π― What it does: Proposes a 4D occupancy world model called SparseWorld based on sparse dynamic queries for continuous perception, prediction, and planning.
Spatial-Temporal Feedback Diffusion Guidance for Controlled Traffic Imputation
Xiaowei Mao (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
CodeGraph Neural NetworkDiffusion modelTime Series
π― What it does: Proposed a traffic missing value imputation method called FENCE based on space-time feedback diffusion guidance, and verified its superiority on real traffic datasets.
π― What it does: Propose the SpatialActor framework to decouple semantics and geometry during robotic manipulation, and build a robust spatial representation through a semantics-guided geometric module and spatial transformer.
π― What it does: Propose a multi-frame infrared small target detection network named TDCNet, which achieves high-precision detection of moving infrared small targets by using a TDCR module that integrates temporal difference convolution with 3D convolution, and a TDC-guided spatiotemporal attention mechanism.
Xintong Li (Renmin University of China), Xiao Zhou (Renmin University of China)
CodeGraph Neural NetworkGraphTime Series
π― What it does: Propose a Spatio-Temporal Hierarchical Causal Model (ST-HCM) that achieves causal inference in observational data with time-invariant unobserved confounding and spatial interventions through hierarchical structures and spatiotemporal dependencies.
π― What it does: This paper addresses the spatial and temporal long-tail distribution problems in video depth super-resolution by proposing a dual-branch SpatioTemporal Difference Network (STDNet), achieving high-quality high-resolution depth video restoration.
π― What it does: Proposes a novel spectral-based learning method (LLwLC), which enhances the expressiveness of graph neural networks by embedding feature subgraphs with linear constraints (such as vertex deletion subgraphs and Neumann constraints), and applies it to the link prediction task.
π― What it does: Propose a spectral attribute-driven data augmentation framework (SPDDA), which generates diverse and realistic extended samples by simulating variations in the number of channels and mixing of adjacent channels in single-source hyperspectral images, thereby enhancing single-source domain generalization capabilities.
Speech-Aware Long Context Pruning and Integration for Contextualized Automatic Speech Recognition
Yiming Rong (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences), Bo Xu (Key Laboratory of Cognition and Decision Intelligence for Complex Systems Institute of Automation Chinese Academy of Sciences)
CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningTextAudio
π― What it does: Investigated a two-stage speech-aware long context pruning and integration framework, SAPΒ², which can actively filter and compress long text contexts to enhance speech recognition accuracy.
π― What it does: Proposed and implemented the Spike Stream Memory Transfer (SSMT) framework to reconstruct high-frame-rate, clear dynamic scene images from discrete pulse streams of neuromorphic spike cameras.
π― What it does: Proposed SpikingHAN, which integrates spiking neural networks (SNN) with heterogeneous graph attention networks, achieving node classification through single-layer shared convolution, semantic-level attention, and SNN to generate 1-bit representations;
π― What it does: Propose a hybrid ANN-SNN network called SWS-Net, combining the noise suppression and time integration capabilities of Spiking Neural Networks (SNNs) with the feature extraction advantages of traditional Artificial Neural Networks (ANNs) to achieve efficient and robust WiFi sensing;
π― What it does: Proposed Spikingformer, a full-spike Transformer architecture integrating MS Residual and self-attention, addressing the non-spike computation issues in models like Spikformer;
SPIRAL: Symbolic LLM Planning via Grounded and Reflective Search
Yifan Zhang (Vanderbilt University), Achille Fokoue (IBM T.J. Watson Research Center)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: Propose a framework named SPIRAL that integrates three types of LLM agents (Planner, Simulator, Critic) into the MCTS loop, achieving symbolic planning through 'rooted reflective search';
π― What it does: Proposed an efficient network called SPJFNet for dark image recovery, which significantly reduces the number of parameters and computational cost while maintaining or improving visual quality.
π― What it does: This study proposes the SplatSSC framework for 3D semantic scene completion from monocular images, addressing the issues of redundancy and 'floating bodies' caused by traditional random initialization of Gaussian primitives;
π― What it does: Proposes the Split-Layer mechanism, which splits each fully connected layer into multiple parallel branches and aggregates them through Hadamard product to construct a high-order polynomial feature space, thereby significantly enhancing the representational power of implicit neural representations (INR).
Spontaneous Yet Predictable: Shapelet-Driven, Channel-Aware Intention Decoding from Multi-Region ECoG
Keren Cao (Xi'an Jiaotong University), Liangjun Chen (Xi'an Jiaotong University)
CodeClassificationExplainability and InterpretabilityTransformerContrastive LearningBiomedical Data
π― What it does: This study proposes a shapelet-driven, channel-aware dual-region ECoG intent decoding framework that can accurately predict spontaneous vocalizations of marmosets 200 ms in advance.
π― What it does: This paper proposes a semi-pull supervised contrastive learning (SPP-SCL) framework, which first performs contrastive alignment of same-sentiment category samples within the image and text modalities individually, then conditionally pulls cross-modal sentiment representations closer based on similarity threshold conditions, thereby achieving intra-modal and cross-modal consistency in the sentiment embedding space before fusion; meanwhile, hierarchical attention (HA) and cross-modal fusion (CMF) modules are combined to enhance feature expression and fusion effects.
SRD: Reinforcement-Learned Semantic Perturbation for Backdoor Defense in VLMs
Shuhan Xu (Wuhan University), Dacheng Tao (AGH University of Krakow)
CodeAdversarial AttackTransformerReinforcement LearningVision Language ModelMultimodality
π― What it does: To defend against backdoor attacks on vision-language models, the authors propose a semantic reward defense framework based on reinforcement learning, which generates red mask perturbations on input images to disrupt the model's attention distribution, thereby suppressing trigger activation and reducing attack success rate.
SSHPool: The Separated Subgraph-based Hierarchical Pooling
Zhuo Xu, Edwin R. Hancock (University Of York)
CodeClassificationGraph Neural NetworkGraph
π― What it does: Proposed a hierarchical pooling method called SSHPool based on separated subgraphs for graph classification, and constructed a complete end-to-end GNN framework.
SSL-CST: Cell Segmentation for Single-Cell Spatial Transcriptome Based on Self-Supervised Learning
Weiliang Huo (Hainan University), Qingchen Zhang (Hainan University)
CodeSegmentationTransformerMultimodalityBiomedical Data
π― What it does: Proposed a self-supervised learning-based method for single-cell spatial transcriptomics cell segmentation called SSL-CST, which accurately segments cells by leveraging nuclear staining images and gene expression information;
π― What it does: Develop a retrieval-based semi-supervised Ultra-High-Resolution (UHR) image segmentation framework called SSR-SAM based on the Segment Anything Model (SAM), which generates visual semantic prompts using locally annotated regions and retrieves similar pixels across the entire image, further achieving consistency regularization through prompt layer perturbation;
Stabilizing Policy Gradient Methods via Reward Profiling
Shihab Ahmed (University of Central Florida), Aritra Dutta (University of Central Florida)
CodeRobotic IntelligenceReinforcement Learning
π― What it does: Proposed and implemented a general reward profiling framework to selectively update policies in policy gradient (PG) methods based on high-confidence performance evaluation, thereby reducing gradient estimation variance, improving convergence speed, and enhancing learning stability.
π― What it does: Proposes a self-supervised diffusion model training method based on Latent Space Filtering (LSF), which utilizes the low-dimensional structural degradation in the latent space to filter out unrealistic synthetic samples, thereby suppressing model collapse and maintaining generation quality.
Stable and Adaptive Fusion for Multi-domain Multi-task Recommendation
Ke Fei (Tencent), Jingjing Li (Tencent)
CodeRecommendation SystemMixture of ExpertsTabular
π― What it does: Designed and implemented the Stable and Adaptive Fusion (SAF) framework to address the negative transfer problem in multi-domain multi-task recommendation.
Wesley H. Holliday (University of California, Berkeley), Cynthia Wang (Carnegie Mellon University)
Code
π― What it does: Investigated the relationship between Simple Stable Voting (SSV) and Splitting Cycle (SC) methods, proving that when the number of candidates does not exceed 6, the winner of SSV is necessarily the winner of SC, and provided a minimal counterexample for 7 or more candidates.
Stage-Aware Graph Contrastive Learning with Node-oriented Mixture of Experts
Xiangkai Zhu, Longsheng Su (Shandong University Of Science And Technology)
CodeRepresentation LearningGraph Neural NetworkLarge Language ModelMixture of ExpertsContrastive LearningGraph
π― What it does: Propose Stage-Aware Graph Contrastive Learning (SAGCL), which combines multilingual models through Node-oriented Mixture of Experts (NodeMoE), and uses self-supervised contrastive learning to align LLM embeddings with graph structures during the feature transformation and propagation stages of GNN.
Qingyang Yan (Huazhong University of Science and Technology), Yixiong Zou (Huazhong University of Science and Technology)
CodeObject DetectionReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
π― What it does: For the visual localization task, the impact of Chain-of-Thought (CoT) generation on model performance is studied, and a curriculum learning strategy named CuRPO based on CoT length and reward is proposed. Reinforcement learning (GRPO) is used to gradually introduce data with increasing difficulty, improving localization accuracy.
π― What it does: This paper proposes an adaptive fine-tuning method called SPA based on state proficiency, aimed at performing online fine-tuning from offline-trained policies, with the goal of significantly improving sample efficiency and final performance while maintaining training stability.
π― What it does: This paper proposes a multivariate time series anomaly detection framework named SDA-D that simultaneously leverages time-point reconstruction error and system state derivatives.
State-Space Hierarchical Compression with Gated Attention and Learnable Sampling for Hour-Long Video Understanding in Large Multimodal Models
Geewook Kim (NAVER Cloud AI), Minjoon Seo (KAIST AI)
CodeCompressionLarge Language ModelVideoMultimodalityBenchmark
π― What it does: Propose a multi-level compression framework named MambaMia, which compresses a large number of frame features from hour-level long videos before inputting them into large-scale multimodal models, thereby alleviating the token explosion problem.
π― What it does: Designed and implemented SODEC, a single-step diffusion image compression model that combines high-information potential codes generated by VAE with one-time diffusion decoding, along with a fidelity guidance module and rate annealing training.
CodeGenerationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: This study proposes a lightweight dynamic steering mechanism that injects a steering vector into the pre-trained drafter by leveraging the hidden representations of the verifier, significantly improving the token acceptance rate and overall inference throughput during speculative decoding.
Steering Visuomotor Policy in Open Worlds via Cross-View Goal Alignment
Shaofei Cai (Peking University), Yitao Liang (Peking University)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision-Language-Action ModelImageVideoBenchmark
π― What it does: Propose a cross-view goal alignment framework that allows humans to specify goals using segmentation masks from their own camera views, training visual-motor policies that can complete tasks based on their own observations, significantly improving human-robot interaction efficiency.
π― What it does: Proposes two zero-shot music editing methods, SteerMusic and SteerMusic+, which directly edit raw music in the data space using score distillation and delta denoising score.
Step Back to Leap Forward: Self-Backtracking for Symbolic Reasoning and Planning in Language Models
Xiao-Wen Yang (Nanjing University), Yu-Feng Li (Nanjing University)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextSequential
π― What it does: Propose a self-backtracking mechanism that enables large language models to automatically determine when to backtrack and improve search paths in reasoning and planning tasks.
Faizan Farooq Khan, Balaji Vasan Srinivasan (King Abdullah University of Science and Technology)
CodeSegmentationGenerationData SynthesisTransformerLarge Language ModelVision Language ModelDiffusion modelMultimodalityBenchmark
π― What it does: This study proposes a system called SLEDGE that can update graphic designs based on step-by-step instructions, and constructs the IDeation dataset and benchmark for training and evaluation.
StepFun-Formalizer: Unlocking the Autoformalization Potential of LLMs Through Knowledge-Reasoning Fusion
Yutong Wu (Institute of Computing Technology, Chinese Academy of Sciences), Xing Hu (Institute of Computing Technology, Chinese Academy of Sciences)
CodeKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: This paper proposes a data synthesis and training pipeline named ThinkingF to improve the accuracy of large models in automatic formalization of natural language mathematical expressions (autoformalization).
Stepwise Contrastive Reasoning for Retrieval-Augmented Generation over Knowledge Graphs
Chenxiao Lin, Qingqiang Wu (Xiamen University)
CodeExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerContrastive LearningGraphRetrieval-Augmented Generation
π― What it does: This paper proposes a lightweight retrieval-augmented generation framework called Stepwise Contrastive Reasoning (SCR), which achieves interpretable reasoning and knowledge retrieval on knowledge graphs by progressively aligning the semantic embeddings of questions and graph entities;
π― What it does: This paper theoretically explains the transferability of universal adversarial perturbations (UAP), proposes and analyzes the relationship between transferability gap and algorithm stability, and designs an Expected Constraint and Noisy Stochastic Universal Adversarial Perturbation (SUAP) algorithm based on this.
STOLA: Self-Adaptive Touch-Language Framework for Tactile Commonsense Reasoning in Open-Ended Scenarios
Ning Cheng (Beijing Jiaotong University), Wenjuan Han (Beijing University of Posts and Telecommunications)
CodeTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsMultimodalityPhysics Related
π― What it does: Proposes STOLAβan adaptive tactile-language framework that utilizes Mixture of Experts (MoE) and two-stage training to achieve open-ended tactile common-sense reasoning.
Stop Mixing Things Up! BISCUIT Teaches Vision-Language Models to Learn New Concepts from Images on the Spot
Jiahua Bao (Harbin Institute of Technology), Jie Liu (Harbin Institute of Technology)
CodeTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextBenchmark
π― What it does: BISCUIT proposes a two-step training method that enables vision-language models to instantly learn and utilize new concepts appearing in images during inference, without relying on text injection or vocabulary expansion.
Streaming Generated Gaussian Process Experts for Online Learning and Control
Zewen Yang (Technical University Of Munich), Sami Haddadin (Technical University Of Munich)
CodeComputational EfficiencyMixture of ExpertsTabular
π― What it does: Proposed the SkyGP framework, which can online generate and maintain a finite number of Gaussian Process experts in a streaming data environment, achieving efficient online learning and control;
Streaming Generation of Co-Speech Gestures via Accelerated Rolling Diffusion
Evgeniia Vu (Constructor University), Dmitry Vetrov (Constructor University)
CodeGenerationDiffusion modelMultimodality
π― What it does: This study proposes a framework for real-time generation of co-speech gestures, achieving continuous long-sequence synthesis using accelerated rolling diffusion technology.
StreamKV: Streaming Video Question-Answering with Segment-based KV Cache Retrieval and Compression
Yilong Chen (Peking University), Ming Lu (Peking University)
CodeRetrievalCompressionTransformerLarge Language ModelPrompt EngineeringVision Language ModelVideo
π― What it does: This paper proposes StreamKV, a no-training framework that provides dynamic semantic segmentation, KV cache retrieval, and compression for video large language models, thereby achieving efficient streaming video question answering.
StreamSTGS: Streaming Spatial and Temporal Gaussian Grids for Real-Time Free-Viewpoint Video
Zhihui Ke (Tianjin University), Tie Qiu (Tianjin University)
CodeCompressionTransformerGaussian SplattingVideo
π― What it does: A dynamic 3D Gaussian representation based on StreamSTGS (Stream-based Spatiotemporal Gaussian Grids) is proposed for real-time free-viewpoint video streams.
π― What it does: Propose Strip R-CNN, a framework that utilizes large-scale strip convolution to enhance high aspect ratio object detection in remote sensing images.
π― What it does: Proposed a two-stage vector sketch generation framework called StrokeFusion, which first uses dual-modal stroke-UDF encoding to jointly represent stroke geometry and unsigned distance fields, then generates strokes through an unordered, variable-length latent space diffusion model, enabling synchronized learning of position, scale, and trajectory while achieving stroke interpolation and editing.
Structure Detection for Contextual Reinforcement Learning
Tianyue Zhou (Massachusetts Institute of Technology), Cathy Wu (Massachusetts Institute of Technology)
CodeReinforcement LearningBenchmark
π― What it does: Studied the SD-MBTL framework for multi-strategy transfer learning based on structural detection in multi-dimensional context reinforcement learning, and proposed M/GP-MBTL to automatically identify and leverage the MOUNTAIN structure to select source tasks, significantly improving transfer efficiency.
Structure-based RNA Design by Step-wise Optimization of Latent Diffusion Model
Qi Si (Shanghai Academy of Artificial Intelligence for Science), Yuan Cheng (Shanghai Academy of Artificial Intelligence for Science)
CodeGenerationOptimizationGraph Neural NetworkTransformerSupervised Fine-TuningReinforcement LearningDiffusion modelGraphBiomedical Data
π― What it does: Proposed a framework named SOLD that combines implicit diffusion models with reinforcement learning for RNA inverse folding, directly optimizing 2D and 3D structural metrics through one-step sampling and segmented rewards, thereby improving the accuracy of RNA sequence design.
π― What it does: Proposed a training-free diffusion-based arbitrary style transfer method called StyleFM, combining three-frequency operations in the frequency domain and recursive attention to achieve high-quality content preservation and style embedding.
StyleTailor: Towards Personalized Fashion Styling via Hierarchical Negative Feedback
Hongbo Ma (Tsinghua University), Ming Li (Hangzhou Dianzi University)
CodeGenerationRetrievalRecommendation SystemVision Language ModelImageTextMultimodality
π― What it does: Propose the StyleTailor framework to realize a closed-loop system for personalized clothing design, shopping recommendations, virtual try-on, and evaluation.
SubGCache: Accelerating Graph-based RAG with Subgraph-level KV Cache
Qiuyu Zhu (Nanyang Technological University), Jie Zhang (Nanyang Technological University)
CodeRetrievalComputational EfficiencyGraph Neural NetworkLarge Language ModelGraphRetrieval-Augmented Generation
π― What it does: Designed and implemented SubGCache, a graph structure retrieval-augmented generation framework based on subgraph-level KV caching, significantly reducing inference latency in batch query scenarios.
Supervised Dynamic Dimension Reduction with Deep Neural Network
Zhanye Luo (University of Chicago), Xiufan Yu (University of Notre Dame)
CodeRepresentation LearningConvolutional Neural NetworkRecurrent Neural NetworkTime Series
π― What it does: Propose a supervised deep dynamic principal component analysis (SDDP) framework that integrates the target variable and lagged observations into factor extraction, generating target-aware low-dimensional features for nonlinear dynamic prediction.
Surgical AI Copilot: Energy-Based Fourier Gradient Low-Rank Adaptation for Surgical LLM Agent Reasoning and Planning
Jiayuan Huang (University College London), Mobarak I. Hoque (University of Manchester)
CodeTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringTextMultimodalityBiomedical Data
π― What it does: Propose Surgical AI Copilotβa Planner-Worker architecture-based LLM robot designed to provide real-time task planning, dialogue, and multimodal execution for endonasal pituitary surgery;
SurgPub-Video: A Comprehensive Surgical Video Framework for Enhanced Surgical Intelligence in Vision-Language Model
Yaoqian Li (Chinese University of Hong Kong), Pheng-Ann Heng (Chinese University of Hong Kong)
CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextBiomedical DataBenchmark
π― What it does: Constructed the SurgPub-Video dataset, a large-scale collection of peer-reviewed journal videos, and designed the SurgLLaVA-Video model based on TinyLLaVA-Video to support video-level inputs, aiming to enhance the performance of surgical vision-language models in tasks such as video scene understanding, VQA, action recognition, and technical evaluation.
π― What it does: Propose MetaDist, a framework that reinterprets graph structure poisoning attacks as self-supervised knowledge distillation, inducing misclassification by maximizing prediction distribution differences through teacher-student model interactions.
SVD-NO: Learning PDE Solution Operators with SVD Integral Kernels
Noam Koren (Technion Israel Institute of Technology), Daniel Freedman (Tel Aviv University)
CodeComputational EfficiencyPhysics Related
π― What it does: This paper proposes a new neural operator, SVD-NO, which directly parameterizes and learns the integral kernel using singular value decomposition (SVD), thereby efficiently approximating the solver of PDEs.
Syllogism-Inspired TableQA: Evidentialization Makes Decomposition Reasoning and Answer Verification More Reliable
Zhe Zhang (Northeastern University), Jie Song (Northeastern University)
CodeTransformerLarge Language ModelPrompt EngineeringTabularBenchmarkChain-of-Thought
π― What it does: This paper proposes a syllogism-based TableQA method called SIRV, which can utilize factual evidence from tables to enhance the reliability of reasoning and answer verification in LLMs.
π― What it does: This paper proposes a synthetic prediction method called SACP based on symmetric aggregation of non-consistency scores, which can effectively aggregate prediction uncertainties from multiple models while maintaining precise coverage.
π― What it does: Propose Symmetrical Flow Matching (SymmFlow), a unified flow matching framework that can accomplish three tasksβsemantic segmentation, classification, and image generationβin the same model.
Andrew Cropper (ELLIS Institute Finland), Matti JΓ€rvisalo
CodeComputational EfficiencyBenchmark
π― What it does: Proposed and implemented a method in inductive logic programming (ILP) that breaks rule space symmetries by constraining safe variables.
π― What it does: Studied and proposed a symmetry-aware Transformer training method to improve plan generation and heuristic prediction in automated planning tasks.
SyncBrain: Exploring Brain Functional Dynamics Through Neural Oscillatory Synchronization
Jiaqi Ding (University of North Carolina at Chapel Hill), Guorong Wu (University of North Carolina at Chapel Hill)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraphBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
π― What it does: Proposed SyncBrain, a physics-informed deep model based on Kuramoto coupled oscillators, to simulate brain function dynamics and decode brain states.
Synthetic Forgetting Without Access: A Few-Shot Zero-Glance Framework for Machine Unlearning
Qipeng Song (Royal Melbourne Institute Of Technology), Feng Xia (Royal Melbourne Institute Of Technology)
CodeClassificationSafty and PrivacySupervised Fine-TuningGenerative Adversarial NetworkImage
π― What it does: Designed and implemented the GFOES framework to achieve machine unlearning under the premise of having only a small amount of retained data and being unable to access forgotten data.
π― What it does: This paper constructs the first unified multi-region, multi-variable synthetic weather observation dataset, SynWeather, and proposes a text-prompt-based general diffusion transformer, SynWeatherDiff, for generating various weather variables;
T-Retriever: Tree-based Hierarchical Retrieval Augmented Generation for Textual Graphs
Chunyu Wei (Renmin University of China), Yueguo Chen (Renmin University of China)
CodeRetrievalGraph Neural NetworkTransformerLarge Language ModelGraphRetrieval-Augmented Generation
π― What it does: Propose the T-Retriever framework, transforming attribute graph retrieval into tree structure retrieval and constructing an information-theoretically optimal encoding tree.
π― What it does: Embed the Sampling Kaczmarz-Motzkin (SKM) method into a neural network to construct a trainable linear constraint satisfaction network, T-SKM-Net, achieving input-dependent dynamic linear constraint hard constraint satisfaction.
π― What it does: Propose a multi-source misinformation detection agent T2 Agent based on tool expansion, which identifies hybrid forgeries through dynamic programming and multi-modal verification.
π― What it does: Propose a temporal graph classification framework named T3FORMER, which extracts temporal information via a sliding window, combines topological (persistent homology) and spectral (Laplacian DOS) descriptors, and employs GraphSAGE to encode global structure; subsequently, it fuses multi-modal features through Transformer and Descriptor-Attention, ultimately achieving temporal graph classification.
π― What it does: Proposes the T4NMTD framework, achieving task decomposition and parallel training through DFA conversion based on LTLf, addressing sparse rewards and long-term dependencies in non-Markov tasks.
π― What it does: Proposed the Tab-PET framework, which generates position encodings for Tabular Transformers using graph-based spectral methods, thereby enhancing the model's representation capability for tabular data.
TabFlash: Efficient Table Understanding with Progressive Question Conditioning and Token Focusing
Jongha Kim, Hyunwoo J. Kim (Google Cloud AI)
CodeRecognitionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodalityTabularBenchmark
π― What it does: Designed and implemented an efficient table image understanding model called TabFlash, combining progressive question conditioning, background token pruning, and token focusing training strategies.
π― What it does: Designed and implemented a personalized federated learning method called FedCSPACK based on parameter package levels, which selects important parameter packages using cosine similarity and aggregates them through a dual-weight mask, significantly reducing communication overhead and enhancing model robustness.
TAdaRAG: Task Adaptive Retrieval-Augmented Generation via On-the-Fly Knowledge Graph Construction
Jie Zhang (Fuzhou University), Yanchao Tan (MemTensor (Shanghai) Technology Co., Ltd.)
CodeGenerationRetrievalTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextGraphBenchmarkRetrieval-Augmented Generation
π― What it does: Propose the TAdaRAG framework, which utilizes intent-driven extraction templates and two-phase training (supervised fine-tuning + reinforcement learning) to instantaneously construct task-adapted knowledge graphs during inference, thereby improving the generation quality of RAG.
Talk2Code: A Multi-Turn Interaction Benchmark with Dual-Track Evaluation for Code Generation
Weibin Yang (Xidian University), Hao Wang (Xidian University)
CodeAI Code AssistantLarge Language ModelTextBenchmark
π― What it does: This paper proposes the Talk2Code benchmark, which constructs a code generation task involving six types of users (two each for beginners, intermediates, and experts) with multi-round interactions, aiming to evaluate the ability of LLMs to generate code based on user proficiency levels in multi-round dialogues.
Taming Cascaded Mixture-of-Experts for Modality-missing Multi-modal Salient Object Detection
Kunpeng Wang (Anhui University), Keke Chen (Anhui University)
CodeObject DetectionTransformerMixture of ExpertsMultimodality
π― What it does: Propose a Cascaded Mixture-of-Experts (CMoE) framework for simultaneously handling missing and complete input scenarios in multi-modal salient object detection, which includes Missing-aware MoE (MaMoE) and Multi-modal MoE (MmMoE).
π― What it does: Propose Target-Balanced Score Distillation (TBSD), dynamically balancing shape and texture optimization within Score Distillation Sampling (SDS), addressing the trade-off between shape distortion and suboptimal texture caused by negative prompts.
π― What it does: Proposes the Task-aware Meta-learning on Heterogeneous Knowledge Graph (TMHKG) framework, treating POI recommendation as a meta-learning task for each user, and dynamically capturing user interest evolution by fusing geographic proximity and category transfer relationships through dual-view knowledge graphs and interaction graphs.
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: To address memory and communication bottlenecks in training large models with long contexts, we propose TawPipe, a topology-based weight pipeline parallel framework.
TechCoach: Towards Technical-Point-Aware Descriptive Action Coaching
Yuan-Ming Li (Sun Yat-sen University), Weishi Zheng
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelVision Language ModelVideoText
π― What it does: Proposes the Descriptive Action Coaching task, constructs the EE4D-DescCoach dataset and the TechCoach framework, which can generate detailed 'well done/needs improvement' comments and overall scores based on action videos.
Tell as You Want: Customizing Image Narrative with Knowledge and Thoughts
Ziwei Yao (Chinese Academy of Sciences), Xilin Chen (Chinese Academy of Sciences)
CodeGenerationRetrievalExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposed a novel image narrative generation framework called T4, capable of generating knowledge-rich and interpretable image narratives based on user interests;
TEMPLE: Incentivizing Temporal Understanding of Video Large Language Models via Progressive Pre-SFT Alignment
Shicheng Li (Peking University), Xu Sun (Kuaishou Technology)
CodeOptimizationTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVideoTextMultimodality
π― What it does: Propose the TEMPLE framework, leveraging direct preference optimization (DPO) and Pre-SFT alignment to significantly enhance the temporal understanding capabilities of video LLMs.
Temporal Dynamics Enhancer for Directly Trained Spiking Object Detectors
Fan Luo (Institute of Automation, Chinese Academy of Sciences), Yanfeng Lu (Institute of Automation, Chinese Academy of Sciences)
CodeObject DetectionSpiking Neural NetworkImageTime Series
π― What it does: Proposed the Temporal Dynamics Enhancer (TDE) module to enhance the temporal information modeling capability of directly trained spiking neural networks (SNNs) in object detection.
π― What it does: Propose a super-resolution video quality assessment framework based on time inconsistency guidance, quantifying and utilizing inter-frame inconsistency information to enhance SR video quality prediction.
Temporal Object-Aware Vision Transformer for Few-Shot Video Object Detection
Yogesh Kumar (Indian Institute of Technology Jodhpur), Anand Mishra (Indian Institute of Technology Jodhpur)
CodeObject DetectionTransformerVision Language ModelVideo
π― What it does: This paper proposes a few-shot video object detection framework based on language-aligned vision transformers and object-aware temporal fusion mechanisms, which can achieve high-precision detection of novel targets in videos with only a few support images provided.
TermGPT: Multi-Level Contrastive Fine-Tuning for Terminology Adaptation in Legal and Financial Domains
Yidan Sun (Zhejiang University), Shenglin Ben (Zhejiang University)
CodeDomain AdaptationRepresentation LearningTransformerSupervised Fine-TuningContrastive LearningTextFinance Related
π― What it does: Propose the TermGPT framework, achieving term adaptation in legal and financial domains through multi-layer contrastive learning, addressing term ambiguity caused by isotropy in LLM embedding spaces.
π― What it does: Proposes a test-driven reinforcement learning (TdRL) framework, using a set of pass-fail and indicative test functions evaluated through trajectories to replace traditional scalar reward functions, defining task objectives while providing learning guidance; implements a complete algorithm under maximum entropy RL (SAC/PPO);
Test-time Diverse Reasoning by Riemannian Activation Steering
Ly Tran Ho Khanh (Chinese University of Hong Kong), Viet Anh Nguyen (University of Hong Kong)
CodeOptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelTextBenchmark
π― What it does: Proposed an unsupervised activation regulation method for language models during inference (SPREAD) to enhance the diversity and accuracy of best-N sampling.
Test-Time Reinforcement Learning for GUI Grounding via Region Consistency
Yong Du (Zhejiang University), Yongliang Shen (SF Technology)
CodeObject DetectionReinforcement LearningVision Language ModelImage
π― What it does: Proposes two unsupervised test-time improvement methodsβGUI-RC (spatial voting based on multiple sampling) and GUI-RCPO (test-time reinforcement learning with region consistency rewards)βwhich can enhance GUI localization accuracy without increasing labeled data or retraining.