Conference on Neural Information Processing Systems Β· 2283 papers
Gaussian Regression-Driven Tensorized Incomplete Multi-View Clustering with Dual Manifold Regularization
Zhenhao Zhong (Hebei Normal University), Ruiqiang Guo (Hebei Normal University)
CodeGaussian SplattingMultimodality
π― What it does: A tensor-based incomplete multi-view clustering framework GUITAR is proposed, which is based on Gaussian regression norm, improved βΞ΄ norm, and double manifold regularization.
GaussianFusion: Gaussian-Based Multi-Sensor Fusion for End-to-End Autonomous Driving
Shuai Liu (Sun Yat-sen University), Kai Huang (Sun Yat-sen University)
CodeAutonomous DrivingOptimizationExplainability and InterpretabilityComputational EfficiencyGaussian SplattingPoint CloudBenchmark
π― What it does: A multi-sensor fusion framework called GaussianFusion based on 2D Gaussian distribution is proposed for perception and path planning in end-to-end autonomous driving.
Gaze-VLM: Bridging Gaze and VLMs through Attention Regularization for Egocentric Understanding
Anupam Pani (Hong Kong University), Yanchao Yang (Hong Kong University)
CodeRecognitionGenerationTransformerVision Language ModelOptical FlowVideoMultimodality
π― What it does: This paper proposes using human gaze as an attention regularization signal during the training phase of VLM to enhance activity understanding and future action prediction in first-person videos.
GEM: Empowering MLLM for Grounded ECG Understanding with Time Series and Images
Xiang Lan (National University of Singapore), Mengling Feng (National University of Singapore)
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodalityTime SeriesBiomedical DataElectrocardiogramBenchmark
π― What it does: Proposes GEM, the first multimodal large language model that combines ECG time series, 12-lead images, and text for evidence-based electrocardiogram interpretation.
π― What it does: This paper studies the generalization mechanism of Kolmogorov-Arnold Networks (KAN), defining Lipschitz complexity for the first time as a measure of structural complexity of KAN, and deriving a generalization upper bound based on this; subsequently, it proposes the LipKAN architecture, which inserts Lip layers between each activation layer and employs L1^5 regularization, significantly reducing Lipschitz complexity and thereby enhancing the model's generalization performance.
π― What it does: This study investigates the generalization error of Selective State Space Models (Selective SSM) in sequence modeling and provides a generalization bound based on covering numbers.
π― What it does: This paper constructs a theoretical framework that proves when the scoring function of branch-and-cut (B&C) decisions has a piecewise polynomial structure, the overall performance metrics (such as tree size) are piecewise constant with respect to the parameters, and provides upper bounds for pseudo-dimension and sample complexity.
Generalized and Invariant Single-Neuron In-Vivo Activity Representation Learning
Wei Wu (Peking University), Jinzhuo Wang (Peking University)
CodeRepresentation LearningAdversarial AttackTransformerAuto EncoderContrastive LearningBiomedical Data
π― What it does: This paper proposes a general adversarial training framework to eliminate batch effects in single-cell activity characterization, thereby enhancing the generalization ability of single-cell characterization models under different animals and stimulation conditions.
π― What it does: This paper proposes a non-Euclidean gradient norm clipping method that combines conditional gradient and steepest descent, and proves its descent property under (L0, L1)-smoothness.
Generalized Top-k Mallows Model for Ranked Choices
Shahrzad Haddadan (Rutgers Business School), Sara Ahmadian (Google)
CodeRecommendation SystemOptimizationTabular
π― What it does: A weighted TopKGMM (Generalized Top-k Mallows Model) is proposed, along with three efficient algorithms: Profile-Based Repeated Insertion Sampling (PRIM), a dynamic programming method for calculating selection probabilities called DYPCHIP, and an active learning center approach named BUCCHOI.
Generalizing Experience for Language Agents with Hierarchical MetaFlows
Shengda Fan (Renmin University of China), Yankai Lin (Renmin University of China)
CodeComputational EfficiencyMeta LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This paper proposes the MetaFlowLLM framework, which constructs an experience hierarchy tree to enable large language model agents to reuse experiences in multi-step tasks through MetaFlow (including static steps and dynamic subtasks), significantly improving task success rates and execution efficiency.
Generalizing Single-Frame Supervision to Event-Level Understanding for Video Anomaly Detection
Junxi Chen (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
CodeAnomaly DetectionTransformerVideo
π― What it does: This paper proposes a Single Frame Supervised Video Anomaly Detection (SF-VAD) paradigm and designs a Frame-guided Progressive Learning (FPL) framework, utilizing only one frame annotation per anomalous video to achieve event-level anomaly understanding.
π― What it does: A new class of linear Generalized Bradley-Terry models (Linear GBT with Diffusion Prior) is proposed, which retains comparative data while utilizing embedding information to generalize to uncomparable objects, and provides monotonicity guarantees under specific embeddings (such as diffusion embedding and one-hot encoding).
π― What it does: This paper proposes a proof format APTP and a lightweight proof checker APTPchecker that are independent of existing DNN verification tools, capable of independently verifying UNSAT proofs provided by DNN verification tools and achieving scalable checking on large-scale models.
Generating Computational Cognitive models using Large Language Models
Milena Rmus (Helmholtz Munich), Eric Schulz (Helmholtz Munich)
CodeOptimizationTransformerLarge Language ModelText
π― What it does: Developed the GeCCo pipeline, utilizing open-source LLMs to generate cognitive models and optimize them through iterative feedback, applied in four cognitive domains: decision-making, learning, planning, and working memory;
π― What it does: Proposes a method that uses loss information from a pre-trained model to guide a score-based generative model in generating high-loss samples, thereby performing risk-averse CVaR optimization in downstream tasks;
π― What it does: This paper proposes GenSCO, a framework that views the generation process of diffusion models as a search step, achieving efficient solution through iterative perturbation-enhancement-post-processing for combinatorial optimization.
π― What it does: A three-stage generative data augmentation framework called DAR-GDA is proposed, which first compresses a multi-step diffusion model into a single-step generator using score distillation, then aligns the real distribution through adversarial training, and finally performs importance reweighting using the probability output of the discriminator.
π― What it does: Designed and evaluated a fully Transformer-based, message-passing-free graph pre-training framework G2PM, which utilizes substructure sequences generated by random walks for masked substructure reconstruction to learn graph representations.
π― What it does: This paper studies generative model inversion from a geometric perspective, analyzes the mechanism of gradient projection onto the generator manifold, and proposes gradient-manifold alignment metrics, gradient alignment training objectives, and the non-training AlignMI method.
GenIR: Generative Visual Feedback for Mental Image Retrieval
Diji Yang (University of California Santa Cruz), James Davis (University of California Santa Cruz)
CodeGenerationRetrievalVision Language ModelDiffusion modelImage
π― What it does: Designed and implemented an interactive visual retrieval framework GenIR, which utilizes a text-to-image diffusion generator to create synthetic images as visual feedback for users, thereby helping them approach their target images in multi-round retrieval; simultaneously defined the Mental Image Retrieval (MIR) task and proposed an automated pipeline for constructing multi-round datasets.
Geo-Sign: Hyperbolic Contrastive Regularisation for Geometrically Aware Sign Language Translation
Edward Fish (University of Surrey), Richard Bowden (University of Surrey)
CodeImage TranslationPose EstimationGraph Neural NetworkLarge Language ModelContrastive LearningVideoText
π― What it does: Proposes the Geo-Sign framework, using hyperbolic geometry regularization for skeletal representation to enhance sign language translation quality.
π― What it does: A remote sensing foundational model named GeoLink has been constructed, which directly enhances the image encoder using OpenStreetMap (OSM) vector data, achieving multimodal fusion in pre-training and downstream tasks.
Geometric Mixture Models for Electrolyte Conductivity Prediction
Anyi Li (Renmin University of China), Wenbing Huang (Renmin University of China)
CodeGraph Neural NetworkGraphTabular
π― What it does: The GeoMix framework is proposed, utilizing Set-SE(3) equivalence and geometric graph representation to predict the conductivity of electrolyte systems, and achieving fine-grained message passing of inter-molecular geometric information through the GIN module.
π― What it does: A lightweight multi-solution optimization framework GACβMSO based on gradient flow and geometric structure is proposed for efficiently fine-tuning large-scale foundational models with parameter efficiency, generating diverse and collaborative solution sets.
π― What it does: Two edge contraction graph pooling layers based on graph size (Magnitude) or spread (Spread) are proposed (MagEdgePool and SpreadEdgePool).
GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization
Pengyue Jia (City University of Hong Kong), Sharon Li (University of Wisconsin-Madison)
CodeRetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes GeoRanker, a distance-aware ranking framework based on a large visual-language model, designed to select locations from a candidate set that are closest to the geographic location of a query image.
π― What it does: This paper proposes GeoSVR, an explicit surface reconstruction framework based on sparse voxels, which achieves high-precision, complete, and efficient geometric reconstruction by utilizing voxel uncertainty depth constraints and voxel dropout regularization.
GLID$^2$E: A Gradient-Free Lightweight Fine-tune Approach for Discrete Biological Sequence Design
Hanqun Cao (Chinese University of Hong Kong), Pheng-Ann Heng (Chinese University of Hong Kong)
CodeGenerationOptimizationReinforcement LearningDiffusion modelBiomedical Data
π― What it does: A lightweight reinforcement learning framework GLID E is proposed for fine-tuning pre-trained discrete diffusion models to generate DNA and protein sequences with target functions.
Global Minimizers of $\ell^p$-Regularized Objectives Yield the Sparsest ReLU Neural Networks
Julia B Nakhleh, Robert D Nowak
CodeOptimizationTabular
π― What it does: This paper proposes and proves a method for βp path norm regularization based on 0 < p < 1, which can directly obtain the sparsest interpolation solution for single hidden layer ReLU networks through gradient descent.
Globally Optimal Policy Gradient Algorithms for Reinforcement Learning with PID Control Policies
Vipul Kumar Sharma, S Sivaranjani
CodeOptimizationReinforcement LearningTime Series
π― What it does: This paper proposes a global optimal optimization framework that combines the policy gradient method in reinforcement learning with the parameterization of PID controllers, providing the gradient expression for the PID control problem and designing both model-based and model-free policy gradient algorithms based on this.
Glocal Information Bottleneck for Time Series Imputation
Jie Yang (University of Illinois Chicago), Kaize Ding (Northwestern University)
CodeTransformerTime Series
π― What it does: A new training paradigm for missing value imputation in time series, Glocal-IB, is proposed, which incorporates global alignment loss into the standard information bottleneck framework to address the issues of model overfitting to local noise and inability to capture global structure under high missing rates.
GLSim: Detecting Object Hallucinations in LVLMs via Global-Local Similarity
Seongheon Park (University of Wisconsin Madison), Sharon Li (University of Wisconsin Madison)
CodeRecognitionObject DetectionTransformerVision Language ModelImageMultimodality
π― What it does: The GLSIM framework is proposed, which uses global and local embedding similarity within the model to detect object hallucinations in large visual-language models.
π― What it does: A unified and efficient graph multi-view learning framework GMV is proposed, which enhances the generalization and robustness of GNN/GT in graph classification tasks by utilizing structure-enhanced subgraph sampling and mixing, multi-view decomposition, and dual-head prediction.
GnnXemplar: Exemplars to Explanations - Natural Language Rules for Global GNN Interpretability
Burouj Armgaan (Indian Institute of Technology Delhi), Sayan Ranu (Indian Institute of Technology Delhi)
CodeExplainability and InterpretabilityGraph Neural NetworkLarge Language ModelGraph
π― What it does: This paper proposes GNNXEMPLAR, a global explanation framework based on exemplars, which utilizes natural language rules to explain the predictions of GNN in node classification tasks.
Go With the Flow: Fast Diffusion for Gaussian Mixture Models
George Rapakoulias (Georgia Institute of Technology), Panagiotis Tsiotras (Georgia Institute of Technology)
CodeGenerationData SynthesisOptimizationDiffusion modelAuto EncoderImageBiomedical Data
π― What it does: This paper proposes a training-free, low-complexity analytical parameter method that decomposes the SchrΓΆdinger bridge problem into a series of Gaussian bridge subproblems and solves the mixed strategy using linear programming, thereby achieving distribution transfer from one Gaussian mixture model to another.
π― What it does: This paper proposes and experiments with a self-gated activation function called GoLU based on the Gompertz function, which can reduce feature variance and smooth the loss surface through right-skewed asymmetrical gating.
π― What it does: Proposes the GoRA framework, which utilizes gradient information to dynamically allocate the rank of LoRA before training and provides non-zero initialization for low-rank adapters;
GPAS: Accelerating Convergence of LLM Pretraining via Gradient-Preserving Activation Scaling
Tianhao Chen (Hong Kong University of Science and Technology), Can Yang (Hong Kong University of Science and Technology)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes Gradient-Preserving Activation Scaling (GPAS), which accelerates pre-training convergence by applying learnable scaling to intermediate activations in Pre-LN Transformers while maintaining gradient magnitude.
π― What it does: A two-stage low-bit quantization framework called GPLQ has been developed, which first performs a single round of Quantization-Aware Training (QAT) on the activations of the Vision Transformer and then applies Post-Training Quantization (PTQ) to the weights, achieving a high-precision model with 4-bit numerical accuracy.
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
π― What it does: This paper proposes the GPO (Guided Pivotal Optimization) strategy, which enhances multi-step reasoning performance by identifying and focusing on critical steps in the reasoning trajectory generated by LLMs for fine-grained model tuning.
π― What it does: This paper proposes GPSToken, a method that utilizes two-dimensional Gaussian parameterization to achieve non-uniform, spatially adaptive image segmentation, which is used for image representation and generation.
Gradient-Weight Alignment as a Train-Time Proxy for Generalization in Classification Tasks
Florian A. HΓΆlzl (Institute for Artificial Intelligence in Medicine Technical University of Munich), Georgios Kaissis (Institute for Artificial Intelligence in Medicine Technical University of Munich)
CodeClassificationExplainability and InterpretabilityTransformerImage
π― What it does: A gradient-weight alignment (GWA) metric is proposed to evaluate model generalization during training and identify important training samples.
Grammars of Formal Uncertainty: When to Trust LLMs in Automated Reasoning Tasks
Debargha Ganguly (Case Western Reserve University), Vipin Chaudhary (Case Western Reserve University)
CodeTransformerLarge Language ModelText
π― What it does: This study investigates the implicit uncertainty of large language models in generating formal reasoning (SMT-LIB), constructs a uncertainty quantification framework based on Probabilistic Context-Free Grammar (PCFG), and achieves selective verification through lightweight signal fusion, significantly reducing the error rate.
Graph Data Selection for Domain Adaptation: A Model-Free Approach
Ting-Wei Li, Hanghang Tong
CodeDomain AdaptationGraph Neural NetworkGraph
π― What it does: A model-free graph data selection framework called GRADATE is proposed, which selects the most beneficial training graph samples from the source domain for the target domain using Graph Data Distribution Distance (GDD);
π― What it does: This paper proposes GRIDDD, a discrete graph diffusion probability model that supports dynamic insertion and deletion of nodes during the diffusion process for variable-sized molecular generation.
π― What it does: The researchers proposed a new topological descriptor called SpectRe, which combines spectral information with RePHINE for graph representation learning.
π― What it does: The GraphKeeper framework is proposed to address the problem of catastrophic forgetting in incremental learning for graphs (Domain-IL), maintaining performance on previous domains while continuously adding new graph domains.
GraphMaster: Automated Graph Synthesis via LLM Agents in Data-Limited Environments
Enjun Du (Beijing Institute of Technology), Guoren Wang (Beijing Institute of Technology)
CodeGenerationData SynthesisGraph Neural NetworkLarge Language ModelAgentic AITextGraphRetrieval-Augmented Generation
π― What it does: A multi-agent framework called GraphMaster is proposed, which utilizes LLM to generate semantically rich and structurally consistent text attribute graphs.
Graphs Help Graphs: Multi-Agent Graph Socialized Learning
Jialu Li (Tianjin University), Qinghua Hu (Tianjin University)
CodeGraph Neural NetworkPrompt EngineeringGraph
π― What it does: This paper proposes the Graph Socialized Learning (GSL) framework and its implementation method, Graphs Help Graphs (GHG), to achieve efficient collaborative learning among multiple agents in heterogeneous dynamic environments.
GraSS: Scalable Data Attribution with Gradient Sparsification and Sparse Projection
Pingbang Hu (University of Illinois Urbana-Champaign), Jiaqi W. Ma (University of Illinois Urbana-Champaign)
CodeOptimizationComputational EfficiencyText
π― What it does: This paper proposes two gradient compression algorithms, GRASS and FACTGRASS, which significantly reduce the memory and computational costs of large-scale model data attribution by leveraging the natural sparsity of gradients and parameters.
GRAVER: Generative Graph Vocabularies for Robust Graph Foundation Models Fine-tuning
Haonan Yuan (Beihang University), Philip S. Yu (University of Illinois)
CodeClassificationDomain AdaptationRepresentation LearningGraph Neural NetworkLarge Language ModelMixture of ExpertsContrastive LearningGraph
π― What it does: Proposes the GRAVER framework, which enhances the support set using a generative graph dictionary to achieve robust and efficient fine-tuning of graph-based models under multi-domain pre-training.
GRE Suite: Geo-localization Inference via Fine-Tuned Vision-Language Models and Enhanced Reasoning Chains
Chun Wang (Zhejiang University), Yiren Song (LibLib AI)
CodeTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelImageChain-of-Thought
π― What it does: The GRE Suite framework is proposed, combining visual language models with multi-stage reinforcement learning to enhance inference for global image geolocation.
GRIFFIN: Effective Token Alignment for Faster Speculative Decoding
Shijing Hu (Fudan University), Pan Zhou (Singapore Management University)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
π― What it does: This paper proposes a new speculative decoding framework called GRIFFIN, which explicitly addresses the token misalignment issue between the training and inference phases, significantly improving the generation speed of large language models.
π― What it does: An end-to-end framework is proposed to directly train RL agents using a pre-trained symbolic labeler from a limited number of annotated trajectories, followed by high-level tasks described by Reward Machines (RM).
Group-in-Group Policy Optimization for LLM Agent Training
Lang Feng (Nanyang Technological University), Bo An (Skywork AI)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: This paper proposes GiGPO, a group-based reinforcement learning algorithm that enables fine-grained credit allocation in multi-step LLM agent training.
π― What it does: An efficient group-level data selection framework called Group-MATES is proposed, which utilizes a relational data influence model to achieve speed-quality trade-off optimization during pre-training.
GSAlign: Geometric and Semantic Alignment Network for Aerial-Ground Person Re-Identification
Qiao Li (Wuhan University), Jiayi Ji
CodeRetrievalTransformerImage
π― What it does: This paper proposes a Geometric and Semantic Alignment Network (GSAlign) specifically designed to address the issues of geometric distortion and semantic misalignment caused by extreme viewpoint differences in aerial-ground person retrieval (AG-ReID).
CodeConvolutional Neural NetworkRecurrent Neural NetworkTabularTime SeriesElectronic Health Records
π― What it does: The GST-UNet framework is proposed to achieve single-trajectory spatiotemporal causal inference, combining a U-Net encoder with iterative G-computation, capable of simultaneously handling spatial interference, temporal confounding, and spatiotemporal lag effects.
π― What it does: This paper proposes GTR-Loc, a LiDAR positioning framework that utilizes geospatial text assistance to achieve accurate pose regression on single-frame point clouds.
Guard Me If You Know Me: Protecting Specific Face-Identity from Deepfakes
Kaiqing Lin (Shenzhen University), Bin Li (Shenzhen University)
CodeRecognitionSafty and PrivacyExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringImageMultimodalityBenchmark
π― What it does: The VIPGuard framework is proposed for personalized deepfake detection and interpretable reasoning for known identities.
GUARDIAN: Safeguarding LLM Multi-Agent Collaborations with Temporal Graph Modeling
Jialong Zhou (King's College London), Xiao Yang (Tsinghua University)
CodeAnomaly DetectionGraph Neural NetworkTransformerLarge Language ModelTextGraph
π― What it does: This paper proposes a framework called GUARDIAN, designed to detect and mitigate the issues of hallucination amplification and error injection and propagation in multi-agent collaboration of large language models.
GuardReasoner-VL: Safeguarding VLMs via Reinforced Reasoning
Yue Liu (National University of Singapore), Bryan Hooi (National University of Singapore)
CodeSafty and PrivacyTransformerReinforcement LearningVision Language ModelImageTextMultimodality
π― What it does: A multimodal reasoning safety guardian model, GuardReasoner-VL, has been constructed, which can perform reasoning before determining whether the input and output are harmful.
GUI-G1: Understanding R1-Zero-Like Training for Visual Grounding in GUI Agents
Yuqi Zhou (Renmin University of China), Jun Xu (Huawei)
CodeObject DetectionTransformerLarge Language ModelReinforcement LearningMultimodality
π― What it does: This paper studies and improves the R1-Zero-like training framework for graphical user interface (GUI) visual localization tasks. By systematically analyzing the three core components: input templates, reward functions, and policy updates, we propose the Fast Thinking template, Box Size constraint reward, and the GRPO improvement method that removes length bias and incorporates difficulty weighting. Ultimately, we train GUI-G1-3B on 17K public samples.
Guided Diffusion Sampling on Function Spaces with Applications to PDEs
Jiachen Yao (California Institute of Technology), Anima Anandkumar (California Institute of Technology)
CodeRestorationGenerationDiffusion modelTime SeriesPhysics Related
π― What it does: A discretization-invariant function space diffusion model called FunDPS is proposed to recover the posterior distribution of PDE solutions from extremely sparse or noisy measurements.
Guiding LLM Decision-Making with Fairness Reward Models
Zara Hall (Columbia University), Richard Zemel (Columbia University)
CodeClassificationRecommendation SystemTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: A general Fairness Reward Model (FRM) is constructed and trained to score the fairness of each step in the chain of thought (CoT) of large language models (LLMs) during the inference phase, thereby emphasizing fair reasoning paths in final decisions and improving the fairness of high-risk decisions (such as judicial risk assessment, social media content moderation, and job screening) without compromising accuracy.
π― What it does: This paper proposes a generation and annotation pipeline for fine-grained visual differences, and constructs the Micro Edit Dataset (MED) along with corresponding evaluation benchmarks.
π― What it does: A PDE solver based on the Hamiltonian framework is proposedβHamiltonian Neural Solver (HNS), which approximates the Hamiltonian functional through a learnable Integral Kernel Functional (IKF) and uses automatic differentiation to obtain functional derivatives for predicting the time evolution of infinite-dimensional systems.
π― What it does: The IDO framework is proposed, which achieves instance-level difficulty modeling and optimization for noisy label learning through two-stage training and dynamic weighted loss.
Hankel Singular Value Regularization for Highly Compressible State Space Models
Paul Schwerdtner (Courant Institute of Mathematical Sciences New York University), Benjamin Peherstorfer (Courant Institute of Mathematical Sciences New York University)
CodeCompressionSequentialBenchmark
π― What it does: This paper proposes a method for regularizing state space models (Hankel structure) during the training process, allowing the model to be efficiently compressed while maintaining high accuracy.
π― What it does: This paper presents HAODiff, a single-step diffusion model for portrait images that can achieve high-quality recovery in the presence of both global noise and human motion blur.
π― What it does: The Hierarchical Sparse Attention (HSA) mechanism is proposed, and based on this, the RAMba model is constructed, integrating RNN backbone, sparse attention, and memory reset mechanism to achieve efficient random access and length generalization for long contexts.
Harmony in Divergence: Towards Fast, Accurate, and Memory-efficient Zeroth-order LLM Fine-tuning
Qitao Tan (University of Georgia), Geng Yuan (University of Georgia)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: An efficient LLM fine-tuning method based on zero-order optimization, DiZO, is proposed. By comparing the hierarchical update differences between FO and ZO, a hierarchical diversification projection mechanism is designed to achieve learning effects similar to FO while significantly reducing memory usage.
Harnessing Feature Resonance under Arbitrary Target Alignment for Out-of-Distribution Node Detection
Shenzhi Yang (Zhejiang University), Haobo Wang (Zhejiang University)
CodeAnomaly DetectionGraph Neural NetworkGraph
π― What it does: This paper proposes an unsupervised, label-free, and pre-training-free method for out-of-distribution (OOD) detection of graph nodes, called RSL. It aligns the features of known in-distribution (ID) nodes to random targets and utilizes the differences in 'feature resonance' between ID nodes and unknown ID/OOD nodes in the single-step gradient direction to filter reliable OOD candidate nodes. It also uses Stochastic Gradient Langevin Dynamics (SGLD) to synthesize more realistic out-of-vocabulary (OOV) samples for training a binary classifier.
Hawaii: Hierarchical Visual Knowledge Transfer for Efficient Vision-Language Models
Yimu Wang (University of Waterloo), Krzysztof Czarnecki (University of Waterloo)
CodeKnowledge DistillationRepresentation LearningTransformerMixture of ExpertsVision Language ModelMultimodality
π― What it does: Proposes the HAWAII framework, which distills the knowledge of multiple visual experts into a single visual encoder to enhance the visual understanding capabilities of VLM;
CodeCompressionTransformerLarge Language ModelText
π― What it does: This paper proposes HBLLM, a 1-bit post-training quantization framework based on Haar wavelet transform, aimed at compressing large language models (LLMs) while maintaining high inference accuracy.
HermesFlow: Seamlessly Closing the Gap in Multimodal Understanding and Generation
Ling Yang (Princeton University), Bin CUI
CodeGenerationOptimizationTransformerLarge Language ModelReinforcement LearningImageTextMultimodality
π― What it does: This paper proposes the HermesFlow framework, which optimizes data through self-generated comparative advantages and disadvantages using Pair-DPO, achieving simultaneous improvements in understanding and generation capabilities in multimodal large language models (MLLMs) while narrowing the performance gap between the two.
Heterogeneous Swarms: Jointly Optimizing Model Roles and Weights for Multi-LLM Systems
Shangbin Feng (University of Washington), Tomas Pfister (Google)
CodeOptimizationLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: Designed the HETEROGENEOUS SWARMS algorithm to jointly optimize the model roles (DAG structure) and weights of multiple LLM systems using particle swarm optimization;
HetSyn: Versatile Timescale Integration in Spiking Neural Networks via Heterogeneous Synapses
Zhichao Deng (Tianjin University), Qiang Yu (Tianjin University)
CodeSpiking Neural NetworkTime SeriesAudio
π― What it does: Designed and implemented the HetSyn framework, introducing adjustable time constants at the synaptic level to achieve multi-time scale integration, and validated its effectiveness on multiple tasks using the HetSynLIF model.
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes Hierarchical Balance Packing (HBP), which addresses the workload imbalance issue in long-context LLM training through multi-layer data packing and dynamic training pipelines.
Hierarchical Demonstration Order Optimization for Many-shot In-Context Learning
Yinhan He (University of Virginia), Jundong Li (University of Virginia)
CodeOptimizationLarge Language ModelText
π― What it does: This paper studies the issue of demonstration order instability in many-shot in-context learning (ICL), proposing an information-theoretic measure called ICD-OVI and a hierarchical optimization framework (HIDO) that can efficiently search within a large-scale demonstration arrangement space.
Hierarchical Frequency Tagging Probe (HFTP): A Unified Approach to Investigate Syntactic Structure Representations in Large Language Models and the Human Brain
Jingmin An (Peking University), Fang Fang (Peking University)
CodeLarge Language ModelTextAudio
π― What it does: Proposed and implemented the Hierarchical Frequency Tagging Probe (HFTP) to detect the hierarchical structural representations of sentences and phrases in large language models (LLMs) and the human brain, and aligned them across various LLMs and human brain data.
Ruoxi Jiang (Fudan University), Rebecca Willett (University of Chicago)
CodeTime SeriesPhysics Related
π― What it does: A multi-scale implicit neural simulator is proposed, which significantly improves the stability and accuracy of long-term predictions by using multi-layer low-dimensional future state representations during prediction.
π― What it does: An end-to-end graph kernel network based on hierarchical shortest path graph kernels (HSP-GKN) is proposed, combining graph kernels with neural networks to achieve task-related graph representation learning.
π― What it does: This paper proposes a Time Encoding (TE) burst camera that utilizes a clock cycle counter to record super-threshold moments, significantly enhancing the dynamic range of the burst camera, and designs a complete image reconstruction network for TE burst streams.
π― What it does: We propose a high-order SE(3) symmetric flow matching framework called QHFLOW, which is used to predict the Kohn-Sham Hamiltonian matrix in density functional theory (DFT), generating it directly rather than through regression, significantly reducing the number of iterations required in the SCF cycle.
π― What it does: A differentiable architecture search framework ARITH-DAS is proposed, which directly performs fine-grained optimization of the interconnections of arithmetic circuits on multi-relation directed acyclic graphs.
π― What it does: A hierarchical encoding method based on hypergraphs (such as Hodge-Laplacian, random walk, discrete curvature, and local degree) is proposed to inject high-order structural information into traditional graph neural networks, enhancing the performance of multi-relational learning.
Highlighting What Matters: Promptable Embeddings for Attribute-Focused Image Retrieval
Siting Li (University of Washington), Simon Shaolei Du (University of Washington)
CodeRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
π― What it does: Construct the COCO-FACET benchmark dataset to evaluate attribute-focused text-image retrieval, and propose the use of promptable embeddings generated by multimodal large language models to enhance retrieval performance.
CodeClassificationRepresentation LearningGraph Neural NetworkPoint CloudBiomedical Data
π― What it does: Designed and implemented HiPoNet, an end-to-end differentiable high-dimensional point cloud network that utilizes multi-view reweighted features, Vietoris-Rips simplicial complex construction, and simplicial wave-particle transforms for multi-scale feature extraction, applied to regression, classification, and representation learning.
Hippocampal-like Sequential Editing for Continual Knowledge Updates in Large Language Models
Quntian Fang (National University of Defense Technology), Guotong Geng
CodeLarge Language ModelSupervised Fine-TuningText
π― What it does: A hippocampus-inspired sequential model editing framework (HSE) has been designed and implemented to continuously update knowledge without retraining large language models (LLMs), addressing the issues of parameter drift and catastrophic forgetting.
HM3: Hierarchical Multi-Objective Model Merging for Pretrained Models
Yu Zhou (Hong Kong Polytechnic University), KC Tan
CodeOptimizationTransformerLarge Language ModelReinforcement LearningImageText
π― What it does: This paper proposes and implements HM3βa hierarchical multi-objective model merging framework that can simultaneously search in the parameter space and architecture space to generate customizable high-performance merged models.