International Conference on Machine Learning Β· 722 papers
Efficient Parallel Training Methods for Spiking Neural Networks with Constant Time Complexity
Wanjin Feng (Institute of Microelectronics Chinese Academy of Sciences), Chunyan Miao (Nanyang Technological University)
CodeSpiking Neural NetworkImageSequential
π― What it does: Proposes a Fixed Point Parallel Training (FPT) method, which reduces the time complexity of SNN from O(T) to O(K) through fixed point iteration, achieving parallel training;
Efficiently Serving Large Multimodal Models Using EPD Disaggregation
Gursimran Singh (Huawei Technologies), Zhenan Fan (Huawei Technologies)
CodeOptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoTextMultimodality
π― What it does: An EPD Disaggregation framework is designed, deploying the encoding, pre-filling, and decoding stages of large-scale multimodal models (LMM) on separate resources to reduce latency and memory usage.
π― What it does: A macro cell placement method called EGPlace is proposed, which is based on evolutionary search and greedy relocation-guided mutation for efficient macro cell layout.
Eigenspectrum Analysis of Neural Networks without Aspect Ratio Bias
Yuanzhe Hu (Dartmouth College), Yaoqing Yang (Dartmouth College)
CodeOptimizationHyperparameter SearchConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningImageText
π― What it does: The FARMS method is proposed and validated, utilizing fixed aspect ratio submatrix sampling and average ESD to eliminate the bias of the weight matrix aspect ratio on the analysis of significant feature spectra, and improving hierarchical hyperparameter tuning effects in various scenarios such as LLM pruning, image classification, and scientific machine learning fine-tuning.
ELMO : Efficiency via Low-precision and Peak Memory Optimization in Large Output Spaces
Jinbin Zhang (Aalto University), Rohit Babbar (University of Bath)
CodeOptimizationComputational EfficiencyLarge Language ModelTextBenchmark
π― What it does: Proposed the ELMO framework, which uses pure low precision (BF16/FP8) to train extremely large output space models, achieving significant reductions in memory usage and computation.
π― What it does: A low-rank adaptation PEFT method ELoRA for SO(3) equivariant graph neural networks is proposed and implemented, and fine-tuned on pre-trained atomic potentials.
Embedding Safety into RL: A New Take on Trust Region Methods
Nikola Milosevic (Max Planck Institute for Human Cognitive and Brain Sciences), Nico Scherf (Max Planck Institute for Human Cognitive and Brain Sciences)
CodeOptimizationSafty and PrivacyReinforcement LearningTabularBenchmark
π― What it does: In the framework of Constrained Markov Decision Processes (CMDP), the authors propose Constrained Trust Region Policy Optimization (C-TRPO) and Constrained Natural Policy Gradient (C-NPG), which construct a safe trust region in the policy space to ensure that the training process always meets safety constraints.
Emoji Attack: Enhancing Jailbreak Attacks Against Judge LLM Detection
Zhipeng Wei (International Computer Science Institute), N. Benjamin Erichson (Lawrence Berkeley National Laboratory)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
π― What it does: The study investigates and verifies the vulnerability of Judge LLM under token segmentation bias and proposes an attack method utilizing emoji injectionβEmoji Attackβto enhance the success rate of jailbreak evasion detection.
π― What it does: A framework that combines adversarial training with conformal prediction is proposed, designing the OPSA attack that maximizes the size of the prediction set without requiring coverage information, along with its corresponding OPSA-AT defense method.
Enhancing Cooperative Multi-Agent Reinforcement Learning with State Modelling and Adversarial Exploration
Andreas Kontogiannis (National Technical University of Athens), George Vouros (University of Piraeus)
CodeReinforcement LearningAuto Encoder
π― What it does: A multi-agent reinforcement learning framework called SMPE 2 is proposed, which is based on state modeling and adversarial exploration to enhance cooperation in partially observable environments without communication.
Enhancing Decision-Making of Large Language Models via Actor-Critic
Heng Dong (Tsinghua University), Chongjie Zhang (Washington University in St. Louis)
CodeRobotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: A LLM-based Actor-Critic framework called LAC is designed, integrating LLM prior strategies with action evaluation for long-term planning to achieve efficient multi-step decision-making.
Enhancing Statistical Validity and Power in Hybrid Controlled Trials: A Randomization Inference Approach with Conformal Selective Borrowing
Ke Zhu (North Carolina State University), Xiaofei Wang (Duke University)
CodeBiomedical Data
π― What it does: In a mixed control trial, a precise randomization inference framework based on Fisher's randomization test is proposed, and selective borrowing from external controls (EC) is achieved through an adaptive conformal selection method, thereby maintaining precise control of the first type error and enhancing test power in small sample environments.
Enhancing Treatment Effect Estimation via Active Learning: A Counterfactual Covering Perspective
Hechuan Wen (University of Queensland), Hongzhi Yin (University of Queensland)
CodeTabular
π― What it does: Under the constraint of a limited label budget, an active learning strategy is utilized to select samples from observational data to enhance the accuracy of treatment effect estimation.
Talor Abramovich (Tel Aviv University), Ofir Press (Princeton Language and Intelligence)
CodeOptimizationTransformerLarge Language ModelAgentic AITextBenchmark
π― What it does: EnIGMA is a language model agent designed for Capture The Flag (CTF) challenges, which automates the process of finding and exploiting security vulnerabilities by integrating interactive tools (such as the gdb debugger and remote connection tools) within a Docker environment.
EPIC: Efficient Position-Independent Caching for Serving Large Language Models
Junhao Hu (Peking University), Tao Xie (Peking University)
CodeComputational EfficiencyTransformerLarge Language ModelText
π― What it does: The EPIC system and LegoLink algorithm are proposed to achieve position-independent (PIC) caching in large language model services, significantly improving inference efficiency by reusing cached KV vectors without the need for precise prefix matching.
EpiCoder: Encompassing Diversity and Complexity in Code Generation
Yaoxiang Wang (Xiamen University), Scarlett Li (Microsoft)
CodeGenerationData SynthesisAI Code AssistantTransformerLarge Language ModelTextBenchmark
π― What it does: A code synthesis framework based on feature trees is proposed, utilizing hierarchical code features to achieve controllable complexity and diversity in instruction data generation.
π― What it does: This paper proposes the EraseAnything method, specifically designed for Transformer visual-text models like Flux, to delete undesirable concepts while maintaining generation quality.
Evaluating Judges as Evaluators: The JETTS Benchmark of LLM-as-Judges as Test-Time Scaling Evaluators
Yilun Zhou (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: This paper constructs the JETTS benchmark to systematically evaluate the performance of LLM judges in testing scenarios involving expansion (reordering, beam search, critical improvement).
Evaluating Neuron Explanations: A Unified Framework with Sanity Checks
Tuomas Oikarinen (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
CodeExplainability and InterpretabilityConvolutional Neural NetworkTransformerVision Language ModelImageText
π― What it does: A unified neuron explanation evaluation framework, NeuronEval, is proposed, and within this framework, theoretical and experimental validation of 18 existing evaluation metrics is conducted, introducing two testing methods: missing label testing and additional label testing.
EVOLvE: Evaluating and Optimizing LLMs For In-Context Exploration
Allen Nie (Stanford University), Minmin Chen (Google DeepMind)
CodeOptimizationKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
π― What it does: This paper proposes the BanditBench benchmark, conducting unsupervised 'self-exploration' experiments with large language models (LLMs) in multi-armed and contextual bandit environments, and investigates two methods to enhance LLM exploration capabilities: algorithm guidance and algorithm distillation.
EvoPress: Accurate Dynamic Model Compression via Evolutionary Search
Oliver Sieberling (ETH Zurich), Dan Alistarh (IST Austria)
CodeCompressionOptimizationLarge Language ModelText
π― What it does: EvoPress is proposedβa dynamic compression framework based on evolutionary search, designed to automatically find the optimal compression configuration for each layer (including layer pruning, sparsification, and quantization) while satisfying overall compression constraints.
π― What it does: A generalized error upper bound based on binary JensenβShannon divergence is proposed, constructing a theoretically achievable precise and tight upper bound in random learning algorithms;
ExLM: Rethinking the Impact of $\texttt{[MASK]}$ Tokens in Masked Language Models
Kangjie Zheng (Peking University), Ming Zhang (Peking University)
CodeTransformerLarge Language ModelTextBiomedical Data
π― What it does: This paper systematically studies the 'semantic corruption' problem caused by the [MASK] token in masked language models and proposes the EXLM model, which enhances the model's contextual understanding and downstream performance by performing multi-state expansion for each [MASK] token and explicitly modeling the semantic dependencies between states.
π― What it does: The concept of 'exogenous isomorphism' is proposed, and ~EI-identifiability is introduced to simplify the complete causal identifiability analysis of the L3 layer; identifiability conditions are provided for two types of models: Bijective SCM and Triangular Monotonic SCM, and based on this, a neural TM-SCM is constructed for counterfactual consistency inference in practice.
π― What it does: The LieNLSD method is proposed, which can explicitly discover nonlinear symmetries from dynamic data and automatically determine the number and expressions of generators.
Exploiting Curvature in Online Convex Optimization with Delayed Feedback
Hao Qiu (Universita degli Studi di Milano), Mengxiao Zhang (University of Iowa)
CodeOptimizationTabular
π― What it does: This study investigates online convex optimization under delayed feedback, utilizing loss curvature to improve the regret upper bound.
Exploiting Presentative Feature Distributions for Parameter-Efficient Continual Learning of Large Language Models
Xin Cheng (Southeast University), Lei Feng (Southeast University)
CodeClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningTextSequentialBenchmark
π― What it does: This paper proposes a method for achieving parameter-efficient online continual learning by utilizing the distribution of performance features from pre-trained large language models, aiming to avoid information leakage (IL) and enhance CL performance.
Exploring Criteria of Loss Reweighting to Enhance LLM Unlearning
Puning Yang (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
CodeOptimizationSafty and PrivacyTransformerLarge Language ModelText
π― What it does: This paper explores the role of loss reweighting in the context of zero-shot learning in large language models through experiments, proposing two types of reweighting objectives: Saturation and Importance, and based on this, designs a new SatImp scheme.
Extracting Rare Dependence Patterns via Adaptive Sample Reweighting
Yiqing Li (Mohamed Bin Zayed University of Artificial Intelligence), Kun Zhang (Carnegie Mellon University)
CodeAnomaly DetectionOptimizationTabularTime SeriesFinance Related
π― What it does: A method that combines kernel-based independence testing with adaptive sample importance reweighting is proposed to detect rare dependence patterns that are significant only in small regions but exist in the overall sample; this test is embedded in the PC algorithm to achieve causal structure learning under rare dependence.
FairPFN: A Tabular Foundation Model for Causal Fairness
Jake Robertson (University of Freiburg), Frank Hutter
CodeTransformerTabular
π― What it does: This paper proposes FairPFN, a Transformer-based foundational model for tabular data, aimed at eliminating the causal influence of protected attributes under the condition of having only observational data, thereby enhancing algorithmic fairness.
π― What it does: A training-free, partially denoised acceleration framework called Falcon is proposed, which can leverage the sequential dependencies of historical actions in visual motion tasks, significantly reducing the number of denoising steps and achieving real-time control.
Fast and Low-Cost Genomic Foundation Models via Outlier Removal
Haozheng Luo (Northwestern University), Han Liu (Northwestern University)
CodeClassificationOptimizationTransformerSupervised Fine-TuningBiomedical Data
π― What it does: A genome-based model named GERM has been constructed, utilizing an outlier-free attention layer to eliminate outliers in attention, thereby achieving fast low-rank adaptation and robust post-training quantization.
Fast Estimation of Partial Dependence Functions using Trees
Jinyang Liu (University of Copenhagen), Munir Hiabu (University of Copenhagen)
CodeExplainability and InterpretabilityComputational EfficiencyTabularBenchmark
π― What it does: The FastPD algorithm is proposed for the rapid and consistent estimation of the bias function of any subset in tree models, thereby obtaining complete functional decomposition and SHAP values.
π― What it does: Proposes Sliding Tile Attention (STA), which replaces full 3D attention in video diffusion Transformers, significantly reducing inference time.
π― What it does: This paper proposes an accelerated version of the Karger and Karger-Stein minimum cut algorithms using machine learning predictions, which can significantly reduce the time complexity of solving the global minimum cut under the premise of known prediction error rates.
π― What it does: A feature upsampling framework named FeatSharp is proposed, which can elevate the feature maps of low-resolution visual encoders to higher resolutions at a lower cost while maintaining the original model representation, and utilizes tile information for fine-grained refinement.
π― What it does: This paper proposes a framework based on representational geometry and set capacity to quantify the depth, strategy, and stages of neural networks in the feature learning process, and validates its effectiveness through experiments.
Features are fate: a theory of transfer learning in high-dimensional regression
Javan Tahir (Stanford University), Grant M. Rotskoff
CodeTabular
π― What it does: This paper constructs a theoretical framework for transfer learning in high-dimensional regression tasks by analyzing deep linear networks and shallow ReLU networks. It derives a phase diagram of transfer performance and verifies that the overlap of feature spaces is a key indicator determining the success of transfer.
Federated Incomplete Multi-view Clustering with Globally Fused Graph Guidance
Guoqing Chao (Harbin Institute of Technology), Dianhui Chu (Harbin Institute of Technology)
CodeFederated LearningGraph Neural NetworkGraph
π― What it does: This paper proposes a Federated Incomplete Multi-View Clustering method (FIMCFG), which achieves end-to-end feature extraction and clustering through a globally fused graph-guided dual-head graph convolutional encoder and fusion module.
π― What it does: This paper proposes FedSMU, an algorithm that symbolically updates client models in federated learning (using sign operations) and employs the Lion optimizer for split execution, which reduces communication costs and enhances the generalization performance of the global model.
π― What it does: An AI-generated image detector based on few-shot learning (Few-Shot Detector, FSD) is proposed, which can identify fake images generated by a given small number of samples from unknown generative models.
FG-CLIP: Fine-Grained Visual and Textual Alignment
Chunyu Xie (Beihang University), Yuhui Yin (360 AI Research)
CodeClassificationRecognitionRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes Fine-Granularity CLIP (FG-CLIP), which achieves fine-grained alignment between images and text through a three-stage training process (global contrast, regional contrast, and hard negative sample learning);
Fixing the Double Penalty in Data-Driven Weather Forecasting Through a Modified Spherical Harmonic Loss Function
Christopher Subich (Environment and Climate Change Canada), Jing Yang (Environment and Climate Change Canada)
CodeGraph Neural NetworkTime Series
π― What it does: By rewriting the mean squared error loss function and utilizing spherical harmonic decomposition to separate amplitude errors from correlation errors, the double-penalty problem in data-driven weather forecasting is eliminated.
FlashTP: Fused, Sparsity-Aware Tensor Product for Machine Learning Interatomic Potentials
Seung Yul Lee (Seoul National University), Jae W. Lee (Seoul National University)
CodeComputational EfficiencyTabular
π― What it does: A GPU-accelerated library named FlashTP has been developed to efficiently implement the tensor product layer in machine learning interatomic potentials (MLIP), significantly improving inference and training speed while reducing memory usage.
π― What it does: This paper studies a novel post-training quantization method called FLATQUANT, which enhances the flatness of LLM weights and activations through learnable fast affine transformations, significantly reducing quantization error.
π― What it does: A multi-agent framework named FOA is proposed, utilizing LLMs as agents to perform autonomous exploration and resampling based on genetic particle filtering in dynamic tree search, addressing complex reasoning and decision-making tasks.
Flexibility-conditioned protein structure design with flow matching
Vsevolod Viliuga (Heidelberg Institute for Theoretical Studies), Frauke GrΓ€ter (Max Planck Institute for Polymer Research)
CodeProtein Structure PredictionFlow-based ModelBiomedical Data
π― What it does: This paper proposes a scalable protein structure generation framework based on flow matching (FliPS), which can synthesize protein backbones under a given flexible distribution for each residue and filter them using BackFlip;
π― What it does: A new grammar-constrained decoding (GCD) algorithm called GREATGRAMMA is proposed and implemented, which reduces offline preprocessing time by constructing a token spanner table while maintaining online masking efficiency.
π― What it does: A dynamic delay poisoning attack (DDPA) targeting machine learning models is proposed, which specifies the attack target after model training and uses poisoned data to trigger malicious adjustments to model parameters when the model loses specific samples.
Yue Liu (National University of Singapore), Bryan Hooi (National University of Singapore)
CodeAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
π― What it does: A black-box LLM jailbreak method called FlipAttack is proposed, which utilizes left-side perturbations and flipping techniques to bypass security protections.
π― What it does: This paper proposes a social recommendation framework called RecFlow based on flow matching, which uses flow matching to denoise the noise in social graphs.
Flow of Reasoning: Training LLMs for Divergent Reasoning with Minimal Examples
Fangxu Yu (University of California San Diego), Lianhui Qin (University of California San Diego)
CodeTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
π― What it does: Proposes Flow of Reasoning (FOR), a method for fine-tuning large language models using the principles of GFlowNet to generate diverse reasoning paths.
π― What it does: A framework for image generation called FlowAR is proposed, which combines scale autoregression with flow matching, capable of generating high-quality images at each scale while maintaining a two-dimensional structure.
Forest-of-Thought: Scaling Test-Time Compute for Enhancing LLM Reasoning
Zhenni Bi (Huawei Noah's Ark Lab), Yunhe Wang (Huawei Noah's Ark Lab)
CodeTransformerLarge Language ModelTextChain-of-Thought
π― What it does: The Forest-of-Thought (FoT) framework is proposed, which integrates multiple reasoning trees and enhances the complex reasoning ability of LLMs through three mechanisms: sparse activation, dynamic self-correction, and consensus decision-making.
Foundation Model Insights and a Multi-Model Approach for Superior Fine-Grained One-shot Subset Selection
Zhijing Wan (Wuhan University), Shin'ichi Satoh (National Institute of Informatics)
CodeClassificationContrastive LearningImage
π― What it does: This paper studies the use of Foundation Models (FMs) as Information Extractors (IE) for one-time subset selection and proposes a multi-model fusion method called RAM-APL to enhance the subset selection performance for fine-grained image classification.
CodeRetrievalOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper analyzes the frequency domain properties of RoPE from the perspective of discrete signal processing, pointing out that spectral damage caused by linear layers, activation functions, and undertrained low-frequency components leads to this issue. It then proposes Fourier Position Embedding (FoPE) β using Fourier series in each dimension and setting low-frequency components to zero to enhance the periodic extension and length generalization of attention, and conducts pre-training, continued pre-training, and fine-tuning experiments on multi-scale models.
π― What it does: This paper proposes the FrameBridge framework, modeling frame-to-frame generation from images to videos as a data-to-data bridge process, replacing the traditional noise-to-data diffusion generation.
From Complex to Atomic: Enhancing Augmented Generation via Knowledge-Aware Dual Rewriting and Reasoning
Jinyu Wang (Microsoft Research Asia), Jiang Bian (Microsoft Research Asia)
CodeGenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
π― What it does: An iterative RAG framework based on knowledge atomization, query generation, atomic retrieval, and atomic selection (KAR-RAG) is proposed to achieve adaptive decomposition and reasoning for complex multi-hop questions.
From Language Models over Tokens to Language Models over Characters
Tim Vieira (ETH ZΓΌrich), Ryan Cotterell (ETH ZΓΌrich)
CodeGenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A theoretical and algorithmic framework is proposed to transition from token-based language models to character-based interfaces, addressing the prompt boundary issue and achieving character-level probability estimation and conditional generation.
From Local Details to Global Context: Advancing Vision-Language Models with Attention-Based Selection
Lincan Cai (Beijing Institute of Technology), Jian Liang (Kuaishou Technology)
CodeClassificationDomain AdaptationRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: By utilizing the attention map from DINO to guide image cropping and performing cropping in the feature space to retain global semantics, combined with soft matching to filter fine-grained descriptions generated by LLM, the ABS (Attention-Based Selection) method is proposed to enhance the zero-shot and OOD generalization performance of CLIP.
From Pixels to Perception: Interpretable Predictions via Instance-wise Grouped Feature Selection
Moritz Vandenhirtz (ETH Zurich), Julia E Vogt
CodeClassificationExplainability and InterpretabilityTransformerSupervised Fine-TuningImage
π― What it does: By performing instance-level binarization on images and sparsifying them based on semantically meaningful regions, interpretable predictions are achieved; at the same time, the sparsity threshold is dynamically adjusted during inference based on classifier confidence.
Fully Dynamic Euclidean Bi-Chromatic Matching in Sublinear Update Time
Gramoz Goranci (University of Vienna), Da Wei Zheng
CodeTime Series
π― What it does: This paper presents the first fully dynamic bi-color matching algorithm that achieves sublinear update time in Euclidean two-dimensional space for rapid estimation of the W1-Wasserstein distance.
Fully Heteroscedastic Count Regression with Deep Double Poisson Networks
Spencer Young (Delicious AI), Hua Wei (Arizona State University)
CodeImageTextPoint CloudTabular
π― What it does: The Deep Double Poisson Network (DDPN) is proposed, achieving complete heteroscedastic uncertainty quantification in count regression and enhancing performance through integration.
G-Adaptivity: optimised graph-based mesh relocation for finite element methods
James Rowbottom (University of Cambridge), Chris Budd (University of Bath)
CodeOptimizationGraph Neural NetworkMesh
π― What it does: A grid relocation method based on graph neural networks, G-Adaptivity, has been developed to directly minimize FEM errors, significantly improving the accuracy of finite element solutions while keeping the number of grid nodes unchanged.
π― What it does: A Gamma Distribution-based Principal Component Analysis (Ξ PCA) method is proposed to extract features that are insensitive to observation angles, thereby enhancing the robustness of Synthetic Aperture Radar (SAR) target recognition.
π― What it does: A Gaussian Mixture Flow Matching (GMFlow) model is proposed, which predicts flow velocity using a multimodal Gaussian mixture distribution to improve the quality of few-step sampling.
GCAL: Adapting Graph Models to Evolving Domain Shifts
Ziyue Qiao (Great Bay University), Hui Xiong (Hong Kong University of Science and Technology)
CodeDomain AdaptationOptimizationGraph Neural NetworkContrastive LearningGraphTime Series
π― What it does: Proposes the GCAL framework, achieving continuous adaptation on graph data with continuous unsupervised domain drift, utilizing a dual-layer optimization of adaptation and memory replay.
GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation
Yi Cai (Freie Universitaet Berlin), Gerhard Wunder (Freie Universitaet Berlin)
CodeExplainability and InterpretabilityImageText
π― What it does: A general feature attribution framework GEFA based on proxy gradient estimation is proposed, which can explain any model in a black-box environment.
π― What it does: This paper proposes a theoretical analysis framework for ensemble clustering and designs a new ensemble clustering algorithm based on this framework. First, the upper bounds for the generalization error and overfitting risk of ensemble clustering are provided, proving them to be O(β(log n m) + 1/βn), along with sufficient conditions for consistency (m β« log n). Secondly, the error is decomposed into a Bias-Diversity structure, indicating that minimizing the bias of the base distribution and maximizing diversity can reduce error, and equating diversity enhancement to robust min-max optimization. Finally, an algorithm based on high-confidence elements for K* approximation and weight learning is implemented, using reduced gradient descent for solving.
Generalization Principles for Inference over Text-Attributed Graphs with Large Language Models
Haoyu Peter Wang (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
CodeClassificationGraph Neural NetworkLarge Language ModelPrompt EngineeringTextGraph
π― What it does: This paper proposes an unsupervised, zero-shot graph classification method called LLM-BP, which utilizes large language models for task-adaptive text embedding and Bayesian propagation aggregation of graph structures.
π― What it does: A recursive learning and category distribution regularization (RLCD) framework is proposed to improve the performance of base class discrimination and new class discovery in the general category discovery (GCD) task.
Dimitri von RΓΌtte (ETH Zurich), Thomas Hofmann (ETH Zurich)
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelText
π― What it does: A Generalized Interpolated Discrete Diffusion (GIDD) framework is proposed, extending masked diffusion to arbitrary mixed distributions, providing closed-form forward/backward transitions and ELBO; and achieving self-correction under a masked + uniform noise mixture.
π― What it does: A framework for audio generation based on the Transformer decoder is proposed, using continuous-valued audio tokens and introducing a Masked Next Token Prediction (MNTP) task.
π― What it does: This paper proposes DiffMove, a conditionally diffusion model based on embedding space, designed to recover missing locations in sparse human trajectories.
π― What it does: This paper proposes a Noise Conditioned Graph Network (NCGN) and its specific implementation, Dynamic Message Passing (DMP), to dynamically adjust the connectivity range and resolution of the graph based on different noise levels during the generation process in stream-based generative models (diffusion, flow-matching), thereby improving the quality of geometric shape generation (3D point clouds, spatial transcriptomics, images).
π― What it does: Proposes the GeoRCG framework, which generates semantic representations using a pre-trained geometric encoder before molecular generation.
π― What it does: This paper studies the complexity and optimality of the Follow-the-Perturbed-Leader (FTPL) strategy in the K-armed bandit problem. A new technique called Conditional Geometric Resampling (CGR) is proposed for unbiased loss estimation, significantly reducing the average complexity per round.
Geometry Informed Tokenization of Molecules for Language Model Generation
Xiner Li (Texas A&M University), Shuiwang Ji (Texas A&M University)
CodeGenerationDrug DiscoveryTransformerLarge Language ModelDiffusion modelGraph
π― What it does: Proposes the Geo2Seq method, which discretizes 3D molecular geometric structures into a one-dimensional token sequence invariant to SE(3), and utilizes large language models to generate 3D molecules.
GeoPixel: Pixel Grounding Large Multimodal Model in Remote Sensing
Akashah Shabbir (Mohamed bin Zayed University of Artificial Intelligence), Salman Khan (Australian National University)
CodeObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageTextMultimodality
π― What it does: This paper proposes GeoPixel, which enables pixel-level visual language dialogue and multi-object segmentation for high-resolution remote sensing images, generating fine-grained natural language descriptions.
GLGENN: A Novel Parameter-Light Equivariant Neural Networks Architecture Based on Clifford Geometric Algebras
Ekaterina Filimoshina (HSE University), Dmitry Shirokov (Institute for Information Transmission Problems of the Russian Academy of Sciences)
Code
π― What it does: A parameter-lightweight equivariant neural network architecture GLGENN is proposed, which utilizes geometric (Clifford) algebra to achieve equivariance for arbitrary pseudo-orthogonal transformations and implements parameter compression through weight sharing.
π― What it does: This paper proposes a new global optimization method called GS-PowerOpt, which first amplifies the weight of the global maximum point of the objective function through power transformation (or exponential power transformation), then applies Gaussian smoothing to the transformed function, and uses zero-order stochastic gradient ascent iteration to find the maximum point.
π― What it does: This paper proposes Gradient Aligned Regression (GAR), which improves regression models in the label space through two pairwise losses (difference matching and normalized difference matching).
Gradient-based Explanations for Deep Learning Survival Models
Sophie Hanna Langbein (Leibniz Institute for Prevention Research and Epidemiology), Marvin N. Wright (Leibniz Institute for Prevention Research and Epidemiology)
CodeExplainability and InterpretabilityMultimodalityTabular
π― What it does: A temporal gradient interpretability method for deep survival models is proposed and implemented, including GradSHAP(t), etc.;
GRAM: A Generative Foundation Reward Model for Reward Generalization
Chenglong Wang (Northeastern University), JingBo Zhu
CodeRecommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: A Generative Reward Model (GRAM) is proposed, which learns the input-response relationship using unlabeled data through unsupervised pre-training and supervised fine-tuning in two stages, and then fine-tunes with a small amount of labeled preference data.
Graph Adaptive Autoregressive Moving Average Models
Moshe Eliasof (University of Cambridge), Carola-Bibiane SchΓΆnlieb
CodeGraph Neural NetworkGraphTime Series
π― What it does: A framework called GRAMA is proposed, which converts graph structures into time series and learns selectable autoregressive moving average (ARMA) models on them, seamlessly integrating with any GNN architecture;
π― What it does: A self-regressive graph generation model G2PT based on Transformer is proposed, which uses tokenized graph sequence representation and can directly predict the graph structure by generating the next token.
Graph Inverse Style Transfer for Counterfactual Explainability
Bardh Prenkaj (Technical University of Munich), Gjergji Kasneci (Technical University of Munich)
CodeGenerationExplainability and InterpretabilityGraph Neural NetworkTransformerGraph
π― What it does: A reverse generation framework GIST based on graph style transfer is proposed, which generates graph counterfactual explanations that conform to structure and semantics using a backward backtracking approach.
Graph Neural Network Generalization With Gaussian Mixture Model Based Augmentation
Yassine ABBAHADDOU, Michalis Vazirgiannis
CodeClassificationGraph Neural NetworkGraph
π― What it does: A Gaussian Mixture Model-based graph data augmentation method called GRATIN is proposed to enhance the generalization and robustness of graph neural networks in graph classification tasks.
π― What it does: This paper presents GraphGPTβa self-supervised generative pre-trained graph model based on Graph Euler Transformer (GET), which can directly input any graph into a standard Transformer for learning after serializing it through an Euler path.
Gridded Transformer Neural Processes for Spatio-Temporal Data
Matthew Ashman (University of Cambridge), Richard E. Turner (University of Cambridge)
CodeTransformerTime Series
π― What it does: This paper proposes 'gridded TNP'βa neural process framework that maps unstructured spatiotemporal data to a structured pseudo-label grid, and utilizes efficient Transformers (ViT, Swin) for encoding, processing, and decoding on this grid.
GS-Bias: Global-Spatial Bias Learner for Single-Image Test-Time Adaptation of Vision-Language Models
Zhaohong Huang (Xiamen University), Rongrong Ji (Xiamen University)
CodeDomain AdaptationTransformerVision Language ModelContrastive LearningImage
π― What it does: A testing-time adaptive method named GS-Bias is proposed, which directly adjusts the output logits of the CLIP model by learning global and spatial biases during testing on a single image, in order to enhance zero-shot generalization performance.
GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning
Xiangheng Wang (Zhejiang University), Yunjun Gao (Zhejiang University)
CodeRepresentation LearningGraph Neural NetworkTransformerMixture of ExpertsTime SeriesSequential
π― What it does: Proposes the GTR framework, which combines a multi-view encoder (road network view + free space grid view) with position and time embeddings to learn a global multi-dimensional representation of trajectories;
Guarantees of a Preconditioned Subgradient Algorithm for Overparameterized Asymmetric Low-rank Matrix Recovery
Paris Giampouras (University of Warwick), Rene Vidal
CodeAnomaly DetectionOptimizationTabular
π― What it does: A over-parameterized preconditioned subgradient algorithm (OPSA) is proposed for robust low-rank matrix recovery in the presence of outliers, unknown rank, and asymmetric matrices.
Gumiho: A Hybrid Architecture to Prioritize Early Tokens in Speculative Decoding
Jinze Li (Advanced Micro Devices), Emad Barsoum (Advanced Micro Devices)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: A hybrid model called Gumiho, which combines sequence Transformer and parallel MLP, is proposed to enhance the efficiency of Speculative Decoding during inference.
π― What it does: The H-Tuning framework is proposed, which first uses low-cost mix-order optimization and LoRA low-rank adaptation for efficient fine-tuning of large pre-trained models, and then compresses the fine-tuned teacher model into a very small student model through knowledge distillation, for multi-label cardiovascular disease (CVD) detection on mobile ECG.
HALoS: Hierarchical Asynchronous Local SGD over Slow Networks for Geo-Distributed Large Language Model Training
Geon-Woo Kim (University of Texas at Austin), Aditya Akella (University of Texas at Austin)
CodeLarge Language ModelText
π― What it does: A hierarchical asynchronous local SGD framework named HALoS is proposed to address communication bottlenecks and hardware heterogeneity in geographically distributed LLM training.
Handling Imbalanced Pseudolabels for Vision-Language Models with Concept Alignment and Confusion-Aware Calibrated Margin
Yuchen Wang (Harbin Institute of Technology), Xinyang Chen (Harbin Institute of Technology)
CodeClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelImage
π― What it does: In unsupervised, semi-supervised, and transductive zero-shot learning tasks for VLM, a solution to the pseudo-label imbalance problem is proposed through concept alignment and confusion-based calibrated margins.