π― What it does: Propose VisionDrop, an untrained vision-only attention-based token sparsification framework that progressively prunes and fuses residual information in visual encoders and LLMs stage-by-stage.
RetouchAgent: Towards Interactive and Explainable Image Retouching with MLLM Agents
Shuo Zhang (Xi'an Jiaotong University), Xinyu Yang (Xi'an Jiaotong University)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringVision Language ModelImageTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose the RetouchAgent framework, which utilizes multimodal large language models to collaboratively achieve interactive and interpretable image retouching;
RetroLM: Retrieval-Augmented KVs for Long-Context Processing
Kun Luo (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences), Kang Liu (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences)
CodeRetrievalComputational EfficiencyTransformerLarge Language ModelContrastive LearningTextRetrieval-Augmented Generation
π― What it does: RetroLM achieves retrieval-augmented long-context reasoning by implementing retrieval enhancement at the KV cache level of LLMs, splitting long texts into pages and retrieving critical pages on demand.
CodeAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningTextBenchmarkChain-of-Thought
π― What it does: Proposes RetrySQL, a training paradigm that utilizes retry data for self-correction during the text-to-SQL generation process. After training, the model can identify and correct erroneous reasoning steps during generation, thereby producing more accurate SQL queries.
Revisiting (Un)Fairness in Recourse by Minimizing Worst-Case Social Burden
Ainhize Barrainkua (Basque Center for Applied Mathematics), Novi Quadrianto (Basque Center for Applied Mathematics)
CodeOptimizationExplainability and InterpretabilityTabular
π― What it does: Proposes a unified algorithmic fairness explanation framework and minimizes social burden via the MISOB method without using sensitive attributes;
π― What it does: This paper proposes a Norm-Ratio Attention and Semantic Recovery Distillation Network, addressing the modeling and recovery of missing semantic information in pedestrian re-identification under extremely low illumination conditions.
π― What it does: Propose the LightCSCF method, which suppresses structurally similar negative samples by using boundary-constrained cosine similarity in contrastive learning, addressing the GCN over-smoothing and contrastive learning gradient saturation issues, thereby improving recommendation effectiveness.
Revisiting Differentiable Structure Learning: Inconsistency of L1 Penalty and Beyond
Kaifeng Jin (University of Illinois Urbana-Champaign), Biwei Huang (University of California San Diego)
CodeOptimizationRepresentation LearningGraph
π― What it does: Investigated the inconsistency problem of L1 penalty in differentiable structure learning, and proposed the CALM method based on L0 penalty, hard DAG constraints, and moral graphs.
Revisiting Fairness-aware Interactive Recommendation: Item Lifecycle as a Control Knob
Yun Lu (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences), Mingsheng Shang (Chongqing Institute of Green and Intelligent Technology, Chinese Academy of Sciences)
π― What it does: This paper proposes using the lifecycle of short video content (rapid growth, stable period, decline period) as a control mechanism to improve fairness and accuracy in interactive recommendations;
Revisiting MLLM Based Image Quality Assessment: Errors and Remedy
Zhenchen Tang (New Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences), Jing Dong (New Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences)
CodeLarge Language ModelSupervised Fine-TuningVision Language ModelImage
π― What it does: Propose the Q-Scorer framework to address conversion errors and semantic confusion caused by discrete tokens in MLLMs during image quality assessment.
π― What it does: Proposed a lightweight infrared-visible image fusion method called AIDFusion, aiming to address the network laziness problem and enhance fusion performance.
Revisiting the Data Sampling in Multimodal Post-training from a Difficulty-Distinguish View
Jianyu Qi (Central South University), Rongchang Zhao (Central South University)
CodeData-Centric LearningSupervised Fine-TuningReinforcement LearningVision Language ModelMultimodalityBenchmark
π― What it does: Propose two difficulty-aware sampling strategies (PISM and CMAB), and construct a multi-modal post-training framework based on difficulty hierarchy, demonstrating that training with GRPO alone can surpass the traditional SFT+GRPO scheme.
Revitalizing Canonical Pre-Alignment for Irregular Multivariate Time Series Forecasting
Ziyu Zhou (Hong Kong University of Science and Technology), Yuxuan Liang (Hong Kong University of Science and Technology)
CodeComputational EfficiencyConvolutional Neural NetworkTransformerTime SeriesBiomedical Data
π― What it does: Proposes KAFNet, an irregular multivariate time series (IMTS) prediction model that leverages Canonical Pre-Alignment (CPA), integrating pre-convolution, temporal kernel aggregation, and frequency-domain linear attention modules.
Reward Redistribution via Gaussian Process Likelihood Estimation
Minheng Xiao (Ohio State University), Xian Yu (Ohio State University)
CodeReinforcement Learning
π― What it does: Propose a Gaussian Process-based Likelihood Reward Redistribution framework (GP-LRR), which generates dense reward signals by maximizing the likelihood of the entire trajectory through a leave-one-out strategy, and combines it with Soft Actor-Critic (SAC) to significantly improve sample efficiency and final performance in environments with sparse terminal rewards.
RFI: Rectified Flow Intervention for Mitigating Object Hallucination in Large Vision-Language Models
Junyu Cheng (Xiamen University), Shuangyin Li (South China Normal University)
CodeObject DetectionVision Language ModelRectified FlowImageMultimodality
π― What it does: Propose RFI (Rectified Flow Intervention), a lightweight method that dynamically predicts hidden space intervention vectors in large vision-language models to suppress object hallucinations.
RflyPano: A Panoramic Benchmark for Ultra-low Altitude UAV Localization Powered by RflySim
Dun Dai (Beihang University), Quan Quan (Central South University)
CodeData SynthesisPose EstimationRetrievalAutonomous DrivingConvolutional Neural NetworkTransformerSimultaneous Localization and MappingImageBenchmark
π― What it does: This paper constructs the first panoramic visual localization benchmark for ultra-low altitude (<120 m) drones - the RflyPano dataset, and verifies two categories of localization methods (image retrieval and pose regression) on it.
RFNNS: Robust Fixed Neural Network Steganography with Universal Text-to-Image Models
Yu Cheng (East China Normal University), Xinpeng Zhang (Fudan University)
CodeGenerationSafty and PrivacyDiffusion modelImage
π― What it does: Proposed a robust steganography method RFNNS based on a fixed neural network and a general text-to-image model, which embeds robust perturbations only in texture-complex regions of the cover image to achieve steganography and recovery.
π― What it does: Proposed the RGMP framework, achieving a geometry-prior multi-modal policy based on voice commands, which includes a Geometric-prior Skill Selector and an Adaptive Recursive Gaussian Network, completing the full process from semantic parsing to trajectory generation;
RI-Loss: A Learnable Residual-Informed Loss for Time Series Forecasting
Jieting Wang (Shanxi University), Furong Peng (Shanxi University)
CodeTransformerTime Series
π― What it does: Propose a loss function called RI-Loss based on the residual-random noise mutual information, utilizing the Hilbert-Schmidt Independence Criterion (HSIC) to model the statistical independence between prediction residuals and noise, thereby simultaneously suppressing observational noise and capturing long-term dependencies in time series.
RiemanLine: Riemannian Manifold Representation of 3D Lines for Factor Graph Optimization
Yan Li, Gim Hee Lee (University Of Edinburgh)
CodeOptimization
π― What it does: Proposed a Riemannian manifold-based 3D line minimal parameterization (RiemanLine) that simultaneously handles single lines and parallel line sets, significantly reducing parameter dimensionality;
Right Branches Matter in Failure-based Variable Ordering Heuristics
Yang Zhang (Northeast Normal University), Hongbo Li (Northeast Normal University)
CodeOptimizationBenchmark
π― What it does: Improved and experimented with a variable ordering heuristic based on failure rate and failure length using right branch failure information
Alonso Granados (University of Arizona), Jason Pacheco (University of Arizona)
CodeReinforcement Learning
π― What it does: Proposed the risk-sensitive exponential actor-critic (rsEAC) algorithm, combining a new gradient theorem and numerically stable exponential value function estimation to achieve deep reinforcement learning with entropy risk measures.
π― What it does: Propose a dual-branch U-Net combined with reinforcement learning (PPO) to achieve cross-modal feature alignment and fusion between CT and MRI, thereby enabling precise 3D whole-heart segmentation.
π― What it does: This paper proposes a reinforcement learning-based knowledge distillation framework, RLKD, which utilizes a Generative Structural Reward Model (GSRM) to transfer the teacher LLM's implicit multi-branch reasoning structure to the student LLM, addressing the issue that traditional SFT can only replicate surface-level reasoning paths.
π― What it does: Proposed a lightweight recursive multi-scale feature atmospheric turbulence suppressor (RMFAT), which performs two-frame recursive restoration using only the current frame and the previous frame;
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposed the Reward Margin Optimization (RMO) framework, which enhances preference alignment in large language models (LLMs) by reshaping the reward margin distribution during data preparation, batch construction, and training processes.
RoadSceneVQA: Benchmarking Visual Question Answering in Roadside Perception Systems for Intelligent Transportation System
Runwei Guan (Hong Kong University of Science and Technology), Yutao Yue (Hong Kong University of Science and Technology)
CodeAutonomous DrivingTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
π― What it does: This paper proposes the RoadSceneVQA dataset and the RoadMind model, aiming to enhance visual question answering and reasoning capabilities for roadside perception.
Robust Causal Discovery Under Imperfect Structural Constraints
Zidong Wang (City University of Hong Kong), Xiaoguang Gao (Northwestern Polytechnical University)
CodeOptimizationExplainability and InterpretabilityGraphBiomedical Data
π― What it does: Propose a robust causal discovery framework named RoaDs, based on prior alignment and multi-task learning, for causal graph learning under imperfect structural constraints.
Robust Integrative Analysis of Multi-omics Datasets via Nuclear-norm Maximization
Meng-Zhu Wang (Hebei University of Technology), Hongxing Zhang (National Center for Protein Sciences)
CodeOptimizationRepresentation LearningGraph Neural NetworkContrastive LearningBiomedical Data
π― What it does: Developed a spatial multi-omics integration method RIA based on batch nuclear norm maximization (BNM), leveraging adaptive graph learning and dynamic prototype contrastive learning to construct more discriminative and diverse latent representations.
Robust Lazy Conflict Detection via Multi-Conflict Extraction and Genetic Diversity Control
Viet-Man Le (Graz University of Technology), Alexander Felfernig (Graz University of Technology)
CodeOptimizationTabular
π― What it does: This study proposes an improved lazy conflict detection method, enhancing conflict set coverage and robustness through multi-conflict extraction and genetic diversity control.
π― What it does: This paper proposes a robust multi-agent combined path planning framework named Robust CBSS, and implements two algorithms, RCbssBase and RCbssEff, which can accomplish dynamic task allocation and path planning in real robot environments with directional constraints and probabilistic delays.
Robust Pseudo-Labeling via Decoupled Class-Aware Filtering and Dynamic Category Correction
Jianghang Lin (Xiamen University), Liujuan Cao (Xiamen University)
CodeSegmentationTransformerVision Language ModelImage
π― What it does: To address the pseudo-label noise issue in semi-supervised instance segmentation, the PL-DC framework is proposed, which includes decoupled filtering, dynamic class correction, and pixel-level uncertainty weighting.
Robust Semi-paired Multimodal Learning for Cross-modal Retrieval
Yang Qin, Peng Hu (Sichuan University)
CodeRetrievalContrastive LearningMultimodality
π― What it does: This paper studies image-text retrieval in the semi-aligned multimodal learning (SPL) scenario and proposes a robust cross semi-aligned learning (RCSL) framework;
π― What it does: Proposed a robust multi-modal 3D detector called RobusTor3D, which enhances model robustness against long-tail distributions, adverse weather, sensor errors, modality missing, and cross-domain scenarios through knowledge from vision-language models at both structural and supervisory layers.
RoSA: Enhancing Parameter-Efficient Fine-Tuning via RoPE-aware Selective Adaptation in Large Language Models
Dayan Pan (Beihang University), Xiangyu Zhao (City University of Hong Kong)
CodeOptimizationComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Designed a new parameter-efficient fine-tuning framework named RoSA, which integrates RoPE-aware attention enhancement and dynamic layer selection;
CodeRecognitionGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Propose a multi-event argument extraction model called RoSE, which explicitly models the heterogeneity and overlap of event structures through role-related structures to improve the accuracy of multi-event argument extraction.
RouterNet: Hierarchical Point Routing Network for Robust Vertebral Landmark Localization on AP X-ray Images
Yingjie Guo (Huazhong University of Science and Technology), Zhiwei Wang (Huazhong University of Science and Technology)
CodePose EstimationConvolutional Neural NetworkImageBiomedical Data
π― What it does: Proposed a hierarchical point routing network called RouterNet for precise localization of vertebral landmark points in AP X-ray images, and utilized the localization results for automatic assessment of scoliosis.
RPTS: Tree-Structured Reasoning Process Scoring for Faithful Multimodal Evaluation
Haofeng Wang (Harbin Institute of Technology), Yu Zhang (Harbin Institute of Technology)
CodeExplainability and InterpretabilityLarge Language ModelMultimodalityBenchmarkChain-of-Thought
π― What it does: This paper proposes a tree-structured reasoning process scoring metric called RPTS, and builds a multimodal reasoning benchmark named RPTS-Eval based on it.
π― What it does: This study first formalizes shilling attacks targeting citation recommendation in academic papers and proposes the RSA-CR algorithm, which achieves global citation recommendation with robustness against attacks through dual-layer academic graphs and confidence-guided aggregation (Dumbbell Inductive Learning).
π― What it does: Proposes the RSOD semi-supervised sonar image object detection framework, which uses a teacher-student model to evaluate the reliability of pseudo-labels and generate hybrid pseudo-labels, significantly enhancing detection performance even with minimal annotations.
π― What it does: This paper proposes a macro placement method called RSPlace based on reinforcement learning, which first incorporates the rotation angle of macros into the search space;
RTMol: Rethinking Molecule-text Alignment in a Round-trip View
Letian Chen (Shanghai Innovation Institute), Yang Yang (Shanghai Jiao Tong University)
CodeDrug DiscoveryTransformerLarge Language ModelReinforcement LearningTextGraph
π― What it does: Through the self-supervised cyclic learning framework RTMol, the tasks of molecular-to-text description and text-to-molecular generation are unified, achieving bidirectional alignment between molecules and text.
Run, Ruminate, and Regulate: A Dual-process Thinking System for Vision-and-Language Navigation
Yu Zhong (Chinese Academy of Sciences), Yunji Chen (Chinese Academy of Sciences)
CodeAutonomous DrivingRobotic IntelligenceGraph Neural NetworkTransformerLarge Language ModelVision Language ModelVision-Language-Action ModelImageTextMultimodalityChain-of-Thought
π― What it does: Designed and implemented a dual-process thinking framework called R3, consisting of three modules: a lightweight Runner, an LLM-driven Ruminator, and a Regulator, achieving efficient and accurate path planning in visual language navigation tasks.
S-DAG: A Subject-Based Directed Acyclic Graph for Multi-Agent Heterogeneous Reasoning
Jiangwen Dong (Hong Kong Polytechnic University), Mingjin Zhang (Hong Kong Polytechnic University)
CodeGraph Neural NetworkTransformerLarge Language ModelAgentic AIMixture of ExpertsTextBenchmark
π― What it does: Construct a topic-based directed acyclic graph (S-DAG) to fine-grainedly identify relevant knowledge domains in interdisciplinary problems, and deploy domain-expert large language models (LLMs) collaboratively on this graph to achieve multi-agent reasoning.
S2C: A Noise-Resistant Difference Learning Framework for Unsupervised Change Detection in VHR Remote Sensing Images
Lei Ding (Information Engineering University), Jicang Lu (Information Engineering University)
CodeTransformerContrastive LearningImageBenchmark
π― What it does: Developed an unsupervised change detection framework S2C based on visual foundation models and contrastive learning, achieving semantic-to-change mapping in very high-resolution (VHR) remote sensing images.
π― What it does: Propose a self-supervised hypergraph sequential recommendation framework S HyRec, integrating global intent tendency, temporal context intent, sequence dependency awareness, and cross-perspective self-supervised learning to enhance sequential recommendation performance.
SΒ²Teacher: Step-by-step Teacher for Sparsely Annotated Oriented Object Detection
Yu Lin (Xiamen University), Liujuan Cao (Xiamen University)
CodeObject DetectionImage
π― What it does: Studies directional object detection under sparse annotation, proposing the S Teacher framework to achieve high-precision detection in scenarios with only a few annotated instances.
SΒ³: Spiking Neurons as an Isolating Segmenter for Brain Signal Decoding
Qian Zheng (Zhejiang University), Gang Pan (Zhejiang University)
CodeSegmentationSpiking Neural NetworkBiomedical Data
π― What it does: Proposed the S3 model, which uses spiking neurons for adaptive segmentation of EEG signals, taking into account individual and task differences while preserving temporal patterns.
π― What it does: A framework named S3Net is proposed for processing dynamic visual sensor (DVS) event streams, achieving efficient asynchronous processing through learnable voxel sparse coding and a spatiotemporal separation dual-branch network.
CodeSegmentationTransformerMixture of ExpertsImageBenchmark
π― What it does: Propose a scalable semi-supervised semantic segmentation framework S5, which pretrains an RS base model using a large amount of unlabeled remote sensing images.
π― What it does: SACodec is a low-bitrate neural speech codec that can simultaneously provide high-fidelity audio reconstruction and semantically rich discrete tokens at 1.5 kbps.
Safe Multi-Agent Reinforcement Learning via Distributional Safety Critic and Maximum Entropy Optimization
Qiwei Liu (East China University of Science and Technology), Huaicheng Yan (East China University of Science and Technology)
CodeOptimizationReinforcement LearningBenchmark
π― What it does: This paper proposes a Worst-Case Multi-Agent Soft Actor-Critic (WCMASAC) framework based on maximum entropy and distributed safety assessment, aimed at learning stochastic policies for multi-agent systems under safety constraints.
Safe RAG by RAG: Untying the Bell That RAG Rang with the RAG Hand
Xun Liang (Renmin University of China), Simin Niu (Renmin University of China)
CodeSafty and PrivacyKnowledge DistillationContrastive LearningTextRetrieval-Augmented Generation
π― What it does: Developed a framework-level security solution for RAG2RAG, incorporating Detective and Judge dual modules to parallel supervise RAG generation, thereby enhancing security.
SAFE: Semantic- and Frequency-Enhanced Curriculum for Cross-Domain Deepfake Detection
Yulin Yao (Beijing University of Posts and Telecommunications), Dan Luo (Beijing University of Posts and Telecommunications)
CodeDomain AdaptationAnomaly DetectionTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningVideoTextMultimodality
π― What it does: Built the SAFE framework, combining semantically enhanced multimodal alignment and dual-score curriculum learning, specifically designed for cross-domain deepfake detection.
SafeNLIDB: A Privacy-Preserving Safety Alignment Framework for LLM-based Natural Language Database Interfaces
Ruiheng Liu (Xi'an Research Institute of High-Tech), Bailong Yang (Xi'an Research Institute of High-Tech)
CodeData SynthesisSafty and PrivacyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextTabularBenchmarkChain-of-Thought
π― What it does: Developed an end-to-end security alignment framework, SAFENLIDB, to prevent privacy leakage in LLM-driven natural language database interfaces while maintaining the reliability of SQL generation.
π― What it does: Perform safe fine-tuning on CLIP, proposing SafeR-CLIP which achieves migration and elimination of NSFW content through neighbor-based safety redirection.
SafetyReminder: Reviving Delayed Safety Awareness of Vision-Language Models to Defend Against Jailbreak Attacks
Peiyuan Tang (Xi'an Jiaotong University), Zijiang James Yang (Singapore Management University)
CodeSafty and PrivacyAdversarial AttackSupervised Fine-TuningPrompt EngineeringVision Language ModelMultimodality
π― What it does: Discover that Vision-Language Models (VLMs) exhibit 'delayed safety awareness' during the generation process, and propose the SafetyReminder framework, which utilizes SAPT to learn soft prompts that activate safety awareness in the intermediate generation stage, preventing the generation of malicious content.
π― What it does: Proposes the SAGA method, which learns a Gaussian distribution with high success rates under text prompts, directly samples from it, and performs fine-tuning to enhance the semantic alignment quality in text-to-image generation.
SAGE: Spuriousness-Aware Guided Prompt Exploration for Mitigating Multimodal Bias
Wenqian Ye (University of Virginia), Aidong Zhang (University of Virginia)
CodeClassificationPrompt EngineeringVision Language ModelContrastive LearningMultimodality
π― What it does: Studied how to suppress multimodal pseudo correlations in zero-shot classification tasks of pre-trained vision-language models such as CLIP, and proposed the SAGE method to enhance model robustness through prompt exploration.
SageLM: A Multi-aspect and Explainable Large Language Model for Speech Judgement
Yuan Ge (Northeastern University), Jingbo Zhu (Northeastern University)
CodeData SynthesisExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodalityAudio
π― What it does: This paper proposes SageLM, an end-to-end, interpretable multi-dimensional speech evaluation model capable of simultaneously assessing semantic and acoustic features.
SAM-DAQ: Segment Anything Model with Depth-guided Adaptive Queries for RGB-D Video Salient Object Detection
Jia Lin (Hangzhou Dianzi University), Jiyong Zhang (Hangzhou Dianzi University)
CodeSegmentationTransformerVideoMultimodality
π― What it does: This paper proposes a RGB-D video salient object detection framework called SAM-DAQ based on SAM2, achieving salient object segmentation without manual prompts through depth-guided parallel adapters and query-driven temporal memory modules.
SAM2-OV: A Novel Detection-Only Tuning Paradigm for Open-Vocabulary Multi-Object Tracking
Yangkai Chen (Xiamen University), Hanzi Wang (Xiamen University)
CodeObject DetectionObject TrackingTransformerVision Language ModelContrastive LearningImageVideo
π― What it does: Adopting a detection-only fine-tuning paradigm, leveraging SAM2's zero-shot cross-frame association to achieve open-vocabulary multi-object tracking.
SAMCL: Empowering SAM to Continually Learn from Dynamic Domains with Extreme Storage Efficiency
Zeqing Wang (Xidian University), Fei Cheng (Xidian University)
CodeSegmentationDomain AdaptationComputational EfficiencyPrompt EngineeringImageBiomedical Data
π― What it does: Propose a continual learning method named SAMCL, enabling SAM to incrementally learn and maintain performance across multiple domains.
π― What it does: This paper proposes a sample weighted incomplete multimodal clustering method based on graph coarsening label extraction (IMC-GCSW), aiming to simultaneously address the incompleteness and quality inconsistency of multimodal data.
π― What it does: Proposed the SMCIR framework, which can diagnose missing modalities for each sample and complete missing modalities through cross-modal enhancement, thereby improving the performance of multimodal sentiment analysis.
π― What it does: Proposes the SC-SSL framework, which suppresses class imbalance in semi-supervised learning through decoupling sampling control, with particular emphasis on feature learning for minority classes and numerical calibration;
SAOT: An Enhanced Locality-Aware Spectral Transformer for Solving PDEs
Chenhong Zhou (Hong Kong Baptist University), Zaifeng Yang (Agency for Science, Technology and Research)
CodeTransformerMeshPhysics Related
π― What it does: Propose a spectral Transformer (SAOT) that integrates wavelet attention and Fourier attention to efficiently approximate the solution operators of PDEs.
π― What it does: In this paper, the authors propose a post-training quantization framework named SAQ-SAM for the Segment Anything Model (SAM), aiming to significantly enhance low-bit quantization performance through semantically consistent pruning and prompt-aware reconstruction.
SAR: A Structure-Aligned Reasoning Framework for Temporal Knowledge Graph Question Answering
Qianyi Hu (Central China Normal University), Shoujin Wang (University of Technology Sydney)
CodeTransformerLarge Language ModelAgentic AIPrompt EngineeringTextGraphRetrieval-Augmented Generation
π― What it does: Propose the SAR framework, integrating LLM reasoning, structural alignment retrieval, and iterative verification to address structural mismatch issues in temporal knowledge graph question answering.
π― What it does: Construct a structure-preserving diffusion framework, Sat2Flow, to generate urban OD flow matrices using satellite imagery as the sole input.
Satisficing and Optimal Generalised Planning via Goal Regression
Dillon Z. Chen (Vector Institute), Sheila A. McIlraith (Vector Institute)
CodeOptimizationBenchmark
π― What it does: Propose a generic planning method called MOOSE based on goal regression, which can generate general plans that are directly executable or used for search pruning from training instances.
Scalable Multi-Objective and Meta Reinforcement Learning via Gradient Estimation
Zhenshuo Zhang (Northeastern University), Hongyang R. Zhang (Northeastern University)
CodeMeta LearningReinforcement Learning
π― What it does: Propose a scalable multi-objective reinforcement learning method that first trains a meta-policy, then rapidly evaluates the adaptability of any task subset using first-order gradient estimation to obtain a task affinity matrix and cluster groups;
Scalable Solutions to Zero-Sum Partially Observable Stochastic Games Through Belief Aggregation with Approximation Guarantees
Kim Hammar (University of Melbourne), Tansu Alpcan (University of Melbourne)
CodeReinforcement LearningBenchmark
π― What it does: Propose a new method called SAB (Shapley Iteration with Aggregated Beliefs) to approximately solve the value function of one-sided zero-sum partially observable stochastic games (POSG).
Harrison H Li (Harvey Mudd College), David B. Lobell (Stanford University)
CodeConvolutional Neural NetworkImageAgriculture Related
π― What it does: Propose combining predictive projection inference (PPI) with computer vision models, leveraging field photos and geographic coordinates to supplement limited crop cutting measurements, thereby improving the accuracy of regional average yield estimates.
π― What it does: Proposed Scale-Net, a unified multi-task vehicle routing neural solver, achieving cross-scale generalization from small-scale training to large-scale instances with thousands of nodes through a hierarchical U-Net framework;
π― What it does: Decompose mathematical problems into subproblems, allocate computational resources according to difficulty, and achieve selective computation during testing.
Scaling and Transferability of Annealing Strategies in Large Language Model Training
Siqi Wang (Hong Kong University of Science and Technology), Xiaomeng Li (Meituan Inc)
CodeOptimizationHyperparameter SearchLarge Language ModelMixture of ExpertsText
π― What it does: Studied and verified the transferability of learning rate annealing strategies in large-scale language model training, and proposed an improved prediction framework;
Scaling Laws for Conditional Emergence of Multilingual Image Captioning via Generalization from Translation
Julian Spravil (Fraunhofer IAIS), Sven Behnke (University of Bonn)
CodeGenerationTransformerSupervised Fine-TuningVision Language ModelContrastive LearningTextMultimodality
π― What it does: Investigated the use of multimodal translation task training data to achieve zero-shot generalization for image captioning in unseen languages within multilingual multitask scenarios.
Scaling LLM Speculative Decoding: Non-Autoregressive Forecasting in Large-Batch Scenarios
Luohe Shi (Wuhan University), Hai Zhao (Shanghai Jiao Tong University)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
π― What it does: Propose SpecFormer, a non-autoregressive draft generation architecture that combines unidirectional and bidirectional attention, designed to accelerate lossless autoregressive inference in large-batch inference scenarios;
Scaling Towards the Information Boundary of Instructions through Data Synthesizing
Li Du (Beijing Academy of Artificial Intelligence), Tengfei Pan (Beijing Academy of Artificial Intelligence)
CodeData SynthesisData-Centric LearningLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposed a closed-loop instruction data construction framework and built a high-quality instruction dataset named Infinity Instruct Subject (InfInstruct-Sub) with approximately 1.5 million entries.
CodeData-Centric LearningTransformerLarge Language ModelVision Language ModelMultimodalityBenchmark
π― What it does: Constructed a large-scale multi-modal instruction database OmniVQA-Chat-400K and a video quality rating dataset OmniVQA-MOS-20K, achieving data scaling through machine + human annotation methods;
π― What it does: Propose a self-calibrating autoregressive visual generative model called SCAN, which can evaluate and refine generated image patches without regenerating the entire image.
scCluBench: Comprehensive Benchmarking of Clustering Algorithms for Single-Cell RNA Sequencing
Ping Xu (Computer Network Information Center, Chinese Academy of Sciences), Yuanchun Zhou (Computer Network Information Center, Chinese Academy of Sciences)
π― What it does: Provides a unified, standardized benchmark framework for scRNA-seq clustering algorithms (scCluBench), aggregating 36 datasets covering 18 human and mouse tissues, spanning various scales and sparsity levels, and conducting systematic evaluations of 36 traditional, deep learning, graph neural network, and biology-based clustering methods.
SceneJailEval: A Scenario-Adaptive Multi-Dimensional Framework for Jailbreak Evaluation
Lai Jiang (Shanghai Jiao Tong University), Li Pan (Shanghai Jiao Tong University)
CodeClassificationSafty and PrivacyLarge Language ModelTextBenchmark
π― What it does: Constructed a scenario-adaptive multi-dimensional evaluation framework named SceneJailEval, and collected a high-quality jailbreak dataset containing 14 scenarios and 1,308 examples based on this framework, used to evaluate the jailbreak success rate and risk level of LLMs.
SciMKG: A Multimodal Knowledge Graph for Science Education with Text, Image, Video and Audio
Tong Lu, Junsheng Du (Beijing Normal University)
CodeTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodalityAudio
π― What it does: This paper proposes an automated framework that utilizes large language models to extract concepts and perform multi-modal alignment on K12 science MOOC resources, constructing a multi-modal educational knowledge graph called SciMKG that includes text, images, videos, and audio.
SCIR: A Self-Correcting Iterative Refinement Framework for Enhanced Information Extraction Based on Schema
Yushen Fang (Huazhong University of Science and Technology), Wenqi Yang (Huazhong University of Science and Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
π― What it does: Propose the SCIR self-correcting iterative refinement framework, integrating Dual-Path Self-Correcting and Feedback-Driven Optimization, combined with the MBSC dataset, to achieve efficient information extraction without fine-tuning across languages;
π― What it does: SCo-Cloud proposes a constellation architecture based on collaboration between central satellites and edge satellites, achieving cloud detection, thin cloud removal, thick cloud repositioning localization, task scheduling, and content-aware downlink;
Scope Delineation Before Localization: A Two-Stage Framework for Enhancing Failure Attribution in Multi-Agent Systems
Kai Sun (Xi'an Jiaotong University), Bin Shi (Xi'an Jiaotong University)
CodeAnomaly DetectionTransformerLarge Language ModelPrompt EngineeringTextSequentialChain-of-Thought
π― What it does: Proposed a two-stage failure attribution framework called Scope Delineation Before Localization (SDBL), which improves failure attribution accuracy in multi-agent systems by first delineating the failure scope and then precisely localizing the failure steps.
π― What it does: This paper proposes a score-matching-based energy model that directly learns the gradient of the joint distribution between tensors and latent factors, achieving low-rank tensor completion and denoising.
SculptDrug: A Spatial Condition-Aware Bayesian Flow Model for Structure-based Drug Design
Qingsong Zhong (East China Normal University), Jilin Hu (East China Normal University)
CodeDrug DiscoveryFlow-based ModelBiomedical Data
π― What it does: Proposed SculptDrug, a spatially condition-aware structural drug design model based on Bayesian flow networks, for generating drug molecules that comply with protein surface constraints.
SD-PSFNet: Sequential and Dynamic Point Spread Function Network for Image Deraining
Jiayu Wang (University of Electronic Science and Technology of China), Shaoning Zeng (University of Electronic Science and Technology of China)
CodeRestorationConvolutional Neural NetworkImage
π― What it does: Propose a multi-stage sequential network SD-PSFNet based on dynamic point spread function (PSF) for single raindrop image de-raining;
SDA: Steering-Driven Distribution Alignment for Open LLMs Without Fine-Tuning
Wei Xia (Peking University), Zhi-Hong Deng (Peking University)
CodeGenerationTransformerLarge Language ModelText
π― What it does: Proposed a framework named SDA for distribution realignment of large language models during inference without training or parameter modifications, which can dynamically reallocate output probabilities during generation according to user instructions, thereby enhancing model consistency with human intentions.