FGD-Align: Pluralistic Alignment for Large Language Models via Fuzzy Group Decision-Making
Weihang Pan (Zhejiang University), Jieping Ye (Zhejiang University)
CodeReinforcement Learning from Human FeedbackLarge Language ModelText
🎯 What it does: Propose the FGD-Align framework, which utilizes fuzzy group decision theory to achieve multi-perspective alignment of large language models, integrating triangular fuzzy numbers, hierarchical aggregation, and probabilistic fuzzy DPO;
🎯 What it does: Propose the FGNet framework, which transfers the pre-trained Segment Anything 2 (SAM2) visual prior to three-dimensional electron microscopy (EM) neuron segmentation. The framework employs feature-guided attention (FGA) to guide the fine-grained encoder (FGE) in extracting details, and uses dual affinity decoders to generate refined segmentation results.
🎯 What it does: This paper proposes a non-reversible text-guided image editing framework called FIA-Edit, achieving high fidelity and semantic accuracy through frequency interaction attention;
CodeRecommendation SystemExplainability and InterpretabilityTabular
🎯 What it does: Proposed SPINRec, a stochastic path integral explanation method for recommendation systems, utilizing random baseline sampling to capture the effects of observed and unobserved interactions;
Filter, Correlate, Compress: Training-Free Token Reduction for MLLM Acceleration
Yuhang Han (Westlake University), Siteng Huang (Zhejiang University)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelImageVideoTextMultimodalityBenchmark
🎯 What it does: Proposed a three-stage 'Filter-Correlate-Compress' framework to reduce multimodal token counts in visual encoders and LLM decoders without training, significantly enhancing large model inference speed;
🎯 What it does: Proposed the Granular-ball One-Class Network (GBOC), which maps time-series data into an adaptive high-density particle ball space via Granular-ball Vector Data Description, combined with an LSTM encoder for unsupervised anomaly detection.
🎯 What it does: This paper designs and implements a hierarchical long video generation world model, which first learns coarse-grained prediction of large motions (Coarse DiT) and fine-grained prediction of continuous detailed motions (Fine DiT). Subsequently, it uses the fine-grained video stream as a self-supervised signal to distill the coarse-grained flow, ultimately achieving high-quality, temporally consistent driving scene video generation.
🎯 What it does: Proposed a framework named CFSG for fine-grained domain generalization, with the core idea of simultaneously decomposing the concept space and feature space into three subspaces: public, specific, and confusing, and dynamically adapting to varying degrees of domain shift by weighted fusion of the three during inference.
Fine-Grained Representation for Lane Topology Reasoning
Guoqing Xu (Chinese Academy of Sciences), Yang Yang (Chinese Academy of Sciences)
CodeAutonomous DrivingTransformerImage
🎯 What it does: Propose an end-to-end fine-grained lane topology reasoning framework called TopoFG, which leverages multi-scale BEV features, position priors, and sequence priors to achieve joint prediction of lane centerlines and topology relationships.
FinRpt: Dataset, Evaluation System and LLM-based Multi-agent Framework for Equity Research Report Generation
Song Jin (Renmin University of China), Rui Yan (Renmin University of China)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AITextBenchmarkFinance Related
🎯 What it does: Proposed the task of automatically generating Equity Research Reports (ERR), constructed the FinRpt benchmark dataset and evaluation system, and developed the multi-agent framework FinRpt-Gen for ERR generation.
First Learn, Then Review: Human-Like Continual Learning for Cross-View Geo-Localization with Limited Field of View
Lei Cheng (Southeast University), Teng Wang (Southeast University)
CodeRetrievalKnowledge DistillationConvolutional Neural NetworkMixture of ExpertsContrastive LearningImage
🎯 What it does: Proposed the HCL-Geo two-stage human-like continual learning framework, addressing the cross-view geolocation task under limited perspectives and unknown directions.
First-Order Error Matters: Accurate Compensation for Quantized Large Language Models
Xingyu Zheng (Beihang University), Xianglong Liu (ETH Zurich)
CodeCompressionTransformerTextBenchmark
🎯 What it does: Propose a new post-training quantization method called FOEM, which improves quantization error compensation by explicitly incorporating first-order gradient information, thereby enhancing the compression efficiency of large language models.
🎯 What it does: FLAG-4D proposes a dual-network framework that utilizes an Instantaneous Deformation Network (IDN) and a Global Motion Network (GMN) to perform fine-grained and globally consistent deformation of 3D Gaussian primitives over time, achieving high-quality 4D reconstruction.
FlashSVD: Memory-Efficient Inference with Streaming for Low-Rank Models
Zishan Shao (Duke University), Hai ¨Helen¨ Li
CodeComputational EfficiencyTransformerText
🎯 What it does: This paper proposes FlashSVD, an end-to-end, low-rank aware streaming inference framework for language models that have already been compressed using SVD, which can significantly reduce activation memory requirements without increasing computational costs.
🎯 What it does: Propose a two-stage text-to-video generation framework called FlashVideo, which first generates content and motion highly aligned with the text at low resolution using a large model, then refines details at high resolution using a lightweight model.
Xingbo Du (Mohamed bin Zayed University of Artificial Intelligence), Rui Zhang (Renmin University of China)
CodeClassificationExplainability and InterpretabilityRepresentation LearningLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a Flexible Concept Bottleneck Model (FCBM) that allows dynamic modification of the concept set during training and inference without retraining the entire network; it generates concept-to-label weights using a hypernetwork and implements sparse concept selection via a sparsemax with a learnable temperature; further demonstrating the model's zero-shot generalization on unseen concepts and rapid adaptation to new concepts with only one round of fine-tuning.
FloorPlanFormer: Multi-Task Transformer Network for Floor Plan Recognition with Outer-to-Inner Feature Refinement
Yun Liang (South China Agricultural University), Yishen Lin (South China Agricultural University)
CodeSegmentationTransformerImage
🎯 What it does: Propose a three-stage Transformer architecture called FloorPlanFormer, which utilizes multi-task learning to simultaneously identify outer contours, inner contours, and entrance doors;
Flora: Effortless Context Construction to Arbitrary Length and Scale
Tianxiang Chen (University of Science and Technology of China), Jieping Ye
CodeData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose the Flora strategy, which constructs arbitrary-length long contexts by concatenating short instructions without using LLMs or human intervention, and enhances the LLM's ability to process long texts through instruction fine-tuning.
🎯 What it does: The paper proposes a novel knowledge graph embedding model called FlorE, which combines the full Lorentz group with directional offset to address the Z-Paradox relation pattern.
🎯 What it does: This paper proposes the Flow-Induced Diagonal Gaussian Processes (FiDGP) framework, which projects neural network weight uncertainty into a low-dimensional induced subspace and combines regularized flow variational posterior to achieve model compression and high-quality uncertainty estimation, while supporting out-of-distribution (OoD) detection in a single forward pass.
🎯 What it does: Propose the FlowPath framework, which utilizes reversible neural flows to learn continuous, data-driven control paths to improve classification of irregularly sampled time series.
Focusing on Language: Revealing and Exploiting Language Attention Heads in Multilingual Large Language Models
Xin Liu (Institute of Information Engineering Chinese Academy of Sciences), Xiaoqi Jia (Institute of Information Engineering Chinese Academy of Sciences)
CodeExplainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes an efficient method called LAHIS to evaluate the importance of each head in multi-head self-attention during multilingual processing, and identifies language-specific and general-purpose attention heads.
🎯 What it does: Designed and implemented Foresight-Conditioned Diffusion (ForeDiffusion), a robotic manipulation strategy that leverages predicted future views combined with a dual loss, enabling the generation of more stable and accurate action sequences in long-horizon complex tasks.
Forest vs Tree: The (N, K) Trade-off in Reproducible ML Evaluation
Deepak Pandita (Rochester Institute of Technology), Christopher M Homan
CodeOptimizationData-Centric LearningText
🎯 What it does: The study evaluates the trade-off between the number of samples N and the number of annotations per sample K when assessing machine learning models under a fixed total annotation budget, and proposes how to allocate human annotation resources to achieve reliable evaluation.
Forgetting by Pruning: Data Deletion in Join Cardinality Estimation
Chaowei He (Soochow University), An Liu (Beijing Jiaotong University)
CodeComputational EfficiencyTabularBenchmark
🎯 What it does: This paper proposes CEP, a chi-square estimation forgetting framework for multi-table machine learning based on distribution-sensitive pruning and domain pruning, designed to quickly delete data under incomplete training scenarios.
Formal Abductive Latent Explanations for Prototype-Based Networks
Jules Soria (Université Paris-Saclay), Daniela Cancila (Université Paris-Saclay)
CodeClassificationExplainability and InterpretabilityImage
🎯 What it does: This paper proposes Abductive Latent Explanations (ALEs), a formal explainable method for constructing explanations in the latent space of case-based reasoning networks;
Format as a Prior: Quantifying and Analyzing Bias in LLMs for Heterogeneous Data
Jiacheng Liu (Wuhan University), Tieyun Qian (Wuhan University)
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelTextGraphTabularRetrieval-Augmented Generation
🎯 What it does: This paper systematically investigates the 'format bias' present in large language models when processing multi-format information (text, tables, infoboxes, knowledge graphs), and reveals its existence, driving factors, and internal mechanisms through a three-phase empirical analysis.
🎯 What it does: A full integer quantization framework is implemented for the PETR series of multi-view 3D detection models, significantly reducing inference latency and memory usage.
Fractured Glass, Failing Cameras: Simulating Physics-Based Adversarial Samples for Autonomous Driving Systems
Manav Prabhakar (University of Michigan), Arpan Kusari (University of Michigan)
CodeObject DetectionAutonomous DrivingAdversarial AttackImagePhysics Related
🎯 What it does: This study develops a physics-based adversarial sample generation method to simulate sensor failure caused by camera glass breakage and evaluate its impact on autonomous driving systems.
FreeAskWorld: An Interactive and Closed-Loop Simulator for Human-Centric Embodied AI
Yuhang Peng (Tsinghua University), Jiangtao Gong (Tsinghua University)
CodeAutonomous DrivingData-Centric LearningRobotic IntelligenceLarge Language ModelWorld ModelVideoTextMultimodalityBenchmark
🎯 What it does: Introduces the FreeAskWorld interactive closed-loop simulator and the Direction Inquiry Task, supporting high-level human-machine interaction and navigation;
🎯 What it does: Proposes an annotation-free Gaussian splatting method called FreeGaussian, which automatically locates interactive objects and achieves controllable view synthesis through differential analysis of optical flow and camera motion.
FreeInpaint: Tuning-free Prompt Alignment and Visual Rationality Enhancement in Image Inpainting
Chao Gong (Fudan University), Tao Mei (HiDream.ai Inc.)
CodeRestorationDiffusion modelImage
🎯 What it does: Propose FreeInpaint, a text-guided image inpainting method that requires no fine-tuning, directly enhancing restoration results during the inference stage by optimizing initial noise and intermediate latent variables.
🎯 What it does: Propose the FreqCycle framework, combining low-frequency cycle extraction with mid-to-high frequency enhancement to achieve efficient time series forecasting.
🎯 What it does: Proposed the FreqTAD model, which achieves dynamic graph anomaly detection by leveraging multi-scale frequency encoding and time-frequency attention.
🎯 What it does: Designed and implemented a frequency-domain aware vision-language multimodal generalization network (FVMGN) for cross-scenario, multimodal transfer in remote sensing image classification.
🎯 What it does: Proposes a full-rank efficient fine-tuning method named FRoD, which leverages hierarchical joint decomposition to extract a shared base and introduces sparse learnable rotation perturbations on this base, achieving efficient adaptation for large pre-trained models.
From Blind Transfer to Wise Selection: Prototype-Driven Neighbor-Domain Adaptation for Fake News Detection
Wayne Lu (Independent Researcher), Yiheng Li (University of International Business and Economics)
CodeClassificationDomain AdaptationTransformerMixture of ExpertsVision Language ModelMultimodality
🎯 What it does: This paper proposes the PANDA framework, which dynamically selects and fuses the most beneficial domain knowledge to address the negative transfer problem in multi-domain, multi-modal fake news detection.
CodeExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmark
🎯 What it does: Designed and implemented the LUCID framework, which automatically generates audit datasets using knowledge graphs to evaluate the machine forgetting effect of LLMs.
From Decision Trees to Boolean Logic: A Fast and Unified SHAP Algorithm
Alexander Nadel (Technion), Ron Wettenstein (Technion)
CodeExplainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: Proposed the WOODELF algorithm, which unifies the implementation of various feature importance metrics for decision tree ensemble models, including SHAP, Shapley interaction values, and Banzhaf values. It provides a pure Python implementation (using NumPy, SciPy, CuPy) and supports parallel computing on both CPU and GPU.
From Diagnosis to Generalization: A Cognitive Approach to Data Selection for Educational LLMs
Yuxiang Guo (University of Science and Technology of China), Shijin Wang (IFLYTEK Research)
CodeData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed the CASS framework, combining cognitive diagnosis, information selection, and hierarchical curriculum to efficiently select and fine-tune data subsets for educational LLMs.
From Dialogue to Destination: Geography-Aware Large Language Models with Multimodal Fusion for Conversational Recommendation
Yeming Li (University of Electronic Science and Technology of China), Yang Yang (University of Electronic Science and Technology of China)
CodeRecommendation SystemTransformerLarge Language ModelMultimodality
🎯 What it does: Propose a geo-aware dialogue recommendation framework called GeoCRS, which collaborates with a frozen LLM and external trainable modules to jointly generate multimodal geo-guided signals.
From Discriminative to Generative: A Diffusion-Based Paradigm for Multi-Agent Collaborative Perception
Kexin Gong (Beijing University of Posts and Telecommunications), Jinglin Li (Beijing University of Posts and Telecommunications)
CodeAutonomous DrivingDiffusion modelPoint Cloud
🎯 What it does: Proposes a two-stage generative supervised collaborative perception framework called DiGS-CP, which utilizes conditional diffusion models during the training phase to guide feature fusion, significantly enhancing perception performance while reducing communication overhead.
From Hypothesis to Premises: LLM-based Backward Logical Reasoning with Selective Symbolic Translation
Qingchuan Li, Tongxuan Liu (University Of Science And Technology Of China)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Proposes the Hypothesis-driven Backward Logical Reasoning (HBLR) framework, combining confidence-aware selective symbolic translation with hypothesis-driven backward reasoning to address redundancy and translation errors in LLM forward reasoning.
From IDs to Semantics: A Generative Framework for Cross-Domain Recommendation with Adaptive Semantic Tokenization
Peiyu Hu (Xi'an Jiaotong-Liverpool University), Jia Wang (Xi'an Jiaotong-Liverpool University)
CodeDomain AdaptationRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a cross-domain generative recommendation framework called GenCDR based on large language models, addressing the issues caused by traditional ID dependencies, such as item ID explosion and insufficient domain personalization.
From Macro to Micro: Probing Dataset Diversity in Language Model Fine-Tuning
Haoyu Li (Beijing Institute of Technology), Kun Liu (Beijing Institute of Technology)
CodeData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper systematically studies the impact of dataset diversity on model performance during the supervised fine-tuning phase of large language models, and classifies and experimentally verifies macro, meso, and micro diversity control strategies.
🎯 What it does: This paper proposes the ReACT method, which achieves controllable model fusion by performing linear correction in the model's final representation space;
🎯 What it does: Propose the BoxPromptIML framework, which utilizes coarse box prompts to generate pseudo masks via SAM and distills a lightweight student model, achieving weakly supervised image tampering localization.
From Pixels to Logic: A Perception-Reasoning Decomposition Framework for Open-World Referring Expression Comprehension
Lihong Huang (Shenzhen University), Yan Liu (Hong Kong Polytechnic University)
CodeSegmentationDepth EstimationTransformerVision Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Proposes a training-free open-world gesture expression understanding framework (PRDF), which separates visual perception from language reasoning. It first generates rich textual scene descriptions using open-source foundation models, then employs a language model to perform logical reasoning for target localization.
🎯 What it does: Propose a transferable adversarial attack method (TVA) based on a video foundation model (VFM) that does not require downstream task knowledge, training data, model queries, or parameters. It directly generates adversarial perturbations on the temporal representations of VFM to attack various downstream video models and multimodal large language models.
From Single to Societal: Analyzing Persona-Induced Bias in Multi-Agent Interactions
Jiayi Li (Peking University), Yansong Feng (Peking University)
CodeExplainability and InterpretabilityTransformerLarge Language ModelAgentic AIPrompt EngineeringText
🎯 What it does: This paper systematically investigates the biases arising from assigning individual personalities in multi-agent interactions of large language models (LLMs), exploring differences in social traits such as trustworthiness and persistence.
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Constructed the StimuliQA dataset containing real psychological scenarios and proposed the Psy-Interpreter bilateral reinforcement learning framework (Trajectory Cache + T-GRPO + Bilateral Reward), evaluated on multiple psychological reasoning benchmarks, and achieved continual learning based on self-labeling.
From Text to Simulation: A Multi-Agent LLM Workflow for Automated Chemical Process Design
Xufei Tian (East China University of Science and Technology), Ke Ye (East China University of Science and Technology)
CodeOptimizationTransformerLarge Language ModelAgentic AITextChain-of-Thought
🎯 What it does: Propose a workflow based on a multi-agent large language model (LLM) that automatically converts natural language chemical process descriptions into complete configuration files executable in professional simulation software, achieving closed-loop iterative optimization during this process.
FT-MoE: Sustainable-learning Mixture of Experts for Fault-Tolerant Computing
Wenjing Xiao (Guangxi University), Min Chen (South China University of Technology)
CodeAnomaly DetectionComputational EfficiencyMixture of ExpertsTime Series
🎯 What it does: Constructed the FT-MoE framework, implementing a dual-path hybrid expert network for edge fault detection and classification, and achieving continual learning through offline training + online fine-tuning.
Full-Atom Peptide Design via Riemannian–Euclidean Bayesian Flow Networks
Hao Qian, Lei Xu (Shanghai Jiao Tong University)
CodeGenerationFlow-based ModelBiomedical Data
🎯 What it does: Propose a full-atom peptide design framework based on Bayesian flow networks (PepBFN), jointly modeling amino acid types, residue orientations, centroid coordinates, and side-chain torsion angles to generate peptide chains in a fully continuous parameter space.
🎯 What it does: Proposes a Bayesian optimization framework (FFBO) for scenarios where both inputs and outputs are functions, constructing a probabilistic model through function-to-function Gaussian processes (FFGP);
🎯 What it does: For embedding communication in distributed deep learning recommendation models, we propose FUSEDREC: fusing multi-class embeddings, deduplication, delayed hashing, and recovery mechanisms to achieve single AlltoAll communication while preserving category information.
Junlin Xu (Wuhan University of Science and Technology), Yajie Meng (Wuhan Textile University)
CodeDrug DiscoveryConvolutional Neural NetworkGraph Neural NetworkTransformerLarge Language ModelMultimodalityBiomedical Data
🎯 What it does: Propose the FuseMine framework, which constructs a multi-modal complex-protein interaction prediction model by jointly encoding molecular structures and sequences through graph neural networks, convolutional networks, and pre-trained language models.
🎯 What it does: Proposed a lightweight visual backbone network called FVNet, which integrates the continuous-time dynamics of liquid neural networks into visual feature extraction to achieve adaptive spatiotemporal feature encoding.
GAGS: Granularity-Aware Feature Distillation for Language Gaussian Splatting
Yuning Peng (Wuhan University), Bisheng Yang (Wuhan University)
CodeSegmentationKnowledge DistillationRepresentation LearningVision Language ModelGaussian SplattingImage
🎯 What it does: Proposes the GAGS framework, distilling 2D CLIP features into a 3D Gaussian splat model to achieve out-of-the-box multi-view semantic queries;
🎯 What it does: This paper proposes GaussianImage++, a method that enhances image representation and compression performance through 2D Gaussian splatting.
GDBA Revisited: Unleashing the Power of Guided Local Search for Distributed Constraint Optimization
Yanchen Deng (Nanyang Technological University), Bo An (Nanyang Technological University)
CodeOptimizationGraphBenchmark
🎯 What it does: Proposed a new distributed guided local search framework called DGLS, addressing the reasons why the original GDBA algorithm performs poorly on general-value DCOPs, solving issues such as over-violation, infinite penalty accumulation, and uncoordinated updates.
CodeAnomaly DetectionVision Language ModelMultimodality
🎯 What it does: This paper addresses the outlier detection problem for graphical user interface (GUI) agents by proposing a Gaussian Mixture Model (GEM) method based on input embedding distance, which is used to identify instructions that fall outside the training distribution;
GEMA-Score: Granular Explainable Multi-Agent Scoring Framework for Radiology Report Evaluation
Zhenxuan Zhang (Imperial College London), Guang Yang (Imperial College London)
CodeExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAgentic AITextBiomedical DataComputed Tomography
🎯 What it does: Proposes Granular Explainable Multi-Agent Score (GEMA-Score) — a fine-grained medical report evaluation framework based on multi-agent collaboration, capable of objectively measuring pathological entities, locations, severity, and uncertainty, while providing interpretable comprehensive scores through subjective expression assessment (completeness, readability, terminology standardization);
Generalizable Drug–Target Interaction Prediction via ESM-2 Representations and Progressive Contrastive Curriculum Learning
Qianyang Wu (Hainan University), Feifei Cui (Hainan University)
CodeDrug DiscoveryTransformerLarge Language ModelContrastive LearningBiomedical Data
🎯 What it does: Proposed the ESP-DTI framework, combining the ESM-2 protein language model, CLIP-style cross-modal alignment, and progressive adaptive curriculum learning to predict drug-target interactions.
Generalized-Scale Object Counting with Gradual Query Aggregation
Jer Pelhan (University of Ljubljana), Matej Kristan (University of Ljubljana)
CodeObject DetectionTransformerImage
🎯 What it does: Propose an end-to-end few-shot counting and detection framework called GECO2, which utilizes scale-specific query encoders and cross-scale aggregation to generate high-resolution global query maps for accurate counting and localization.
Generalizing Vision-Language Models with Dedicated Prompt Guidance
Xinyao Li (University of Electronic Science and Technology of China), Jingjing Li (University of Electronic Science and Technology of China)
CodeDomain AdaptationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageText
🎯 What it does: Propose a two-step domain expert-guided domain generalization framework called GuiDG, first learning the source domain expert through prompt tuning, and then fine-tuning the visual encoder guided by cross-modal attention;
🎯 What it does: This paper generates risky samples using diffusion models and introduces category consistency constraints to ensure the generated samples align with the desired categories.
🎯 What it does: Propose a hierarchical autoregressive sketch generation process called Sketch-HARP, which allows flexible editing, deletion, or insertion of individual strokes during the drawing process, enabling fine-grained control over sketches.
CodeComputational EfficiencyReinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes GenPRM, a process reward model that enhances the reasoning quality of large language models through generative chain reasoning and code verification.
🎯 What it does: Created a video dataset named GenVidBench with a scale of 6.78 million videos, and conducted cross-source and cross-generator detection experiments on multiple video classification models.
GeoGen: A Two-stage Coarse-to-Fine Framework for Fine-grained Synthetic Location-based Social Network Trajectory Generation
Rongchao Xu (Florida State University), Guang Wang (Florida State University)
CodeData SynthesisSafty and PrivacyTransformerDiffusion modelTime SeriesSequential
🎯 What it does: Proposed the GeoGen two-stage coarse-to-fine framework for generating high-fidelity, privacy-safe fine-grained LBSN check-in trajectories.
Zhouhongyuan Hu (Sichuan University), Zhenbin Wang (Sichuan University)
CodeSegmentationDomain AdaptationConvolutional Neural NetworkGenerative Adversarial NetworkBiomedical Data
🎯 What it does: In the retinal fundus image segmentation task under source-free unsupervised domain adaptation (SF-UDA), the Geometric Correspondence Constrained (GCC) framework is proposed. It first stratifies pseudo-labels by entropy for quality assessment, then aligns low-quality samples using geometric correspondence information from high-quality samples, and further corrects high-confidence noise through adaptive Gaussian perturbation (SAPE).
Geometry Meets Light: Leveraging Geometric Priors for Universal Photometric Stereo Under Limited Multi-Illumination Cues
King-Man Tam (Institute of Science Tokyo), Rei Kawakami (National Institute of Informatics)
CodeDepth EstimationTransformerImage
🎯 What it does: Designed and implemented the GeoUniPS network by incorporating pre-trained 3D reconstruction models (e.g., VGGT) as geometric priors into a general photometric stereo framework, and constructed a synthetic dataset PS-Perp with perspective projection.
🎯 What it does: Propose a GAVIM framework that is imputation-free and based on variational autoencoders to complete incomplete multi-view clustering tasks.
GeoMoE: Divide-and-Conquer Motion Field Modeling with Mixture-of-Experts for Two-View Geometry
Jiajun Le (Wuhan University), Jiayi Ma (Wuhan University)
CodePose EstimationGraph Neural NetworkMixture of ExpertsImagePoint Cloud
🎯 What it does: Propose GeoMoE, which utilizes Mixture-of-Experts to perform probabilistic prior-driven decomposition and subfield linearization of two-view motion fields, thereby achieving more accurate and robust motion field estimation in tasks such as relative pose, homography, and point cloud registration.
GeoShield: Safeguarding Geolocation Privacy from Vision-Language Models via Adversarial Perturbations
Xinwei Liu (Institute of Information Engineering, Chinese Academy of Sciences), Xiaochun Cao (Northeastern University)
CodeSafty and PrivacyAdversarial AttackVision Language ModelImage
🎯 What it does: Studies how to prevent VLMs from accurately predicting location information by adding adversarial perturbations to images, thereby protecting geographic privacy.
CodePose EstimationTransformerSupervised Fine-TuningPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Propose the GeoX-Bench benchmark to evaluate the capabilities of large multimodal models in cross-perspective geolocation and pose estimation
🎯 What it does: Propose GEWDiff, a diffusion model based on wavelet encoding and geometric enhancement, achieving four times super-resolution for hyperspectral image reconstruction.
GeWu: A Culturally-Grounded Chinese Benchmark for Multi-Stage Social Bias Evaluation in Large Language Models
Yi Lin (Southern University of Science and Technology), Xuetao Wei (Lingnan University)
CodeTransformerLarge Language ModelTextBenchmark
🎯 What it does: Constructed the GeWu Chinese social bias assessment benchmark, containing 60,192 questions, and selected a high-bias subset of 1,000 questions named GeWu-1K.
Ghost in the Transformer: Detecting Model Reuse with Invariant Spectral Signatures
Suqing Wang (Wuhan University), Zuchao Li (Wuhan University)
CodeAnomaly DetectionTransformerText
🎯 What it does: Propose the GhostSpec method, utilizing the singular value spectrum of the attention weight matrix within Transformers as an immutable fingerprint of the model's origin, achieving lightweight, data-free verification for large language models (LLMs).
GHOST: Solving the Traveling Salesman Problem on Graphs of Convex Sets
Jingtao Tang (Simon Fraser University), Hang Ma (Simon Fraser University)
CodeOptimizationGraphBenchmark
🎯 What it does: Proposed and implemented the GHOST framework for optimally solving the Traveling Salesman Problem (GCS-TSP) on graphical convex sets (GCS), integrating combinatorial path search with convex trajectory optimization.
🎯 What it does: Propose the GIIM framework, which utilizes multi-heterogeneous graphs to simultaneously model intra-view interactions and inter-view dynamics, while providing robust handling for missing views.
GitTaskBench: A Benchmark for Code Agents Solving Real-World Tasks Through Code Repository Leveraging
Ziyi Ni (Institute of Automation, Chinese Academy of Sciences), Pin Lyu (Institute of Automation, Chinese Academy of Sciences)
CodeAI Code AssistantTransformerLarge Language ModelAgentic AITextMultimodalityBenchmark
🎯 What it does: Proposed GitTaskBench, a specialized benchmark to evaluate the ability of code agents to complete end-to-end real-world tasks using real GitHub repositories.
GLANCE: Global Actions in a Nutshell for Counterfactual Explainability
Loukas Kavouras (Information Management Systems Institute, Athena Research Center), Ioannis Emiris (National and Kapodistrian University of Athens)
CodeOptimizationExplainability and InterpretabilityTabularFinance Related
🎯 What it does: Propose a Global Explainable Adversarial Explanation (GCE) algorithm named GLANCE, which can generate efficient, low-cost, and interpretable action sets under a given threshold s;
GlitchCleaner: Lightweight Glitch Tokens Repairing by Lossless Gated LoRA in Large Language Models
Yibo Fan (Nankai University), Huan Li (Nankai University)
CodeAnomaly DetectionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Propose a lightweight, lossless method called GlitchCleaner, which automatically repairs glitch tokens by introducing a gated LoRA branch in the key MLP layers of large language models;
GlitchMiner: Mining Glitch Tokens in Large Language Models via Gradient-based Discrete Optimization
Zihui Wu (Xidian University), Shiguo Lian (China Unicom)
CodeAnomaly DetectionOptimizationTransformerLarge Language ModelText
🎯 What it does: Designed a behavior-driven gradient-guided local search framework called GlitchMiner to discover glitch tokens that cause abnormal behavior in large language models.
GLOBA: Rethinking Parameter Conflicts in Model Merging
Zehao Liu (Chinese Academy of Sciences), Wei Zhou (Chinese Academy of Sciences)
CodeRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Investigated parameter conflicts in multi-task model merging, analyzed the row-column space relationships of task vectors from a geometric perspective, and proposed the GLOBA framework to extract fully orthogonal parameters, classify overlapping parameters, and perform selective fusion based on different types.
🎯 What it does: Aiming at the dynamic cropping strategies in high-resolution vision-language models (HR-LVLM), this paper proposes GlobalCom 2, a zero-training overhead, plug-and-play global-local guided visual token compression framework, which can significantly compress the number of tokens while retaining most of the visual information.