π― What it does: Propose the SDE-HARL framework, which decomposes the policy network of heterogeneous multi-agents into local lightweight networks and edge global networks, achieving distributed execution through compressed hidden layers;
SDEval: Safety Dynamic Evaluation for Multimodal Large Language Models
Hanqing Wang (Shanghai Artificial Intelligence Laboratory), Xiangyang Zhu (ShanghaiTech University)
CodeSafty and PrivacyExplainability and InterpretabilityLarge Language ModelPrompt EngineeringVision Language ModelDiffusion modelImageTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Proposed SDEval, a security dynamic evaluation framework for multimodal large language models;
π― What it does: For the LiDAR semantic scene completion task, this paper proposes the SDNet framework and introduces an input-aware label correction strategy.
SEA-PACE: Semi-Supervised Underwater Image Enhancement via Gaussian ProcessβAssisted Self-Paced Learning
Jingyang Wang (Ocean University of China), Junyu Dong (Ocean University of China)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: Proposes the SEA-PACE framework, which utilizes semi-supervised learning combined with model-aware reliability assessment and adaptive consistency learning to enhance underwater image enhancement.
SEAP: Sparse Expert Activation Pruning Unlocks the Brainpower of Large Language Models
Xun Liang (Renmin University of China), Zhiyu Li (Institute for Advanced Algorithms Research)
CodeComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
π― What it does: Proposes a training-free, task-adaptive sparse expert activation pruning (SEAP) method, which constructs task-specific pruning masks by analyzing hidden state clustering in multi-task calibration data, and dynamically selects corresponding computational paths during inference using a lightweight router, significantly reducing inference cost while maintaining most of the performance.
SEBSFormer: A Spectral-Enhanced Bi-Stream Transformer for Robust EEG Decoding
Lin Zhang (Shanghai Jiao Tong University), Lei Xu (Shanghai Jiao Tong University)
CodeClassificationRecognitionTransformerBiomedical Data
π― What it does: This paper proposes a Spectral Enhanced Dual-Stream Transformer (SEBSFormer), which enhances the spectral diversity and spatiotemporal representation capability of EEG signals through three mechanisms: a spectral compensation module, a multi-scale temporal attention module, and graph-guided dynamic fusion. Additionally, the paper formally proves the low-pass filtering characteristics of the Transformer's self-attention mechanism.
Security Games with Layered Defenses: Adaptive Adversaries and Gittins Indices
Chun Kai Ling (National University of Singapore), Christian Kroer (Columbia University)
CodeOptimizationReinforcement LearningGraph
π― What it does: Proposes a security game model for hierarchical defense, studying the scenario where adaptive attackers dynamically switch between multiple attack paths, and provides corresponding optimal defense strategies;
SEED: Spectral Entropy-Guided Evaluation of Spatial-Temporal Dependencies for Multivariate Time Series Forecasting
Feng Xiong, Jianhong Lin (Tianjin University)
CodeGraph Neural NetworkTransformerTime Series
π― What it does: This paper proposes the SEED framework, which uses spectral entropy to evaluate spatiotemporal dependencies in multivariate time series, dynamically balancing the independence and interdependence of variables, and achieving more accurate predictions through modules such as signed graph construction and context space extraction.
Seeing and Knowing in the Wild: Open-domain Visual Entity Recognition with Large-scale Knowledge Graphs via Contrastive Learning
Hongkuan Zhou (Robert Bosch GmbH), Steffen Staab (University of Stuttgart)
CodeRecognitionTransformerVision Language ModelContrastive LearningImageTextGraph
π― What it does: Proposes the Knowledge-guided Contrastive Learning (KnowCoL) framework for open-domain visual entity recognition, mapping images, text descriptions, and structured information from the Wikidata knowledge graph into a shared semantic space to achieve zero-shot entity recognition.
CodeSegmentationAnomaly DetectionExplainability and InterpretabilityVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Proposes VBackChecker, a reference-free multimodal large language model hallucination detection framework that verifies through pixel-level visual reverse attribution.
π― What it does: This paper proposes a joint de-raining and super-resolution framework called DHGM, which combines diffusion models with high-pass filtering and guided filtering to achieve high-frequency texture restoration and magnification of rainy images.
SEFEL: A Simple Yet Effective Framework for Fast Event Linking
Yinan Liu (Northeastern University), Xiaochun Yang (Northeastern University)
CodeRetrievalComputational EfficiencyRepresentation LearningTransformerLarge Language ModelContrastive LearningText
π― What it does: Proposed an end-to-end event linking framework called SEFEL, which constructs argument-aware event representations by leveraging event argument information and uniformly handles event linking both inside and outside knowledge bases.
π― What it does: Proposed SELDON, a continuous-time variational autoencoder for real-time prediction and physical parameter inference in sparse, heteroscedastic, multi-band light curves.
Selection of LLM Fine-Tuning Data Based on Orthogonal Rules
Xiaomin Li (Harvard University), Hong Hu (Massachusetts Institute of Technology)
CodeData-Centric LearningLarge Language ModelText
π― What it does: This paper proposes a fully automated data selection framework that leverages LLM to automatically generate rules and employs DPP (Determinantal Point Process) for rule orthogonality screening, aiming to select high-quality subsets from large-scale corpora to enhance LLM fine-tuning performance.
Mohit Meena (Fujitsu Research of India), Mahesh Chandran (Fujitsu Research of India)
CodeClassificationComputational EfficiencyRepresentation LearningGraph Neural NetworkMixture of ExpertsGraph
π― What it does: Propose the Self-Adaptive Graph Mixture of Models (SAGMM) framework, which utilizes multiple GNNs as experts and adaptively selects and combines them through topology-aware attention gates to enhance performance in graph learning tasks.
π― What it does: Propose Self-NPO, achieving diffusion model alignment without labels or reward models through self-supervised negative preference optimization (NPO), and significantly reduce training costs via truncated diffusion fine-tuning (TDFT).
CodeExplainability and InterpretabilityRepresentation LearningText
π― What it does: Propose a self-supervised ILP framework named SS-ILP, and design the Poker algorithm, which can automatically generate positive and negative samples and learn recursive logic programs by maximizing the second-order background theory (SONF) under conditions with only positive examples and unlabeled samples.
π― What it does: Proposed a self-supervised one-step diffusion refinement framework (OSD), which efficiently reconstructs multispectral images through an initial predictor and a single-step diffusion network.
π― What it does: Propose AD-L-JEPA, a self-supervised pre-training framework that learns LiDAR object detection features by predicting masked region embeddings in the BEV space.
Semantic Document Derendering: SVG Reconstruction via Vision-Language Modeling
Adam Hazimeh (EPFL), Pascal Frossard (EPFL)
CodeImage TranslationGenerationTransformerVision Language ModelImageMultimodalityBenchmark
π― What it does: Propose a framework called SliDer based on Vision-Language Models (VLM) to restore color slide images into editable SVG format, achieving a complete conversion from pixels to structured vectors.
Semantic Volume: Quantifying and Detecting Both External and Internal Uncertainty in LLMs
Xiaomin Li (Harvard University), Anurag Beniwal (Amazon)
CodeExplainability and InterpretabilityLarge Language ModelTextBenchmark
π― What it does: Propose the Semantic Volume method, which quantifies the internal and external uncertainty of LLMs by calculating the determinant of the Gram matrix on perturbed embeddings of queries or answers.
Semantic-Aware Feature Enhancement for Partial Label Learning
Haowei Mei (Nanjing University), Huaxiong Li (Nanjing University)
CodeClassificationImageTabular
π― What it does: Proposes the Semantic-Aware Feature Enhancement (SAFE) method for addressing the partially labeled learning (PLL) problem, shifting the focus from label denoising to feature recovery and enhancement, and constructing a classifier that jointly integrates complete features with candidate labels.
π― What it does: Proposed a cross-modal hashing framework SCBCH for multi-label noise scenarios, combining CSCC and BSCH modules to achieve robust retrieval.
Semantic-Driven Visual Progressive Refinement for Aerial-Ground Person ReID: A Challenging Large-Scale Benchmark
Aihua Zheng (Anhui University), Bin Luo (Anhui University)
CodeRecognitionRetrievalTransformerPrompt EngineeringMixture of ExpertsVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Propose a semantic-driven visual evolutionary refinement framework, SVPR-ReID, for aerial-ground person re-identification (AGPReID), and contribute a large-scale real-world scenario dataset, CP2108.
CodeClassificationObject DetectionCompressionExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkAuto EncoderImage
π― What it does: Designed and implemented a semantic encoder-decoder framework called SemanticNN for extremely weak embedded devices, capable of withstanding bit errors in wireless transmission, achieving feature compression, and enabling device-edge collaborative inference.
π― What it does: SemanticVLA proposes a semantic alignment sparsification and enhancement framework for efficiently performing robotic grasping and manipulation tasks.
SEMC: Structure-Enhanced Mixture-of-Experts Contrastive Learning for Ultrasound Standard Plane Recognition
Qing Cai (Ocean University of China), Zhi Liu (Shandong University)
CodeClassificationRecognitionConvolutional Neural NetworkMixture of ExpertsContrastive LearningImageBiomedical DataUltrasound
π― What it does: For the task of ultrasound standard plane recognition, the SEMC framework is proposed, combining structure-aware feature fusion (SSFM) and expert-driven hierarchical contrastive learning with classification (MCRM)
Semi-Supervised High Dynamic Range Image Reconstructing via Bi-Level Uncertain Area Masking
Wei Jiang (Huazhong University of Science and Technology), Zhiguo Cao (Huazhong University of Science and Technology)
CodeRestorationImage
π― What it does: This paper proposes a semi-supervised HDR image reconstruction framework that utilizes a teacher-student model combined with an uncertainty mask to achieve high-quality HDR restoration with only a small amount of HDR labeled data.
Semi-Supervised Regression by Preserving Ranking Relationships Between Close Unlabeled Samples
Ximing Li (Jilin University), Renchu Guan (Shenzhen University of Advanced Technology)
CodeImageTextAudio
π― What it does: Propose a new semi-supervised regression method SSR-RCUS, which combines samples generated by closed Mixup with an auxiliary ranking task to enhance regression performance.
Semi-Supervised Semantic Segmentation via Derivative Label Propagation
Yuanbin Fu (Tianjin University), Xiaojie Guo (Tianjin University)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: This paper proposes a semi-supervised semantic segmentation framework called DerProp, which corrects pseudo-labels by performing discrete derivative operations on pixel features and incorporating similarity regularization.
Semore: VLM-guided Enhanced Semantic Motion Representations for Visual Reinforcement Learning
Wentao Wang, Yan Wang (Tsinghua University)
CodeLarge Language ModelReinforcement LearningVision Language ModelImage
π― What it does: Propose the Semore framework, which employs a VLM-guided semantic and motion dual-stream encoder to explicitly supervise and interact features in visual reinforcement learning, thereby enhancing environmental representation capabilities.
Sentient: Detecting APTs via Capturing Indirect Dependencies and Behavioral Logic
Wenhao Yan (Institute of Information Engineering, Chinese Academy of Sciences), Junrong Liu (Institute of Information Engineering, Chinese Academy of Sciences)
π― What it does: Leverage pre-trained graph Transformer and intent analysis module (Bi-Mamba2) to perform global attention learning on the original graph constructed from system audit logs, capturing indirect dependencies and behavioral logic, and detecting APT attacks using only normal logs.
SepPrune: Structured Pruning for Efficient Deep Speech Separation
Yuqi Li (City College of New York), Yao Lu (University of Southern California)
CodeCompressionComputational EfficiencyAudio
π― What it does: Developed a structured pruning framework called SepPrune for compressing deep speech separation models and reducing computational costs.
CodeDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelContrastive LearningMultimodalityBiomedical Data
π― What it does: This paper proposes BioText-CPI, a framework for predicting compound-protein interactions based on protein biomedical text descriptions without requiring protein sequences, supporting multi-modal input.
SERL: Self-Examining Reinforcement Learning on Open-Domain
Weixuan Ou (Zhejiang University), Yifan Qiao (Alibaba Group)
CodeLarge Language ModelReinforcement LearningText
π― What it does: Proposed an unsupervised self-checking reinforcement learning framework called SERL, where large language models simultaneously act as both generator (Actor) and evaluator (Judge), achieving self-improvement through internal comparative evaluation.
SEVADE: Self-Evolving Multi-Agent Analysis with Decoupled Evaluation for Hallucination-Resistant Sarcasm Detection
Ziqi Liu (Xi'an Jiaotong-Liverpool University), Yangbin Chen (Xi'an Jiaotong-Liverpool University)
CodeClassificationLarge Language ModelAgentic AITextRetrieval-Augmented Generation
π― What it does: Proposed a self-evolving framework named SEVADE based on dynamic multi-agent reasoning and decoupled evaluation for detecting sarcastic text.
SeViL: Semi-supervised Vision-Language Learning with Text Prompt Guiding for Moving Infrared Small Target Detection
Weiwei Duan (University of Electronic Science and Technology of China), Sicheng Zhu (University of Electronic Science and Technology of China)
CodeObject DetectionPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
π― What it does: Proposes a semi-supervised vision-language learning framework SeViL, which utilizes text prompts to guide the detection of small moving targets in infrared images.
SFedHIFI: Fire Rate-Based Heterogeneous Information Fusion for Spiking Federated Learning
Ran Tao (Southwestern University of Finance and Economics), Guisong Liu (Southwestern University of Finance and Economics)
CodeFederated LearningSpiking Neural NetworkImage
π― What it does: Proposes the SFedHIFI framework, enabling edge devices to adaptively deploy spiking neural networks (SNNs) with varying widths based on local resources, and aggregates heterogeneous models through channel-level matrix decomposition;
π― What it does: Designed an unsupervised multi-modal graph anomaly detection framework SFGA, which identifies abnormal nodes by combining multi-view graph autoencoders with cross-modal attention fusion and similarity constraints.
π― What it does: Propose a Singularity-enhanced Graph Attention Network (SGAT), which achieves high-precision image feature matching in low-texture and heavily occluded scenarios through a graph attention mechanism guided by single-point singularity and co-potential priors.
π― What it does: Propose the SGP4SR model, which adopts a separated modal framework to learn user preferences, significantly reducing noise impact in multi-modal recommendations.
π― What it does: Propose a multi-modal image matching method called SGPFeat, which leverages semantic and geometric priors to enhance feature extraction and keypoint detection;
π― What it does: Propose SGS-3D, a training-agnostic framework based on the 'split-then-grow' idea, which first uses geometric primitives to separate and clean 2D-uplifted semantic masks, then gradually grows them into complete instances in 3D voxels based on spatial connectivity and feature similarity, achieving high-fidelity 3D instance segmentation.
ShapBPT: Image Feature Attributions Using Data-Aware Binary Partition Trees
Muhammad Rashid (University of Torino), Damiano Verda (Rulex Innovation Labs)
CodeClassificationObject DetectionAnomaly DetectionExplainability and InterpretabilityImage
π― What it does: Proposes ShapBPT, a data-aware hierarchical Shapley attribution method based on binary partition trees (BPT), for pixel-level image explanations.
Shaping Parameter Contribution Patterns for Out-of-Distribution Detection
Haonan Xu (Nanjing University of Science and Technology), Yang Yang (Nanjing University of Science and Technology)
CodeClassificationAnomaly DetectionImage
π― What it does: This paper proposes a regularization method called SPCP, which truncates the contribution of classifier parameters during training. By limiting the upper bound of parameter contributions, it forms a dense and bounded contribution pattern, thereby enhancing the model's robustness to abnormal inputs and reducing over-reliance on a few dominant parameters.
Sheaf Graph Neural Networks via PAC-Bayes Spectral Optimization
Yoonhyuk Choi (Sookmyung Women's University), Chong-Kwon Kim (Korea Institute of Energy Technology)
CodeClassificationGraph Neural NetworkGraph
π― What it does: Designed and implemented the SGPC (Sheaf Graph Neural Networks with PAC-Bayes Calibration) model, integrating cellular sheaf propagation, Wasserstein-Entropic lifting, Stochastic Variance-Reduced Diffusion (SVR), and Adaptive Frequency Mixing (AFM), and achieving end-to-end training through Ξ²-Dirichlet posterior calibration.
SheetBrain: A Neuro-Symbolic Agent for Accurate Reasoning over Complex and Large Spreadsheets
Ziwei Wang (Carnegie Mellon University), Dongmei Zhang (Microsoft Research)
CodeTransformerLarge Language ModelAgentic AITabularBenchmark
π― What it does: Proposes SheetBrain, a multi-stage (understandingβexecutionβverification) framework based on neuro-symbolic reasoning, to accurately answer and operate complex large-scale spreadsheets.
ShoppingBench: A Real-World Intent-Grounded Shopping Benchmark for LLM-based Agents
Jiangyuan Wang (Alibaba International Digital Commercial Group), Xiaoyi Zeng (Alibaba International Digital Commercial Group)
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmarkRetrieval-Augmented Generation
π― What it does: Proposed ShoppingBench, a real-intent-driven end-to-end benchmark for shopping scenarios, containing 3,310 user instructions and a sandbox environment with 25 million real products;
CodeRetrievalRepresentation LearningVision Language ModelContrastive LearningImageBiomedical Data
π― What it does: Propose the Adaptive Teaching Paradigm, utilizing a visual model as a teacher to dynamically shrink and adapt its knowledge to fit EEG students, achieving zero-shot EEG-to-Image retrieval;
CodeDepth EstimationLarge Language ModelReinforcement LearningVision Language ModelImageTextChain-of-Thought
π― What it does: Proposed and implemented the SIFThinker framework, which can dynamically focus on and correct visual attention regions during the reasoning process of multimodal large models, and integrate depth information into the reasoning chain.
π― What it does: Proposes the Sign-Aware Multimodal Graph Recommendation (SiMGR) framework, which integrates multimodal features with a signed user-item interaction graph and uses positive and negative thresholds to separate interactions and enhance with pseudo edges.
π― What it does: Propose the Signal framework, which first uses SIM to select important patches from multi-modal images, then achieves global multi-modal alignment in the Gramian space via GAM, and finally performs pixel-level local alignment using the learnable offsets of LAM, thereby enhancing feature representation in multi-modal ReID.
π― What it does: Proposed a two-step screen-camera (SC) watermark noise approximation framework named Simulation-to-Real (S2R), which first generates rough noise through mathematical modeling and then refines the approximation of real SC noise using an unsupervised image-to-image network, thereby enhancing the robustness of watermarks in real-world scenarios.
Sim4Seg: Boosting Multimodal Multi-disease Medical Diagnosis Segmentation with Region-Aware Vision-Language Similarity Masks
Lingran Song (University of Macau), Jianbing Shen (University of Macau)
CodeSegmentationExplainability and InterpretabilityTransformerLarge Language ModelMultimodalityBiomedical DataUltrasoundBenchmarkChain-of-Thought
π― What it does: This paper proposes the medical diagnosis segmentation (MDS) task and constructs a multimodal, multi-disease medical image segmentation and diagnostic dataset, M3DS. It introduces the Sim4Seg framework and the Region-Aware Vision-Language Similarity to Mask (RVLS2M) module, achieving segmentation and diagnosis of multimodal medical images while providing interpretable diagnostic chains.
π― What it does: Propose the Simba framework, transforming the point cloud completion task into learning the distribution of geometric transformations to achieve high-fidelity geometric consistency.
SimDiff: Simpler Yet Better Diffusion Model for Time Series Point Forecasting
Hang Ding (Shanghai Jiao Tong University), Tao Yao (Shanghai Jiao Tong University)
CodeTransformerDiffusion modelTime Series
π― What it does: Proposed SimDiff, a single-stage end-to-end diffusion model for time series point prediction, completely eliminating pre-trained regressors;
SimLabel: Similarity-Weighted Semi-supervision for Multi-annotator Learning with Missing Labels
Liyun Zhang, Yuta Nakashima (Cincinnati Children's Hospital Medical Center)
CodeClassificationRecognitionVideoMultimodality
π― What it does: Propose a similarity-weighted semi-supervised framework called SimLabel to address the problem of missing labels in multi-annotator learning;
Simulating Dispute Mediation with LLM-Based Agents for Legal Research
Junjie Chen (Tsinghua University), Qingyao Ai (Tsinghua University)
CodeTransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
π― What it does: Developed the AgentMediation framework, utilizing LLM agents to simulate real civil dispute mediation processes, generating multi-round dialogues and solutions.
Simulating Human-Like Counseling: A Path- and Scenario-Guided Framework for Psychological Support Dialogue
Yuanchen Shi (Soochow University), Fang Kong (Jiangsu University)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningAgentic AIPrompt EngineeringText
π― What it does: This paper proposes a path-guided simulation framework, PGSim, to construct a large-scale, structured, and context-rich Chinese psychological support dialogue dataset, CPsDD, and develops a multi-modal intelligent agent system, CADSS, based on this dataset, capable of generating context-controllable psychological counseling dialogues.
CodeOptimizationTransformerReinforcement LearningTabularTime Series
π― What it does: Convert train delay prediction into a stochastic simulation task, training strategies through imitation learning to replicate historical state transitions, thereby achieving temporal prediction of delays.
SinBasis Networks: Matrix-Equivalent Feature Extraction for Wave-Like Optical Spectrograms
Yuzhou Zhu (Dalian University of Technology), Liang Zhou (Dalian University of Technology)
CodeClassificationConvolutional Neural NetworkTransformerImagePhysics Related
π― What it does: Propose Sin-Basis Networks, which map the weights of CNN, ViT, Capsule, and other networks to sine basis functions, enhancing the extraction of periodic features in waveform images.
Cameron Gordon (Australian Institute for Machine Learning, University of Adelaide), Simon Lucey (Australian Institute for Machine Learning, University of Adelaide)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
π― What it does: Propose a SineLoRA β method that employs sine activation on quantized low-rank adapters (LoRA) to achieve differential compression;
SITA: A Framework for Structure-to-Instance Theorem Autoformalization
Chenyi Li (Peking University), Zaiwen Wen (Peking University)
CodeOptimizationAI Code AssistantLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Designed and implemented the SITA framework, achieving automated theorem proving in Lean, including the automatic generation of definitions, instance declarations, and complete proofs from abstract structures to concrete instances.
π― What it does: Proposed a new skeleton-based pre-training framework called CSIP-ReID for video person re-identification (ReID), achieving multi-modal pre-training through contrastive learning on skeleton sequences and image frames.
Skill Path: Unveiling Language Skills from Circuit Graphs
Hang Chen (Xi'an Jiaotong University), Wenya Wang (Nanyang Technological University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelText
π― What it does: Proposed a three-step framework (decomposition, pruning, causal mediation) to extract skill paths (Skill Path) in language models, enabling finer-grained capture of mechanisms underlying specific language skills;
SkillGen: Learning Domain Skills for In-Context Sequential Decision Making
Ruomeng Ding (University of North Carolina at Chapel Hill), Chen Zhao (Baylor University)
CodeLarge Language ModelReinforcement LearningPrompt EngineeringTextSequentialRetrieval-Augmented Generation
π― What it does: Propose the SkillGen framework, which constructs an action-centric domain knowledge graph using trajectories sampled from LLMs. High-value actions are extracted through TD assignment to form 'golden segments' and 'step-by-step skills.' During inference, these structured knowledge elements are combined with recent actions to generate more focused, fine-grained, and label-efficient prompts.
Sliding-Window Merging for Compacting Patch-Redundant Layers in LLMs
Xuan Ding (Chinese University of Hong Kong (Shenzhen)), Yao Zhu (Zhejiang University)
CodeOptimizationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposed and implemented the Sliding-Window Merging (SWM) method to dynamically merge adjacent functionally similar Transformer layers in large language models (LLMs), thereby compressing model depth without significantly compromising performance.
SlimInfer: Accelerating Long-Context LLM Inference via Dynamic Token Pruning
Lingkun Long (Beihang University), Jianlei Yang (Beihang University)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Achieve long-context LLM inference acceleration by dynamically performing fine-grained token trimming on hidden states during the forward propagation process.
Small but Mighty: Dynamic Wavelet Expert-Guided Fine-Tuning of Large-Scale Models for Optical Remote Sensing Object Segmentation
Yanguang Sun (Nanjing University of Science and Technology), Lei Luo (Nankai University)
CodeSegmentationTransformerSupervised Fine-TuningMixture of ExpertsImage
π― What it does: Propose a fine-tuning paradigm based on dynamic waveform expert guidance (WEFT), which significantly enhances the performance of optical remote sensing image (ORSI) target segmentation by freezing a large base model (UniPerceiver-L) and introducing a lightweight task-specific waveform expert extractor and expert-guided conditional adapter.
Small Models Exhibit Limited Answer Consistency in Repetition Trials of the Multiple-Choice MMLU-Redux and MedQA Benchmarks
Claudio Pinhanez (IBM Research Brazil), Marcelo Carpinette Grave (IBM Research Brazil)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical DataBenchmark
π― What it does: Conduct 10 repeated inferences of open-source LLMs (2Bβ80B) on MMLU-Redux and MedQA multiple-choice questions, evaluating answer consistency and correlating it with accuracy.
SMART: A Surrogate Model for Predicting Application Runtime in Dragonfly Systems
Xin Wang (University of Illinois Chicago), Zhiling Lan (University of Illinois Chicago)
CodeOptimizationComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelGraphTime Series
π― What it does: Proposes the SMART model, integrating graph neural networks (GNN) and large language models (LLM) to predict application iteration time in Longfly network systems, achieving high-accuracy and low-latency proxy simulation.
SmartAgent: Chain-of-User-Thought for Embodied Personalized Agent in Cyber World
Jiaqi Zhang (University of Queensland), Hongzhi Yin (University of Queensland)
CodeRecommendation SystemRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelVision-Language-Action ModelMultimodalityBenchmarkChain-of-Thought
π― What it does: Propose the Chain-of-User-Thought (COUT) embodied personalized reasoning paradigm and implement the SmartAgent framework to achieve GUI operation to user preference reasoning and personalized recommendation in networked environments, while constructing the SmartSpot dataset.
Xu Liu (Xidian University), Licheng Jiao (Xidian University)
CodeObject DetectionTransformerVision Language ModelContrastive LearningImage
π― What it does: Proposes SOAR, a semi-supervised open-vocabulary aerial object detection framework, which utilizes dual-perception prior denoising to automatically generate pseudo labels and enhance queries, significantly improving detection performance.
π― What it does: Proposes SODiff, a first-order diffusion model that removes JPEG compression artifacts by leveraging a semantically aligned image prompt extractor and a quality factor-aware temporal predictor.
SOM Directions Are Better than One: Multi-Directional Refusal Suppression in Language Models
Giorgio Piras (University of Cagliari), Battista Biggio (University of Cagliari)
CodeSafty and PrivacyHyperparameter SearchLarge Language ModelTextBenchmark
π― What it does: This paper studies the refusal behavior in large language models (LLMs) and proposes a method that utilizes self-organizing maps (SOM) to extract multidimensional refusal directions from internal representations, reducing the model's refusal rate by suppressing these directions.
π― What it does: Proposes the SOMA framework for precise pixel-level SAR and optical image registration, integrating the Feature Gradient Enhancer (FGE) and Global-Local Affine-Flow Matcher (GLAM).
SoMe: A Realistic Benchmark for LLM-based Social Media Agents
Dizhan Xue (Chinese Academy of Sciences), Changsheng Xu (Chinese Academy of Sciences)
CodeLarge Language ModelAgentic AITextBenchmark
π― What it does: Constructed the SoMe benchmark, evaluating the performance of LLM-driven social media agents on 8 tasks and providing 8 tool platforms.
Sonnet: Spectral Operator Neural Network for Multivariable Time Series Forecasting
Yuxuan Shu (University College London), Vasileios Lampos (University College London)
CodeTime SeriesBenchmark
π― What it does: Proposed the Sonnet model, which captures internal and external dependencies in multivariate time series in the spectral domain using learnable wavelet transforms and Koopman operators, and improves prediction accuracy through multivariate resonance attention (MVCA).
SOSControl: Enhancing Human Motion Generation Through Saliency-Aware Symbolic Orientation and Timing Control
Ho Yin Au (Hong Kong Baptist University), Jie Chen (Hong Kong Baptist University)
CodeGenerationDiffusion modelTextSequential
π― What it does: This paper proposes the Salient Orientation Symbolic (SOS) script and the SOSControl framework, achieving programmable human action generation based on salient orientation symbols. It automatically extracts sparse symbolic scripts from motion data and realizes precise pose and temporal control through ControlNet and iterative optimization in text-driven diffusion models.
SpaceVLLM: Endowing Multimodal Large Language Model with Spatio-Temporal Video Grounding Capability
Jiankang Wang (Renmin University of China), Yongdong Zhang (Renmin University of China)
CodeObject DetectionObject TrackingData SynthesisTransformerLarge Language ModelVideoTextMultimodality
π― What it does: Propose SpaceVLLM, a framework that enables MLLM to perform spatio-temporal video grounding (STVG) through Spatio-Temporal Aware Queries and a Query-Guided Space Head.
SpaCRD: Multimodal Deep Fusion of Histology and Spatial Transcriptomics for Cancer Region Detection
Shuailin Xue (Yunnan University), Wenwen Min (Zhongnan University of Economics and Law)
CodeSegmentationDomain AdaptationTransformerAuto EncoderImageMultimodalityBiomedical Data
π― What it does: Proposes SpaCRD, a framework leveraging multi-modal deep fusion and transfer learning for cross-sample, cross-platform, and cross-batch cancer tissue region (CTR) detection.
SPAN: Benchmarking and Improving Cross-Calendar Temporal Reasoning of Large Language Models
Zhongjian Miao (Li Auto Inc), Chen Wei (Li Auto Inc)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
π― What it does: This paper introduces the SPAN benchmark to evaluate the ability of large language models in cross-calendar time reasoning (covering ten reasoning directions, two reasoning types, and two question formats);
π― What it does: The SPARD framework achieves efficient inference for human motion prediction through a single-step residual diffusion model and an adaptive noise predictor;
SPARE: Single-Pass Annotation with Reference-Guided Evaluation for Automatic Process Supervision and Reward Modelling
Md Imbesat Hassan Rizvi (Technical University of Darmstadt), Iryna Gurevych (Queen's University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
π― What it does: Proposes the SPARE (Single-Pass Annotation with Reference-Guided Evaluation) framework, achieving one-time alignment and evaluation of multi-step reasoning processes generated by LLMs, enabling automated process supervision.
SparK: Query-Aware Unstructured Sparsity with Recoverable KV Cache Channel Pruning
Huanxuan Liao (Chinese Academy Of Sciences), Kang Liu (Chinese Academy Of Sciences)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose the SPARK method, which performs training-free, recoverable channel-level non-structured sparsification on LLM KV cache to reduce memory and computational costs.
Sparse Poisson Gamma Belief Networks for High-Dimensional Sparse Count Data
Rui Huang (Harbin Institute of Technology), Sikun Yang (Great Bay University)
CodeClassificationComputational EfficiencyRepresentation LearningTextBiomedical Data
π― What it does: This paper proposes a Sparse Poisson Gamma Belief Network (SPGBN), which enhances feature extraction and missing value prediction performance for high-dimensional sparse count data by explicitly controlling inter-layer connections through sparse graph structure priors and binary masks, building upon the traditional PGBN.
π― What it does: Proposes MIST (Mutual Information-guided Sparse Tuning), which first performs sparse fine-tuning on pre-trained models (PTM) to enhance plasticity and generalization ability in continual learning.
Sparse-dLLM: Accelerating Diffusion LLMs with Dynamic Cache Eviction
Yuerong Song (Fudan University), Xipeng Qiu (Fudan University)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: Propose Sparse-dLLM, which enhances the inference speed of Diffusion LLM through dynamic bidirectional cache eviction and sparse attention
Sparse-Scale Transformer with Bidirectional Awareness for Time Series Forecasting
Ying Liu (Hefei University of Technology), Richang Hong (Chongqing University)
CodeComputational EfficiencyTransformerTime Series
π― What it does: Propose a sparse scale Transformer (SSformer) that enhances time series forecasting performance by adaptively selecting effective time scales and combining bidirectional scale interactions.
SparseRM: A Lightweight Preference Modeling with Sparse Autoencoder
Dengcan Liu (University of Science and Technology of China), Yongdong Zhang (University of Science and Technology of China)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningAuto EncoderText
π― What it does: Propose SparseRM, a lightweight reward model based on sparse autoencoders (SAE), for modeling and aligning preferences of large language models (LLMs).