π― What it does: Propose the ASI framework, which enhances the exploration and exploitation efficiency of GCHRL by constructing a subgoal potential graph to achieve active boundary exploration and hierarchical self-imitation.
Enhancing Medical Large Vision-Language Models via Alignment Distillation
Aofei Chang (Pennsylvania State University), Fenglong Ma (Pennsylvania State University)
CodeExplainability and InterpretabilityKnowledge DistillationSupervised Fine-TuningVision Language ModelBiomedical Data
π― What it does: Proposes the MEDALIGN framework, which enhances the accuracy and interpretability of image understanding and text generation by transferring visual representations and attention knowledge from a professional CLIP model to a medical large vision-language model (Med-LVLM) through alignment distillation.
Enhancing Meme Emotion Understanding with Multi-Level Modality Enhancement and Dual-Stage Modal Fusion
Yi Shi (Zhejiang University), Wenzhi Chen (Zhejiang University)
CodeClassificationRecognitionTransformerLarge Language ModelVision Language ModelMultimodalityChain-of-Thought
π― What it does: Propose the MemoDetector framework, which utilizes a multi-modal large language model (MLLM) for four-step text enhancement and designs a two-stage cross-modal fusion strategy to achieve more precise identification of meme sentiment.
π― What it does: In the multimodal misinformation detection task, this paper proposes the RETSIMD method: first, the text is divided into several paragraphs, and a pre-trained text-image generator is used to generate a supplementary image for each paragraph; then, a graph structure incorporating the original and generated images is constructed, and graph neural networks are utilized to fuse features from multiple images; finally, the fused image features are combined with text features for authenticity judgment.
Enhancing Pre-training Data Detection in LLMs Through Discriminative and Symmetric Prefix Selection
Kai Sun (Xi'an Jiaotong University), Bin Shi (Xi'an Jiaotong-Liverpool University)
CodeData-Centric LearningTransformerLarge Language ModelText
π― What it does: This paper proposes an unlabeled prefix selection framework called DSC-Prefix to enhance the accuracy and robustness of LLM pre-training data detection methods RECALL and CON-RECALL.
Enhancing Retrieval-Augmented Large Vision Language Models via Knowledge Conflict Mitigation
Wenbin An (Xi'an Jiaotong University), Shijian Lu (Zhejiang University of Technology)
CodeTransformerVision Language ModelMultimodalityRetrieval-Augmented Generation
π― What it does: This paper proposes a training-agnostic, plug-and-play framework called KCM, aiming to reduce knowledge conflicts in retrieval-augmented vision-language models and improve generation quality by modulating attention allocation, pruning parameterized neurons, and reinforcing context logits.
π― What it does: This paper proposes Shadow-informed Pose Feature (SiPF) and Rotation-Invariant Attention Convolution (RIAttnConv), enhancing rotation-invariant learning in 3D point clouds through global pose reference and attention mechanisms.
Enhancing Spatial Reasoning Through Visual and Textual Thinking
Xun Liang (Zhejiang University), Jieping Ye (Alibaba Cloud Computing)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
π― What it does: Propose the SpatialVTS method, combining visual reasoning and textual reasoning stages to enhance the performance of Vision-Language Models (VLMs) in 2D/3D spatial reasoning tasks.
Enumerating Minimal Unsatisfiable Cores of LTLf Formulae
Antonio Ielo (University of Calabria), Francesco Ricca (University of Calabria)
CodeBenchmark
π― What it does: Studied a method that utilizes Answer Set Programming (ASP) to enumerate minimal unsatisfiable cores (MUC) of finite trace linear temporal logic (LTLf) formulas.
EoH-S: Evolution of Heuristic Set Using LLMs for Automated Heuristic Design
Fei Liu (City University of Hong Kong), Mingxuan Yuan (Huawei Noah's Ark Lab)
CodeOptimizationTransformerLarge Language ModelBenchmark
π― What it does: This paper proposes the Automated Heuristic Set Design (AHSD) framework and the EoH-S evolutionary algorithm, which utilize large language models to generate and evolve a set of small-scale complementary heuristics, addressing the generalization issues of traditional single heuristic designs.
EPIC: Explanation of Pretrained Image Classification Networks via Prototypes
Piotr Borycki (Jagiellonian University), Εukasz Struski (Jagiellonian University)
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
π― What it does: Propose EPIC, a post-hoc prototype explanation framework that provides interpretable prototypes for image classification models without modifying the pre-trained network architecture or requiring retraining
π― What it does: Proposes a multi-strategy jailbreak method called EquaCode, which first transforms malicious requests into mathematical equation-solving tasks and then embeds the solution into a Python template via code completion, inducing LLMs to generate harmful content with a single query.
Equivariant Atomic and Lattice Modeling Using Geometric Deep Learning for Crystal Structure Optimization
Ziduo Yang (Jinan University), Lei Shen (Jinan University)
CodeOptimizationGraph Neural NetworkGraphPhysics Related
π― What it does: This study proposes an end-to-end E3Relax network that directly maps unrelaxed crystal structures to their DFT-relaxed lowest energy structures;
ERank: Fusing Supervised Fine-Tuning and Reinforcement Learning for Effective and Efficient Text Reranking
Yuzheng Cai (Fudan University), Weiguo Zheng (Fudan University)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: ERANK is a pointwise text re-ranker based on large language models, employing a two-stage training approach: first using supervised fine-tuning with fine-grained integer scoring, followed by optimization through reinforcement learning with list-wise rewards.
ESCA: An Emotional Support Conversation Agent for Enhancing Reasonable Strategy Planning and Effective Expression
Jing Li (Northeastern University), Gang Wu (Northeastern University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringTextRetrieval-Augmented Generation
π― What it does: This paper proposes ESCA, an emotional support chat agent based on a large language model that simultaneously enhances strategy planning and response generation.
Jinping Wang (Wenzhou-Kean University), Zhiwu Xie (Wenzhou-Kean University)
CodeClassificationOptimizationTransformerImage
π― What it does: A multi-step optimization framework DV-BASI is studied, which uses differential vectors to continuously explore task arithmetic for improving model editing effectiveness.
π― What it does: Proposes the Uncertainty-Guided Reliable Learning (UGRL) framework for weakly supervised semantic segmentation, integrating three modules: Prototype-driven Uncertainty Modeling, Loss Modulation, and Reliable Semantic Enhancement, to improve CAM pseudo-labels and decoder learning.
ESMC: MLLM-Based Embedding Selection for Explainable Multiple Clustering
Xinyue Wang, Junhui Hou (Saint Francis University)
CodeExplainability and InterpretabilityLarge Language ModelPrompt EngineeringVision Language ModelMultimodality
π― What it does: Propose an interpretable multi-clustering framework (ESMC) based on the hidden states of multimodal large language models, achieving user-specified semantic clustering by extracting feature embeddings from text prompts and combining them with a lightweight MLP clustering head.
π― What it does: The paper proposes an analytical formula for maximum likelihood estimation (MLE) based on the randomized response (RR) mechanism and its fast implementation, comparing it with existing iterative estimation methods.
EvalMuse-40K: A Fine-Grained Benchmark with Comprehensive Human Annotations for Text-to-Image Generation Model Alignment Evaluation
Shuhao Han (Nankai University), Chongyi Li (Nankai University)
CodeTransformerPrompt EngineeringVision Language ModelDiffusion modelMultimodalityBenchmark
π― What it does: Constructed the EvalMuse-40K large-scale text-image alignment evaluation dataset with fine-grained human annotations, and proposed an efficient FGA-BLIP2 evaluation method.
Evaluating, Synthesizing, and Enhancing for Customer Support Conversation
Jie Zhu (Soochow University), Fang Kong (Alibaba Cloud Computing)
CodeGenerationData SynthesisLarge Language ModelSupervised Fine-TuningText
π― What it does: Propose the Customer Service Conversation (CSC) task and construct a high-quality evaluation dataset CSConv and a role-play-based synthetic dataset RoleCS; significantly improve the model's strategy consistency and problem-solving ability on CSConv through fine-tuning LLMs on RoleCS.
π― What it does: Achieve super-resolution restoration of text images using event camera data, enhancing text readability in low-light and motion-blurred scenarios.
Event-Guided Super-Resolving Blurry Image via Asymmetric Integral Driven Consistency
Chi Zhang (Peng Cheng Laboratory), Wenhan Yang (Peng Cheng Laboratory)
CodeSuper ResolutionConvolutional Neural NetworkImageMultimodalityTime Series
π― What it does: This paper proposes a self-supervised event-driven blurry image super-resolution framework SE-SRB, achieving high-resolution sharpening by leveraging the spatiotemporal consistency of event streams and RGB blurry images.
Evidence-aware Integration and Domain Identification of Spatial Transcriptomics Data
Wei Zhang (Sichuan University), Wei Ju (Nankai University)
CodeDomain AdaptationRepresentation LearningGraph Neural NetworkBiomedical Data
π― What it does: Proposes the PREST framework for downstream analyses of spatial transcriptomics data, including domain identification, DEG detection, and expression interpolation.
EvoFMVC: Trusted Federated Multi-View Clustering with Evolutionary Fusion
Li Zhang (Shanxi University), Xinyan Liang (Shanxi University)
CodeFederated LearningSafty and PrivacyRepresentation LearningNeural Architecture SearchContrastive Learning
π― What it does: Designed and implemented a trustworthy federated multi-view clustering framework named EvoFMVC, leveraging lightweight trustworthy evidence exchange and evolutionary fusion to achieve privacy protection, low communication cost, and adaptive aggregation.
π― What it does: Proposed a black-box optimization algorithm called EvoGrad2, which combines evolutionary weighted gradient learning with second-order Taylor expansion correction to efficiently search for optimal solutions in high-dimensional non-convex spaces.
Enqiang Zhu (Guangzhou University), Pu Wu (Hithink RoyalFlush Information Network Co., Ltd.)
CodeOptimizationGraph
π― What it does: Developed an exact branch-and-bound algorithm named ParDS to solve the Minimum Dominating Set (MDS) problem, significantly improving solving efficiency through an enhanced linear programming lower bound and multiple instance simplification rules.
Exact Shapley Attributions in Quadratic-time for FANOVA Gaussian Processes
Majid Mohammadi (Vrije Universiteit Amsterdam), Siu Lun Chau (CISPA Helmholtz Center for Information Security)
CodeExplainability and InterpretabilityComputational EfficiencyTabular
π― What it does: Proposed a method to compute exact Shapley values for FANOVA Gaussian processes in quadratic time, for local and global explanations.
ExPairT-LLM: Exact Learning for LLM Code Selection by Pairwise Queries
Tom Yuviler (Technion), Dana Drachsler-Cohen (Technion)
CodeComputational EfficiencyAI Code AssistantTransformerLarge Language ModelContrastive LearningTextBenchmark
π― What it does: Propose ExPairT-LLM, a precise learning code selection algorithm based on LLM, which identifies the correct implementation among candidate programs by utilizing pairwise membership and equivalence queries.
Expand Your SCOPE: Semantic Cognition over Potential-Based Exploration for Embodied Visual Navigation
Ningnan Wang (Xi'an Jiaotong University), Hongbin Sun (Xi'an Jiaotong University)
CodeRobotic IntelligenceVision Language ModelImageTextBenchmark
π― What it does: Proposed the SCOPE framework, achieving zero-shot embedded visual navigation through frontier boundary-driven potential estimation and potential maps;
π― What it does: Propose a scalable, differentiable dual-memory framework (EDD) for sample-free continual learning, utilizing shared memory and task-specific memory, and maintaining efficient knowledge utilization and stability through orthogonal regularization and memory alignment.
Experiential Fairness: Bridging the Gap Between User Experience and Resource-Centric Fairness in Online LLM Services
Jiahua Huang (South China University of Technology), Weiwei Lin (Pengcheng Laboratory)
CodeOptimizationTransformerLarge Language ModelText
π― What it does: Propose an experience fairness concept oriented towards user experience, design and implement the ExFairS scheduling framework, which balances resource utilization and SLO compliance rate, achieving dual optimal scheduling for multi-tenant LLM inference services.
Expert-Inspired Multi-Agent Coordination for Multi-Objective Molecular Optimization
Daojian Zeng, Xiangxiang Zeng (Hunan University)
CodeOptimizationDrug DiscoveryTransformerLarge Language ModelReinforcement LearningAgentic AIMixture of ExpertsBiomedical Data
π― What it does: Proposed a multi-objective molecular optimization framework called MAMO based on multi-agent large language models, which simulates expert collaboration cycles for molecular structure modification and property evaluation.
Explainable Depression Assessment from Face Videos by Weakly Supervised Learning
Rongfan Liao (University of Leicester), Siyang Song (Hohai University)
CodeClassificationExplainability and InterpretabilityTransformerVideoBenchmark
π― What it does: Propose a weakly supervised learning method that extracts and selects depression-related facial behaviors using multi-scale time intervals for video-level automatic depression assessment
CodeAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelDiffusion modelImageMultimodality
π― What it does: This paper proposes a framework called ESIDE, which detects synthetic images by utilizing intermediate noise images from the DDIM inverse process, and generates and optimizes interpretable defect descriptions through a multimodal large model.
Kayla Boggess (University of Virginia), Lu Feng (University of Virginia)
CodeExplainability and InterpretabilityReinforcement LearningBenchmark
π― What it does: This paper studies the explanation and summarization of distributed multi-agent reinforcement learning (MARL) strategies, proposing a policy summarization method based on Hasse diagrams and an explanation framework supporting When/Why Not/What queries.
Explaining Temporal Graph Neural Network via Quantum-Inspired Evolutionary Algorithm
Masahiro Mitani (University of Osaka), Yuya Sasaki (University of Osaka)
CodeExplainability and InterpretabilityGraph Neural NetworkGraph
π― What it does: Propose a unified explanation framework called QIEA-TGX based on a quantum-inspired evolutionary algorithm (QIEA) to generate factual and counterfactual explanations for Temporal Graph Neural Networks (TGNN), capable of handling both existing and non-existing events.
Explanation-Preserving Augmentation for Semi-Supervised Graph Representation Learning
Zhuomin Chen (Florida International University), Dongsheng Luo (Florida International University)
CodeExplainability and InterpretabilityRepresentation LearningGraph Neural NetworkContrastive LearningGraph
π― What it does: Propose a semi-supervised data augmentation method called EPA-GRL based on graph explanation, which uses a small number of labeled samples to train a graph interpreter to extract semantic subgraphs, preserving core semantics while perturbing the remaining structure, thereby improving the quality of graph contrastive learning representations.
π― What it does: Propose the EKD-Trip framework in travel recommendations, enhancing journey recommendation effectiveness through explicit intent-aware knowledge distillation.
Exploiting Inter-Session Information with Frequency-enhanced Dual-Path Networks for Sequential Recommendation
Peng He (University of Electronic Science and Technology of China), Qiao Liu (University of Electronic Science and Technology of China)
CodeRecommendation SystemSequential
π― What it does: Proposed a frequency-enhanced dual-path network called FreqRec for sequential recommendation, which parallelly integrates a global spectral aggregator and a local spectral refiner based on time-domain self-attention, and jointly trains with frequency-domain consistency loss and cross-entropy.
Exploiting Missing Data Remediation Strategies Using Adversarial Missingness Attacks
Deniz Koyuncu (Rensselaer Polytechnic Institute), Moti Yung (Google LLC)
CodeAdversarial AttackTabularBiomedical Data
π― What it does: Proposed a generic two-layer optimization framework to implement adversarial missing attacks under various missing data compensation strategies (complete case analysis, mean imputation, regression imputation, etc.); by learning differentiable missing mechanism parameters, attackers can shift learning models toward their expected targets while introducing only a small amount of missing data;
Exploiting Pre-trained Language Model for Cross-city Urban Flow Prediction Guided by Information-theoretic Analysis
Qiang Zhou (Nanjing University of Aeronautics and Astronautics), Jingjing Gu (Nanjing University of Aeronautics and Astronautics)
CodeDomain AdaptationGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningTime Series
π― What it does: Propose a cross-city traffic flow prediction framework based on pre-trained language models (PLM-CUP), achieving distribution alignment between source domain (city + language) and target domain through constructing a semantic bridge encoder and task-specific adapters.
Exploiting Synergistic Cognitive Biases to Bypass Safety in LLMs
Xikang Yang (Institute of Information Engineering, Chinese Academy of Sciences), Songlin Hu (Institute of Information Engineering, Chinese Academy of Sciences)
CodeSafty and PrivacyAdversarial AttackTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
π― What it does: Proposed and implemented a red team framework called CognitiveAttack based on collaborative cognitive biases, designed to generate adversarial prompts capable of bypassing LLM security mechanisms.
π― What it does: Proposes a parameter-efficient fine-tuning method called MuNG (Multimodal Noise Generator), which generates and injects 'beneficial noise' while keeping the multimodal large language model (MLLM) frozen;
π― What it does: Propose a pose tracking framework called PPF-Tracker based on pairwise features, focusing on 6D pose incremental learning and tracking for category-level articulated objects on the SE(3) manifold.
Exploring Efficient Open-Vocabulary Segmentation in the Remote Sensing
Bingyu Li (University of Science and Technology of China), Junyu Gao (Institute of Artificial Intelligence Teleai)
CodeSegmentationDomain AdaptationTransformerVision Language ModelContrastive LearningImageBenchmark
π― What it does: Proposed a unified open-vocabulary remote sensing image segmentation benchmark named OVRSISBench, and designed an efficient open-vocabulary segmentation framework called RSKT-Seg, specifically addressing rotation invariance and domain adaptation issues in remote sensing scenes.
π― What it does: Proposed a general remote sensing change detection framework GenCD based on vision foundation models (VFM), focusing on the generalizability of cross-domain change detection.
π― What it does: Investigated the use of wide-field fisheye camera systems for end-to-end 3D object detection and proposed a corresponding detection framework and dataset.
Exploring the Potentials of Spiking Neural Networks for Image Deraining
Shuang Chen (Durham University), Amir Atapour-Abarghouei (Durham University)
CodeRestorationSpiking Neural NetworkImage
π― What it does: This work proposes a single-image de-raining method based on spiking neural networks (SNN), introducing visual LIF neurons, a spiking decomposition and enhancement module (SDEM), and a multi-scale unit (SMU) to achieve efficient rain removal.
Exploring Transferability of Self-Supervised Learning by Task Conflict Calibration
Huijie Guo (Institute of Software Chinese Academy of Sciences), Wenwen Qiang (Institute of Software Chinese Academy of Sciences)
CodeMeta LearningContrastive LearningImageVideo
π― What it does: This paper proposes the Task Conflict Correction (TC2) method, which leverages a meta-learning framework to enhance the transfer performance of self-supervised learning models.
Exposing the Cracks: Vulnerabilities of Retrieval-Augmented LLM-based Machine Translation
Yanming Sun (University of Macau), Derek F. Wong (University of Macau)
CodeGenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: Investigate the robustness and reliability of retrieval-augmented large language model (REAL-MT) machine translation under noisy retrieval conditions, proposing a systematic noise synthesis framework and two novel evaluation metrics;
Expressive Temporal Specifications for Reward Monitoring
Omar Adalat (Imperial College London), Francesco Belardinelli (Imperial College London)
CodeReinforcement LearningTime SeriesSequential
π― What it does: This study proposes a method that utilizes quantified linear temporal logic (LTL_f[F]) to construct a reward monitor, which can generate dense numerical rewards for each state trajectory during runtime, thereby alleviating the sparsity problem caused by traditional Boolean rewards.
ExtendAttack: Attacking Servers of LRMs via Extending Reasoning
Zhenhao Zhu (Tsinghua University), Jiaheng Zhang (National University of Singapore)
CodeAdversarial AttackPrompt EngineeringText
π― What it does: For large language models (LLMs), a stealthy slowdown attack method called ExtendAttack is proposed, which embeds multi-base ASCII encoded characters into input prompts to force the model to perform extensive decoding computations before understanding the prompt, significantly increasing computational overhead and latency while maintaining answer accuracy.
F2SST: Frequency-to-Spatial Semantic Transfer for Few-Shot Image Classification
Xueyi Chen (Soochow University), Fanzhang Li (Soochow University)
CodeClassificationMeta LearningImage
π― What it does: Propose a framework named F2SST that leverages frequency domain information as implicit semantics to enhance few-shot image classification.
FaceShield: Explainable Face Anti-Spoofing with Multimodal Large Language Models
Hongyang Wang (Shijiazhuang Tiedao University), Xiaochun Cao (Sun Yat-sen University)
CodeClassificationAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: Proposed a multi-modal large language model called FaceShield specifically for Face Anti-Spoofing (FAS), capable of performing four tasks: coarse classification, fine classification, reasoning, and attack localization; simultaneously constructed two FAS instruction datasets, FaceShield-pre10K and FaceShield-sft45K.
CodeRecognitionObject TrackingTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideo
π― What it does: Propose a dynamic facial expression description task for videos, construct a high-quality annotated dataset with 5,033 entries and over 700,000 words, and perform instruction fine-tuning based on this dataset to develop the FaceTrack-MM model for better tracking and focus on the main character's facial expressions.
Facial-R1: Aligning Reasoning and Recognition for Facial Emotion Analysis
Jiulong Wu (Soochow University), Min Cao (Baidu Inc.)
CodeClassificationRecognitionData SynthesisLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodalityBenchmark
π― What it does: In the facial emotion analysis task, we propose a three-stage alignment framework called Facial-R1. It first performs few-shot instruction fine-tuning, then uses emotional factors (AU and emotion labels) as verifiable rewards for reinforcement learning, and finally iteratively expands the training set through self-supervised data synthesis, constructing a large-scale FEA-20K dataset.
FACTGUARD: Event-Centric and Commonsense-Guided Fake News Detection
Jing He (Yunnan University), Renyang Liu (National University of Singapore)
CodeClassificationExplainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
π― What it does: Propose the FACTGUARD framework, which leverages LLMs to extract core event content and combines common-sense reasoning for fake news detection.
π― What it does: Propose a Proximal Fill-in Minimization (PFM) framework based on graph neural networks, which jointly learns to minimize fill-in through a differentiable reordering layer and factorization-enhanced loss before LU decomposition of sparse matrices.
π― What it does: Propose FLKRβan unsupervised multi-agent code generation failure localization framework, and construct COFL, a fine-grained expert-annotated benchmark.
Failures to Surface Harmful Contents in Video Large Language Models
Yuxin Cao, Jin Song Dong (Csiro's Data61)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelVision Language ModelVideo
π― What it does: Analyze and reveal the blind spots of video large language models (LLMs) in detecting harmful content, and validate this vulnerability through constructing three no-query black-box attacks.
Paul GΓΆlz (Cornell University), Hannane Yaghoubizade (Cornell University)
CodeOptimization
π― What it does: The paper studies the fair division problem of allocating indivisible items among couples and small groups, proving that EF1 is achievable between two couples, but not necessarily for three or more couples.
Hadi Hosseini (Pennsylvania State University), Aditi Sethia (Indian Institute of Science)
CodeOptimization
π― What it does: Proposes two classes of algorithms to minimize the number of envious agents in housing allocation problems: one is a metric-agnostic FPT improvement framework for a given initial allocation, and the other is polynomial/linear time algorithms under single-peaked/single-concave preference domains; and experiments verify the impact of reallocation on fairness and welfare.
Fairness Aware Reinforcement Learning via Proximal Policy Optimization
Gabriele La Malfa (King's College London), Elizabeth Black (King's College London)
CodeReinforcement Learning
π― What it does: A fair reinforcement learning algorithm called Fair-PPO based on Proximal Policy Optimization (PPO) is proposed for multi-agent systems. It introduces two penalty terms (retrospective and prospective) based on group fairness metrics (e.g., demographic parity), encouraging agents to pursue reward maximization while satisfying fairness.
Faithful in Steps: Improving Generalization and Citation in RAG via Query Decomposition
Yue Liu (Hong Kong University of Science and Technology), Xiaofang Zhou (Hong Kong University of Science and Technology)
CodeRetrievalTransformerSupervised Fine-TuningVision Language ModelMultimodalityRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Propose a RAG framework called QDRAG based on query decomposition, which can achieve accurate citations and trustworthy answers in multi-hop, multi-modal question answering.
FAM: Fine-Grained Alignment Matters in Multimodal Embedding Learning with Large Vision-Language Models
Tianhang Xiang (South China University of Technology), Mingkui Tan (South China University of Technology)
CodeObject DetectionSegmentationRetrievalRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes the FAM (Fine-grained Alignment Matters) framework, transforming large-scale vision-language models (LVLM) into high-quality multimodal embedding models. The framework involves two-stage training: β Alignment phase: Multi-grained Alignment Contrastive (MAC) loss is used for coarse-to-fine feature alignment on image-text pairs; β‘ Adaptation phase: Vision Embedding Inversion (VEIN) performs masked reconstruction on visual features, encouraging embeddings to retain fine-grained visual information, followed by contrastive learning on downstream multimodal retrieval/localization tasks.
FaNe: Towards Fine-Grained Cross-Modal Contrast with False-Negative Reduction and Text-Conditioned Sparse Attention
Peng Zhang, Heng Kong (Shenzhen University)
CodeClassificationObject DetectionSegmentationTransformerVision Language ModelContrastive LearningMultimodalityBiomedical DataBenchmark
π― What it does: This study proposes the FaNe framework, addressing issues of pseudo-negative samples, coarse-grained alignment, and intra-modal discriminability in medical image vision-language pre-training through semantically enhanced positive sample mining, text-conditioned sparse attention pooling, and hard negative sample adaptive contrastive loss.
FantasyHSI: Video-Generation-Centric 4D Human Synthesis in Any Scene Through a Graph-Based Multi-Agent Framework
Lingzhou Mu (Tsinghua University), Kai Zhang (AMAP, Alibaba Group)
CodeData SynthesisGraph Neural NetworkReinforcement LearningAgentic AIVision Language ModelDiffusion modelVideo
π― What it does: Designed the FantasyHSI framework to achieve 4D human-computer interaction based on video generation, using a multi-agent system to generate long temporal sequences of animations with human behavior and physical consistency in arbitrary 3D scenes.
FARM: Frame-Accelerated Augmentation and Residual Mixture-of-Experts for Physics-Based High-Dynamic Humanoid Control
Jing Tan (Hong Kong University of Science and Technology), Renjing Xu (Guangdong University of Technology)
CodeRobotic IntelligenceTransformerReinforcement LearningMixture of ExpertsMultimodalitySequentialBenchmarkPhysics Related
π― What it does: Propose the FARM framework, integrating frame acceleration augmentation with residual Mixture-of-Experts (MoE), for unified control of high-dynamic humanoid motions.
π― What it does: Propose the FashionMAC framework, which directly generates extrapolated images from worn images using a deformation-free clothing center generation method to produce high-quality fashion display images;
Fast Conformal Prediction Using Conditional Interquantile Intervals
Naixin Guo (City University of Hong Kong), Zhixin Zhou (Alpha Benito Research)
CodeComputational EfficiencyTabularBiomedical Data
π― What it does: Propose Conformal Interquantile Regression (CIR) and its improved version CIR+, which directly construct prediction intervals satisfying coverage probability through black-box quantile regression, avoiding histogram binning in traditional methods while balancing coverage and interval width.
Fast Guaranteed Robust Local-Smooth Principal Component Separation
Mingdi Hu (Xi'an University of Posts and Telecommunications), Jiangjun Peng (Northwestern Polytechnical University)
CodeRestorationOptimizationImageVideo
π― What it does: Propose a new representative coefficient related total variation (RCCTV) regularizer that combines low-rank and local smoothness for robust principal component analysis;
Linjian Meng (Nanjing University), Yang Gao (Nanjing University)
CodeOptimizationReinforcement LearningBenchmark
π― What it does: In two-player zero-sum imperfect information games, Asymmetric PCFR+ (APCFR+) and its simplified version Simple APCFR+ (SAPCFR+) are proposed, enhancing the robustness of PCFR+ by introducing adaptive or fixed step size asymmetry between implicit and explicit cumulative counterfactual regret updates.
Faster Symmetry Breaking Constraints for Abstract Structures
ΓzgΓΌr AkgΓΌn (University of St Andrews), Christopher Jefferson (University of Dundee)
CodeOptimization
π― What it does: Proposed a symmetry breaking method for abstract structures, using representation-dependent ordering and delaying symmetry application to reduce constraint complexity.
π― What it does: Perform architecture-level pruning on the FLUX text-to-image generation model, proposing the FastFLUX framework. It replaces the residual branch of ResBlock with linear layers using Block-wise Replacement with Linear Layers (BRLL), and adopts Sandwich Training (ST) for local fine-tuning to maintain image quality while significantly improving inference speed.
π― What it does: Propose the FDC-Ground framework, using reinforcement learning to address GUI localization tasks, with improved reward design and pruning strategies;
π― What it does: Propose an unsupervised anomaly detection framework based on frequency-domain decomposition preprocessing (FDP), suppressing lesions through low-frequency reconstruction and preserving structural information via high-frequency components to achieve brain MRI anomaly detection;
Feature Attribution for Human Sensing with Radio Signals
Shuokang Huang (Imperial College London), Julie McCann (Imperial College London)
CodeRecognitionExplainability and InterpretabilityData-Centric LearningTime Series
π― What it does: Propose and implement MatryMask, a feature attribution method for WiFi CSI human perception tasks that can explain the model's attention to human-related features in wireless signals.
Feature Integration Spaces: Joint Training Reveals Dual Encoding in Neural Network Representations
Omar Claflin (Independent Researcher)
CodeRepresentation LearningAuto EncoderText
π― What it does: Investigate and propose a dual encoding space, employing jointly trained sparse autoencoders and neural factor machines (NFM) to simultaneously learn feature identity and feature integration effects, and validate their improvements in reconstruction and behavioral performance through intervention experiments.
Feature-Aware One-Shot Federated Learning via Hierarchical Token Sequences
Shudong Liu (Peking University), Guibo Luo (Peking University)
CodeFederated LearningKnowledge DistillationTransformerImageBiomedical Data
π― What it does: Proposed a one-round communication federated learning framework called FALCON, which generates multi-scale token sequences through hierarchical scale encoding and constructs a global model using a multi-scale autoregressive Transformer generator and knowledge distillation.
π― What it does: Proposed FedAdamW, an adaptive optimizer for large model training in federated learning, addressing issues of high variance in second-moment estimation, client drift caused by local overfitting, and slow convergence during reinitialization in each round.
Federated Graph-level Clustering Network with Attribute Inference
Renda Han (Hainan University), Jieren Cheng (Hainan University)
CodeFederated LearningGraph Neural NetworkGraph
π― What it does: Propose FedAI, integrating graph-level clustering and attribute inference to address knowledge drift caused by missing node attributes on clients.
Federated Vision-Language-Recommendation with Personalized Fusion
Zhiwei Li (University of Technology Sydney), Qiang Yang (Hong Kong Polytechnic University)
CodeRecommendation SystemFederated LearningTransformerMixture of ExpertsVision Language ModelMultimodality
π― What it does: Propose FedVLR under the federated learning framework to address the user personalization fusion problem in vision-language recommendation
FedGRPO: Privately Optimizing Foundation Models with Group-Relative Rewards from Domain Clients
Gongxi Zhu, Yuxing Han (Tsinghua University)
CodeOptimizationFederated LearningSafty and PrivacyTransformerLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: Propose the FedGRPO framework, converting large model optimization into a reward-based RL evaluation process, combining expert selection and group relative reward aggregation to achieve federated learning.
π― What it does: Propose the FedLAGC framework, achieving resource adaptation and communication efficiency in system heterogeneous federated learning through layer-adaptive submodel extraction and gradient correction.
FedMerge: Federated Model Merging for Personalization
Shutong Chen, Chengqi Zhang (University Of Technology Sydney)
CodeFederated LearningImageText
π― What it does: Propose the FedMerge framework, which uses the server side to perform weighted merging of multiple models, generating a single personalized model for each client, rather than having clients download/train multiple models.
FedPΒ²EFT: Federated Learning to Personalize PEFT for Multilingual LLMs
Royson Lee (Samsung AI Center), Timothy Hospedales (Samsung AI Center)
CodeFederated LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Propose FedP EFT, which jointly learns personalized parameter-efficient fine-tuning (LoRA) structures for multilingual LLMs in cross-device federated learning environments, achieving adaptive LoRA rank allocation for each client.
π― What it does: This paper proposes the FedPKDA framework, aiming to achieve privacy-preserving personalized federated learning and enhance the generalization ability of local models on heterogeneous data through dynamic knowledge alignment.
π― What it does: Designed and verified a federated learning algorithm named FedPM, which achieves global second-order optimization by using preconditioned mixed local parameters on the server side, thereby improving convergence speed and accuracy in heterogeneous data environments.
π― What it does: Propose FedRNC, a two-stage federated denoising framework, to address spatiotemporal label errors (STLM) in federated noisy incremental learning (FNCIL), by first coarsely filtering noise through local loss clustering, and then performing fine-grained label correction in the feature space using global prototypes.
FedSDA: Federated Stain Distribution Alignment for Non-IID Histopathological Image Classification
Cheng-Chang Tsai (Academia Sinica), Chun-Shien Lu (Academia Sinica)
CodeClassificationDomain AdaptationFederated LearningDiffusion modelAuto EncoderImageBiomedical Data
π― What it does: This paper proposes the FedSDA method, which addresses the feature distribution shift problem in non-IID histopathology image classification by aligning staining distributions across clients within a federated learning framework.
FedSkeleton: Secure Multi-Party Graph Skeleton Construction for Privacy-Preserving Federated Time-Series Forecasting
Henggang Deng (Tsinghua University), Tao Jiang (Huazhong University of Science and Technology)
CodeFederated LearningSafty and PrivacyGraph Neural NetworkGraphTime SeriesOrdinary Differential Equation
π― What it does: Propose the FedSkeleton framework, leveraging graph skeleton construction and dual-stream prediction for privacy-preserving federated time series forecasting
π― What it does: Propose the FedTopo framework, achieving cross-client feature alignment in non-I.I.D. federated learning through topological information.
FeTS: A Feature-Aware Framework for Time Series Forecasting
Le Wang (Shenzhen University), Songbai Liu (Shenzhen University)
CodeComputational EfficiencyTransformerTime Series
π― What it does: Propose the FeTS framework, which includes two major modules, AdaFE and DSFFN, to address the issue of uneven importance in time series prediction through adaptive feature extraction and dual-scale fusion;