proposed that can simultaneously complete missing views, perform unsupervised feature selection, and learn low-dimensional representations on incomplete multi-view data.
TimeCMA: Towards LLM-Empowered Multivariate Time Series Forecasting via Cross-Modality Alignment
Chenxi Liu (Nanyang Technological University), Rui Zhao (SenseTime Research)
CodeTransformerLarge Language ModelPrompt EngineeringTime Series
π― What it does: This paper proposes the TimeCMA framework, which utilizes cross-modal alignment combined with LLM and time series encoders to achieve multivariate time series forecasting.
TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data
Ege Onur Taga (University of Michigan), Samet Oymak (University of Michigan)
CodeData SynthesisOptimizationTransformerTime Series
π― What it does: A multivariate time series zero/few-shot prediction framework called TimePFN based on synthetic data and Transformer is proposed, along with a unified model training and fine-tuning process.
Times2D: Multi-Period Decomposition and Derivative Mapping for General Time Series Forecasting
Reza Nematirad (Kansas State University), Balasubramaniam Natarajan (Kansas State University)
CodeConvolutional Neural NetworkTransformerTime Series
π― What it does: This paper proposes a time series forecasting framework named Times2D, which transforms 1D sequences into 2D representations through frequency domain periodic decomposition and derivative heatmap mapping, and then utilizes 2D convolution and self-attention networks for prediction.
π― What it does: A forward training algorithm for device learning, TinyFoA, is proposed for efficiently training neural networks on mobile/IoT devices.
π― What it does: A lightweight TinySAM model was constructed using techniques such as full-stage knowledge distillation, hard mask weighting and hard prompt sampling, post-training quantization, and hierarchical everything reasoning, achieving a significant reduction in computational load while maintaining the zero-shot segmentation performance of SAM.