CodeCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoBenchmark
π― What it does: A fast video large language model (FastVID) framework based on dynamic density pruning is proposed, focusing on spatiotemporal redundancy compression during the inference phase, significantly improving the inference speed of video LLMs and reducing computational costs.
π― What it does: The ODECE framework is proposed, which utilizes decision-focused learning to simultaneously predict constraint parameters in prediction-optimization problems, balancing feasibility and suboptimality through adjustable weights.
FEAT: Free energy Estimators with Adaptive Transport
Yuanqi Du (Cornell University), Eric Vanden-Eijnden (Courant Institute of Mathematical Sciences, New York University)
CodeOptimizationComputational EfficiencyReinforcement LearningTabularPhysics Related
π― What it does: This paper proposes the FEAT framework to estimate free energy differences through learning adaptive transport, utilizing the non-equilibrium Jarzynski equality and Crooks theorem.
π― What it does: The FedDISC framework is proposed, which combines federated learning with diffusion model-based modality recovery techniques to address the issue of missing modalities in multimodal emotion recognition.
π― What it does: This paper conducts large-scale experiments on various heterogeneous scenarios of federated learning, finding that existing methods lack robustness, and proposes the FedGPS framework, which integrates collaborative corrections of statistical distribution layers and gradient layers to enhance model performance.
π― What it does: Proposes the FedIGL framework, which identifies and separates irrelevant subgraphs shared by clients and client-specific subgraphs through invariant learning methods in federated graph learning, enhancing graph classification and clustering performance.
FedMGP: Personalized Federated Learning with Multi-Group Text-Visual Prompts
Weihao Bo (Nanjing University of Science and Technology), Zechao Li (Baidu VIS)
CodeFederated LearningSafty and PrivacyPrompt EngineeringVision Language ModelImageMultimodality
π― What it does: In the context of federated learning, a personalized adaptation method for visual-text models called FedMGP is proposed, which achieves the dual goals of local personalization and global generalization while maintaining privacy.
π― What it does: Proposes the FedQS framework, which simultaneously optimizes gradient aggregation and model aggregation in semi-asynchronous federated learning (SAFL).
π― What it does: The FedRAM framework is proposed in federated multi-task learning, achieving task and client weighted adaptation through a three-step training process (reference model, proxy model, proxy model), enhancing both local and global performance.
FedRTS: Federated Robust Pruning via Combinatorial Thompson Sampling
Hong Huang (City University of Hong Kong), Dapeng Wu (City University of Hong Kong)
CodeFederated LearningImageText
π― What it does: A robust pruning framework FedRTS based on combinatorial Thompson sampling is proposed in federated learning, which achieves adaptive adjustment of sparse network topology through the TSAdj module;
FedSVD: Adaptive Orthogonalization for Private Federated Learning with LoRA
Seanie Lee (KAIST), Sung Ju Hwang (KAIST)
CodeOptimizationFederated LearningSafty and PrivacyTransformerLarge Language ModelText
π― What it does: LoRA is introduced for pre-trained language models in federated learning, and the FedSVD method is designed to reparameterize BΒ·A through singular value decomposition (SVD) after each round of aggregation, thereby avoiding the amplification of DP-SGD noise in matrix multiplication while allowing the A matrix to adapt over time.
π― What it does: This paper proposes FedWMSAM, a federated learning framework that combines weighted momentum with Sharpness-Aware Minimization, aimed at addressing two major issues: local-global curvature mismatch and momentum echo oscillation in non-IID environments.
FEEDBACK FRICTION: LLMs Struggle to Fully Incorporate External Feedback
Dongwei Jiang (Johns Hopkins University), Daniel Khashabi (Johns Hopkins University)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelText
π― What it does: This paper systematically evaluates the limits of self-improvement of large models after receiving high-quality external feedback within a controlled experimental framework, and finds that even with nearly perfect feedback, the model still experiences 'feedback friction', making it difficult to fully absorb the feedback.
π― What it does: A dynamic guidance mechanism based on state feedback (FeedBack Guidance, FBG) is proposed, which adaptively adjusts the guidance scale during the inference process of the diffusion model based on the posterior estimation of the current trajectory, thereby improving sample quality while maintaining diversity.
π― What it does: This paper presents FerretNet, a lightweight synthetic image detection model that utilizes local pixel dependency (LPD) features to identify texture and edge anomalies in generative models.
Few-Shot Knowledge Distillation of LLMs With Counterfactual Explanations
Faisal Hamman (University of Maryland), Sanghamitra Dutta (University of Maryland)
CodeExplainability and InterpretabilityKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A few-shot task-aware knowledge distillation framework COD is proposed, which expands the training set by generating and utilizing minimally perturbed counterfactual explanations (CFE) for the teacher model, thereby helping the student model to more accurately approach the decision boundary of the teacher.
Few-Shot Learning from Gigapixel Images via Hierarchical Vision-Language Alignment and Modeling
Bryan Wong (KAIST), Mun Yong Yi (KAIST)
CodeClassificationGraph Neural NetworkPrompt EngineeringVision Language ModelImageMultimodalityBiomedical DataMagnetic Resonance Imaging
π― What it does: A hierarchical audiovisual multimodal multiple instance learning framework HiVE-MIL is proposed for few-shot whole slide image classification.
π― What it does: The FLYT algorithm is proposed, which evaluates the value of CLIP pre-training data by training a scoring model and utilizing downstream task gradients; M-FLYT is implemented to mix various scoring methods into a learnable linear model; a Soft Cap Sampling (SCS) strategy is designed to reduce the repetition of high-scoring samples and generate a high-quality filtered dataset.
Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods
Dennis Wei (IBM Research), Maria Chang (IBM Research)
CodeExplainability and InterpretabilityData-Centric LearningImageTextTabular
π― What it does: For scenarios where only the final model is available, a FiMO setting for training data attribution is proposed, which is reconstructed as a measurement of the model's sensitivity to training instances.
Finding and Reactivating Post-Trained LLMs' Hidden Safety Mechanisms
Mingjie Li (CISPA Helmholtz Center for Information Security), Yisen Wang (Peking University)
CodeOptimizationSafty and PrivacyRepresentation LearningAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningTextFinance Related
π― What it does: This paper investigates the reasons for the decline in safety performance of post-trained LLMs (especially large inference models) and proposes a lightweight method called SafeReAct to reactivate hidden safety mechanisms, thereby enhancing the model's safety under harmful prompts without significantly reducing its performance on specialized tasks.
Fine-grained List-wise Alignment for Generative Medication Recommendation
Chenxiao Fan (University of Science and Technology of China), Fuli Feng (University of Science and Technology of China)
CodeRecommendation SystemDrug DiscoveryTransformerLarge Language ModelReinforcement LearningBiomedical DataElectronic Health Records
π― What it does: A fine-grained list-based drug recommendation framework FLAME is constructed based on large language models, achieving safer and more accurate prescription generation through drug-by-drug decision-making and the integration of multi-source medical knowledge.
Fine-Tuning Discrete Diffusion Models with Policy Gradient Methods
Oussama Zekri (Institut Polytechnique de Paris), Nicolas Boulle
CodeGenerationOptimizationReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelTextSequentialBiomedical Data
π― What it does: This paper studies a fine-tuning method for discrete diffusion models, utilizing policy gradients from reinforcement learning to directly optimize non-differentiable rewards.
Finite Sample Analysis of Linear Temporal Difference Learning with Arbitrary Features
Zixuan Xie (University of Virginia), Shangtong Zhang (University of Virginia)
CodeReinforcement Learning
π― What it does: A finite sample convergence rate analysis of linear TD(Ξ») under arbitrary features is proposed, providing the convergence speed in L2 norm;
First SFT, Second RL, Third UPT: Continual Improving Multi-Modal LLM Reasoning via Unsupervised Post-Training
Lai Wei (Shanghai Jiao Tong University), Lichao Sun (Lehigh University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextMultimodality
π― What it does: This paper proposes an MM-UPT framework that achieves unsupervised post-training of multimodal large language models, enhancing reasoning capabilities through a self-reward mechanism based on majority voting within the GRPO algorithm.
Andrew Liu (Flagship Pioneering), Olivia Viessmann (Flagship Pioneering)
CodeProtein Structure PredictionTransformerReinforcement LearningGraphBiomedical Data
π― What it does: Improved IPA by rewriting it into a linear attention form compatible with FlashAttention, achieving linear memory and time complexity on GPUs, supporting structural modeling of thousands of residues.
Osayamen Jonathan Aimuyo (Cornell University), Rachee Singh (Cornell University)
CodeOptimizationComputational EfficiencyTransformerMixture of Experts
π― What it does: This paper proposes FlashMoE, a method that fully integrates distributed Mixture-of-Experts (MoE) operations into a single persistent GPU kernel, addressing the low utilization and high latency issues caused by CPU scheduling, synchronous AlltoAll communication, and frequent kernel launches in traditional implementations.
π― What it does: This paper systematically studies the loss landscape of neural networks and identifies and describes a special structure called 'channels to infinity', which is formed by at least two neurons whose output weights tend to infinity while their input weights approach equality.
π― What it does: This paper utilizes the grooksing training framework to separate the memory phase and the generalization phase of the network, systematically evaluating the occurrence timing and impact of neural collapse (NC) and the relative flatness (RF) of the loss landscape across different models.
Flatten Graphs as Sequences: Transformers are Scalable Graph Generators
Dexiong Chen (Max Planck Institute of Biochemistry), Karsten Borgwardt (Max Planck Institute of Biochemistry)
CodeGenerationData SynthesisDrug DiscoveryTransformerLarge Language ModelPoint CloudGraphSequential
π― What it does: This paper presents AUTOGRAPH, a Transformer-based autoregressive graph generation framework that transforms graphs into sequences through a reversible Segmented Eulerian Neighborhood Trail (SENT), allowing for direct graph generation using large language models.
FlexAC: Towards Flexible Control of Associative Reasoning in Multimodal Large Language Models
Shengming Yuan (University of Electronic Science and Technology of China), Lianli Gao (University of Electronic Science and Technology of China)
CodeLarge Language ModelPrompt EngineeringMultimodality
π― What it does: This paper proposes FlexAC, a method for controlling the intensity of multimodal large language model associative reasoning that is training-free and lightweight.
π― What it does: A two-stage MOF generation framework called MOFFLOW-2 is proposed, which first generates metal clusters and organic ligands using SMILES, and then predicts their 3D structures (translation, rotation, torsion angles, and lattice parameters) through a flow matching model.
CodeComputational EfficiencyKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: A flexible alignment framework is proposed, which includes training-time alignment (TrRa) and inference-time alignment (InRa), achieving efficient and controllable alignment for large language models.
π― What it does: Achieving the generation of 3D scenes with flexible viewpoints through the gradual expansion and optimization of 3D Gaussian splatting based on a single image, utilizing a video-to-video diffusion model (V2V) to complete new viewpoint synthesis and geometrically integrate the generated content into the scene;
Flick: Empowering Federated Learning with Commonsense Knowledge
Ran Zhu (Delft University of Technology), Qing Wang (Delft University of Technology)
CodeData SynthesisFederated LearningTransformerLarge Language ModelVision Language ModelDiffusion modelImageText
π― What it does: The Flick framework is designed and implemented, which generates synthetic data by combining server-side common knowledge from large language models (LLMs) with low-sensitivity cross-client local summaries to alleviate data heterogeneity (label imbalance and domain shift) in federated learning.
π― What it does: Proposes the FANTOM framework, which jointly infers multi-interval causal graphs and interval boundaries under multi-period non-Gaussian or heteroscedastic noise.
π― What it does: A flow equivariant recurrent neural network (FERNN) is proposed, enabling sequence models to maintain equivariance to time-parameterized symmetric transformations.
π― What it does: A neural process model based on flow matching (FlowNP) is proposed, which can directly generate or estimate the conditional distribution of any target point given context points;
π― What it does: A flow matching-based autonomous driving planning framework called Flow Planner is proposed, focusing on interactive behavior modeling.
Flow-GRPO: Training Flow Matching Models via Online RL
Jie Liu (Chinese University of Hong Kong), Wanli Ouyang (Chinese University of Hong Kong)
CodeGenerationReinforcement Learning from Human FeedbackReinforcement LearningFlow-based ModelImageTextStochastic Differential EquationOrdinary Differential Equation
π― What it does: The Flow-GRPO method is proposed, integrating online policy gradient reinforcement learning into the flow matching model to enhance performance in text-to-image tasks.
FlowCut: Rethinking Redundancy via Information Flow for Efficient Vision-Language Models
Jintao Tong (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
CodeOptimizationComputational EfficiencyTransformerVision Language ModelImageVideoMultimodality
π― What it does: Proposes the FlowCut framework, which cuts visual tokens of LVLM from the perspective of information flow, significantly reducing computational and memory burdens.
π― What it does: A generative data assimilation framework based on stochastic interpolation, FlowDAS, is proposed for estimating the state of PDE control systems under noise and sparse observations.
FlowFeat: Pixel-Dense Embedding of Motion Profiles
Nikita Araslanov (Technical University of Munich), Daniel Cremers (Technical University of Munich)
CodeSegmentationDepth EstimationOptical FlowVideo
π― What it does: A pixel-level high-resolution feature representation called FlowFeat is trained using optical flow and video data through a self-supervised approach.
π― What it does: We propose FLOWINGβa morphology transformation framework based on flow-based implicit neural representations (INR), which utilizes reversible flow models (NODE and NCF) to achieve structure-preserving, continuous, and reversible deformations between 2D images and 3D Gaussian splatting.
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
π― What it does: FlowMoE is proposed, a scalable pipeline scheduling framework for distributed Mixture-of-Experts (MoE) training; it unifies the scheduling of multiple types of tasks (MHA computation, gating, expert computation, A2A communication), and overlaps tensor block priority scheduling with all-reduce communication, while using Bayesian optimization to automatically tune block sizes.
FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts
Heming Zou (Tsinghua University), Xiangyang Ji (Tsinghua University)
CodeLarge Language ModelMixture of ExpertsText
π― What it does: This paper proposes FlyLoRA, an implicit MoE LoRA variant inspired by the olfactory circuits of moths, which uses frozen sparse random projections as built-in routers to achieve more efficient parameter tuning and multi-task integration.
FNOPE: Simulation-based inference on function spaces with Fourier Neural Operators
Guy Moss (University of TΓΌbingen), Cornelius SchrΓΆder (University of TΓΌbingen)
CodeFlow-based ModelTime Series
π― What it does: A simulation inference method based on the Fourier Neural Operator (FNO) called FNOPE is proposed for inferring the posterior distribution of function value parameters.
Focus-Then-Reuse: Fast Adaptation in Visual Perturbation Environments
Jiahui Wang (Nanjing University), Yang Yu (Nanjing University)
CodeDomain AdaptationRobotic IntelligenceReinforcement LearningVision Language ModelVideo
π― What it does: The Focus-Then-Reuse (FTR) method is proposed, which utilizes object-level filters to quickly deploy existing visual reinforcement learning strategies in visually disturbed environments.
π― What it does: A method is proposed to directly derive Riemannian metrics from pre-trained energy-based models (EBMs) to calculate the shortest paths between data points in high-dimensional space.
π― What it does: This study investigates how memorization occurs during the training process of Transformer language models by tracking the dynamics of memorization across different datasets and model sizes.
ForceFM: Enhancing Protein-Ligand Predictions through Force-Guided Flow Matching
Huanlei Guo (Southern University of Science and Technology), Bingyi Jing (Southern University of Science and Technology)
CodeDrug DiscoveryFlow-based ModelBiomedical Data
π― What it does: A molecular docking model based on force-guided flow matching, ForceFM, is proposed, which can use physical energy functions as a force field guide during the generation process to produce low-energy, physically reasonable ligand conformations.
π― What it does: A training-independent adaptive layer reuse method called Foresight is proposed for the Diffusion Transformer (DiT) model in text-to-video generation, which dynamically decides whether to reuse the outputs of each layer during inference to reduce redundant computations and accelerate generation.
ForgerySleuth: Empowering Multimodal Large Language Models for Image Manipulation Detection
Zhihao Sun (Fudan University), Yu-Gang Jiang (Fudan University)
CodeSegmentationAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
π― What it does: Proposes the ForgerySleuth framework, utilizing multimodal large language models and low-level trace encoders to achieve image tampering detection and interpretation.
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningTabularTime SeriesFinance Related
π― What it does: The SDForger framework is proposed, utilizing large language models to generate high-quality multivariate time series in low-sample environments.
π― What it does: This paper proposes the FORLA framework, which utilizes lightweight feature adapters and Slot Attention in federated learning to achieve unsupervised object-level representation learning, avoiding the drawbacks of data sharing and model sharing.
π― What it does: A new neural network called FAN (Fourier Analysis Networks) is proposed to address the shortcomings of existing neural networks in modeling and reasoning about periodic phenomena.
FP64 is All You Need: Rethinking Failure Modes in Physics-Informed Neural Networks
Chenhui Xu (University at Buffalo), Jinjun Xiong (University at Buffalo)
CodeOptimizationComputational EfficiencyPhysics Related
π― What it does: This study investigates the 'failure modes' of Physics-Informed Neural Networks (PINN) under low precision (FP32) training, addressing the issue by increasing the arithmetic precision to double precision (FP64) and proposing the Same-Basin hypothesis, which contradicts the traditional 'loss barrier' assumption.
FracFace: Breaking The Visual CluesβFractal-Based Privacy-Preserving Face Recognition
Wanying Dai (Sichuan University), Jin Song Dong (National University of Singapore)
CodeRecognitionSafty and PrivacyImage
π― What it does: A fractal-based frequency domain privacy-preserving face recognition framework, FracFace, is proposed, which can suppress reconstructable information while ensuring recognition accuracy.
π― What it does: Proposes Fractional Diffusion Bridge Models (FDBM), which approximates fractional Brownian motion (fBM) as a Markov process MA-fBM to generate noise for bridge creation, enabling diffusion bridge learning for paired and unpaired data.
π― What it does: A combination optimization method based on Fractional Lagrangian Dynamics (FLD) is proposed and implemented within a sampling and data-driven framework to address the issue of traditional LD struggling to escape narrow local optima.
FRAM: Frobenius-Regularized Assignment Matching with Mixed-Precision Computing
Binrui Shen (Beijing Normal University), Shengxin Zhu (Beijing Normal University)
CodeOptimizationImageGraph
π― What it does: This paper proposes a graph matching framework called FRAM based on Frobenius regularization linear assignment (FRA), which solves the quadratic assignment problem through continuous relaxation and controls the relaxation error with adjustable parameters.
π― What it does: This paper proposes FrameShield, which addresses the adversarial robustness of weakly supervised video anomaly detection (WSVAD) by employing frame-by-frame adversarial training with pseudo-labels and pseudo-anomaly generation, significantly enhancing the model's detection performance under attacks.
FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens
Yiming Zhong (ShanghaiTech University), Yuexin Ma (ShanghaiTech University)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningDiffusion modelMultimodality
π― What it does: The FreqPolicy proposes a frequency-domain visual-motor strategy that utilizes hierarchical frequency domain modeling and continuous tokens to achieve coarse-to-fine action generation.
π― What it does: A frequency-aware Token reduction method is proposed to reduce the computational complexity of self-attention in Vision Transformers while avoiding rank collapse issues caused by excessive smoothing.
π― What it does: A spectral reconstruction network based on fractal recursion (FRN) is proposed, which can recursively generate hyperspectral images from broadband to narrowband step by step.
From Bytes to Ideas: Language Modeling with Autoregressive U-Nets
Mathurin VIDEAU, David Lopez-Paz (Meta AI)
CodeTransformerLarge Language ModelText
π― What it does: Proposes a self-regressive U-Net architecture that directly processes raw bytes and dynamically aggregates them at multiple levels to form word-level and multi-word-level embeddings, eliminating traditional fixed tokenizers and large-scale embedding tables;
CodeExplainability and InterpretabilityAdversarial AttackTabular
π― What it does: This paper proposes a functional equivalence extraction attack (TRA) for decision trees and tree ensemble models using local optimal counterfactual explanations, and provides a theoretical analysis of its query complexity and competitive ratio.
π― What it does: A two-stage facial age transformation framework called Cradle2Cane is proposed, utilizing a few-step text-image diffusion model to achieve high-fidelity facial aging across the entire life cycle.
π― What it does: This study investigates the anti-tampering capability of machine unlearning in visual classification models, finding that existing methods can restore the performance of forgotten samples by fine-tuning only on the retained set after refinement, and proposes a new unlearning method based on weight space regularization.
From Euler to AI: Unifying Formulas for Mathematical Constants
Tomer Raz (Technion Israel Institute of Technology), Ido Kaminer (Technion Israel Institute of Technology)
CodeOptimizationTransformerLarge Language ModelPrompt EngineeringTextPhysics Related
π― What it does: A complete automation framework has been developed to extract, verify, and unify infinite series, continued fractions, and other formulas of Ο and other constants using large language models, constructing a unified Conservative Matrix Field (CMF) structure, and proving the equivalence of formulas through the UMAPS algorithm.
π― What it does: This paper proposes the Policy World Model (PWM), a unified driving world model that can perform action-independent video prediction of future states and directly plan trajectories based on the predicted future states.
From Indicators to Insights: Diversity-Optimized for Medical Series-Text Decoding via LLMs
Xiyuan Jin (Beijing Jiaotong University), Youfang Lin (Beijing Jiaotong University)
CodeClassificationRecognitionOptimizationTransformerLarge Language ModelPrompt EngineeringTextTime SeriesBiomedical DataElectrocardiogram
π― What it does: A multimodal text and time series joint decoding framework based on clinical decision indicators, InDiGO, is proposed, utilizing LLM for efficient analysis of medical time series.
From Pose to Muscle: Multimodal Learning for Piano Hand Muscle Electromyography
RUOFAN LIU, Hideki Koike (Tokyo Institute of Technology)
CodeRecognitionPose EstimationTransformerContrastive LearningMultimodalityBiomedical Data
π― What it does: This paper presents PianoKPM Netβa multimodal network that infers hand electromyography (EMG) from hand posture and piano key strike actions, and releases the largest professional pianist EMG dataset, the PianoKPM Dataset.
From Sequence to Structure: Uncovering Substructure Reasoning in Transformers
Xinnan Dai (Michigan State University), Jiliang Tang (Michigan State University)
CodeTransformerLarge Language ModelPrompt EngineeringTextGraph
π― What it does: This paper studies how to enable a decoder-only Transformer to extract substructures in a graph structure through text sequences, and proposes the Induced Substructure Filtration (ISF) to explain the mechanism of hierarchical identification of substructures.
π― What it does: This paper proposes that the Transformer can implement various denoising algorithms during forward inference, including manifold-based Laplacian denoising, precise score-based diffusion denoising, and learnable anisotropic diffusion;
π― What it does: A few-shot evolutionary optimization framework (FSEO) is proposed, combining meta-learning deep kernel Gaussian processes (MDKL) as a regression surrogate to address expensive multi-objective optimization problems with a limited number of samples.
FSNet: Feasibility-Seeking Neural Network for Constrained Optimization with Guarantees
Hoang T. Nguyen (Massachusetts Institute of Technology), Priya L. Donti (Massachusetts Institute of Technology)
CodeOptimization
π― What it does: The FSNet framework is proposed, which directly embeds the feasibility finding step into the training and inference process of neural networks to solve constrained parametric optimization problems and ensure that the output meets the constraints.
π― What it does: A fully spiking neural network called SpikeFET is proposed, specifically designed for unified frame-event visual object tracking, integrating convolutional local feature extraction with Transformer global modeling under the spiking paradigm. It also designs a Random Patch Module (RPM) and Spatiotemporal Regularization (STR) to enhance localization accuracy and energy efficiency.
Panqi Chen (Zhejiang University), Shikai Fang (Microsoft Research Asia)
CodeTime SeriesOrdinary Differential Equation
π― What it does: A functionally adaptive complexity tensor decomposition model CATTE is proposed, which utilizes Fourier feature encoding for continuous indexing and learns factor trajectories through neural ODE, while applying a sparse prior on functional factor trajectories to achieve automatic tensor rank inference.
π― What it does: A training-free multi-model feature fusion method called Fuse2Match is designed, utilizing Stable Diffusion 3 (SD3), the older version of Stable Diffusion (SD), and the contrastive learning model DINO to improve zero-shot semantic correspondence tasks.
π― What it does: The SeerDrive framework is proposed, which achieves end-to-end adaptive driving decision-making by predicting future Bird's Eye View (BEV) scenarios and interacting in a closed-loop with trajectory planning.
CodeAutonomous DrivingTransformerLarge Language ModelVision Language ModelImageMultimodalityChain-of-Thought
π― What it does: The FSDrive framework is proposed, allowing VLA models to think about the future and plan trajectories at the image level through visualized spatiotemporal Chain-of-Thought.
G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems
Guibin Zhang (National University of Singapore), Shuicheng YAN
CodeGraph Neural NetworkLarge Language ModelAgentic AITextGraphBenchmark
π― What it does: This paper proposes G-Memory, a three-layer graph structure (insight graph, query graph, interaction graph) for recording, retrieving, and updating the long-term collaboration history of multi-agent systems (MAS), supporting the self-evolution of agent teams.
G-Net: A Provably Easy Construction of High-Accuracy Random Binary Neural Networks
Alireza Aghasi (Oregon State University), Wyatt D. Whiting
CodeClassificationRecognitionOptimizationImage
π― What it does: A G-Net framework based on random binary embedding is proposed, which can directly map the trained floating-point neural network to a high-dimensional binary network while maintaining close prediction accuracy; it also provides theoretical guarantees and large-scale experimental validation.
Gains: Fine-grained Federated Domain Adaptation in Open Set
Zhengyi Zhong (National University of Defense Technology), Ju Ren (Tsinghua University)
CodeDomain AdaptationFederated LearningImage
π― What it does: Proposes the Gains framework to achieve fine-grained knowledge discovery and rapid balance of knowledge adaptation in open-set federated learning;
GASP: Efficient Black-Box Generation of Adversarial Suffixes for Jailbreaking LLMs
Advik Raj Basani (Birla Institute of Technology and Science), Xiao Zhang (CISPA Helmholtz Center for Information Security)
CodeOptimizationAdversarial AttackTransformerLarge Language ModelText
π― What it does: Proposes the GASP framework, which efficiently generates readable malicious suffixes that can bypass security shields in a black-box environment, significantly improving the jailbreak success rate of LLMs.
Gated Integration of Low-Rank Adaptation for Continual Learning of Large Language Models
Yan-Shuo Liang (Nanjing University), Wu-Jun Li (Nanjing University)
CodeTransformerLarge Language ModelText
π― What it does: The GainLoRA method is proposed, which integrates low-rank adaptation branches with a gating mechanism in continual learning to reduce forgetting.
π― What it does: A Contour-Guided Continuous Gaussian Field framework (GauSAM) based on SAM has been developed, capable of generating high-quality medical image segmentation masks at any resolution.
π― What it does: This paper proposes a Gaussian Herder Across Pens (GHAP) framework based on optimal transport, which first compresses the geometric information of the 3D Gaussian Splatting (3DGS) model using Gaussian Mixture Reduction (GMR), and then fine-tunes the color and opacity, achieving nearly lossless rendering quality while retaining only 10% of the Gaussian primitives.