NeurIPS 2025 Papers — Page 31
Conference on Neural Information Processing Systems · 5275 papers
NEED: Cross-Subject and Cross-Task Generalization for Video and Image Reconstruction from EEG Signals
Shuai Huang (University of Shanghai for Science and Technology), Yongxiong Wang (University of Shanghai for Science and Technology)
RestorationGenerationTransformerDiffusion modelImageVideoMultimodality
🎯 What it does: A unified framework named NEED is proposed, achieving zero-shot generalization of EEG-based visual reconstruction across subjects and tasks.
NeedleInATable: Exploring Long-Context Capability of Large Language Models towards Long-Structured Tables
Lanrui Wang (Institute of Information Engineering, Chinese Academy of Sciences), Weiping Wang (Meituan)
Data SynthesisRetrievalTransformerLarge Language ModelSupervised Fine-TuningTabularBenchmarkChain-of-Thought
🎯 What it does: The NEEDLEINATABLE (NIAT) benchmark is proposed, focusing on evaluating the fine-grained cell localization and retrieval capabilities of large language models in long tables, and enhancing the model's understanding of long tables through synthetic data.
Negative Feedback Really Matters: Signed Dual-Channel Graph Contrastive Learning Framework for Recommendation
Leqi Zheng (Tsinghua University), Ziyang Liu (Boston Children's Hospital)
Recommendation SystemGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes a dual-channel symbolic graph contrastive learning framework (SDCGCL) for recommendation systems and its specialized model DualFuse, which fully utilizes positive and negative feedback to enhance recommendation effectiveness.
NegoCollab: A Common Representation Negotiation Approach for Heterogeneous Collaborative Perception
Shao Congzhang, Jinglin Li (Beijing University of Posts and Telecommunications)
Domain AdaptationAutonomous DrivingRepresentation LearningConvolutional Neural NetworkContrastive LearningMultimodality
🎯 What it does: In heterogeneous collaborative perception, a common representation is generated through a negotiator, and a sender-receiver is used for bidirectional mapping of local features, thereby eliminating the domain gap between different perception models and enhancing collaborative performance.
Nemotron-Flash: Towards Latency-Optimal Hybrid Small Language Models
Yonggan Fu (NVIDIA Research), Pavlo Molchanov (NVIDIA Research)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study proposes the Nemotron-Flash series of hybrid small language models, systematically exploring the model's depth-width ratio, operator combinations, and training techniques. It automatically discovers the optimal operator combinations through evolutionary search, ultimately training a small language model (SLM) that achieves both low latency and high accuracy on real-time devices.
NEP: Autoregressive Image Editing via Next Editing Token Prediction
Huimin Wu (State Key Laboratory of General Artificial Intelligence), Qing Li (State Key Laboratory of General Artificial Intelligence)
Image TranslationGenerationTransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: A next editing token prediction (NEP) framework based on autoregressive generation is proposed, enabling localized text-guided image editing that regenerates only specified areas, avoiding ineffective computation and unintended modifications in non-editing regions caused by full image reconstruction.
Neptune-X: Active X-to-Maritime Generation for Universal Maritime Object Detection
Yu Guo (City University of Hong Kong), Yuguang Fang (City University of Hong Kong)
Object DetectionGenerationData SynthesisDiffusion modelImage
🎯 What it does: The Neptune-X framework is proposed, combining the X-to-Maritime generative model and the Attribute-dependent Active Sampling strategy to enhance maritime object detection performance using synthetic maritime scene data.
Nested Learning: The Illusion of Deep Learning Architectures
Ali Behrouz (Google Research), Vahab Mirrokni (Google Research)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a Nested Learning framework, treating models, optimizers, and memory mechanisms as multi-layer nested optimization problems. Based on this, it designs more expressive optimizers, self-modifying sequence models, and continuous memory systems, ultimately presenting the HOPE architecture and evaluating it on language modeling, common sense reasoning, and continual learning tasks.
NestedFP: High-Performance, Memory-Efficient Dual-Precision Floating Point Support for LLMs
Haeun Lee (Seoul National University), Jae W. Lee (Seoul National University)
Large Language ModelText
🎯 What it does: This paper proposes NestedFP, a technique for FP8 and FP16 dual-precision LLM inference by overlapping 8-bit sub-weights within a single FP16 memory.
NeSyPr: Neurosymbolic Proceduralization For Efficient Embodied Reasoning
Wonje Choi (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
Computational EfficiencyRobotic IntelligenceTransformerLarge Language ModelContrastive LearningText
🎯 What it does: A neural symbolic procedural framework named NESYPR is proposed, which utilizes large language models to complete embedded task reasoning without an online symbolic planner.
Network two-sample test for block models
Chung Kyong Nguen (University of California), OSCAR HERNAN MADRID PADILLA (University of California)
Graph Neural NetworkGraph
🎯 What it does: A method for two-sample testing of networks without vertex correspondence and with variable node counts is proposed, based on the Stochastic Block Model (SBM) and spectral matching.
Neural Atlas Graphs for Dynamic Scene Decomposition and Editing
Jan Philipp Schneider (University of Siegen), Felix Heide (Princeton University)
SegmentationGenerationAutonomous DrivingGraph Neural NetworkOptical FlowVideo
🎯 What it does: This paper proposes Neural Atlas Graphs (NAG), a high-resolution representation method that splits dynamic scenes into editable 2D neural layers while maintaining a hierarchical structure in 3D space.
Neural Attention Search
Difan Deng (Leibniz University Hannover), Marius Lindauer (Leibniz University Hannover)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By learning the role of each token (global, local, sliding window) within the Transformer, a learnable sparse attention mask is constructed, which automatically determines which tokens can be discarded during inference, significantly reducing KV cache usage and computational load.
Neural B-frame Video Compression with Bi-directional Reference Harmonization
Yuxi Liu (Nanjing University), Zhan Ma (Nanjing University)
CompressionOptical FlowVideo
🎯 What it does: A B-frame video compression method based on neural networks, BRHVC, is proposed to specifically address the issue of imbalanced bidirectional reference contributions.
Neural Collapse in Cumulative Link Models for Ordinal Regression: An Analysis with Unconstrained Feature Model
Chuang Ma (Kyoto University), Toshiyuki Tanaka (Kyoto University)
ClassificationOptimizationTabular
🎯 What it does: This study investigates the phenomenon of Neural Collapse in Ordinal Regression tasks, proposing and theoretically proving the concept of Ordinal Neural Collapse (ONC) and its three main characteristics.
Neural Collapse is Globally Optimal in Deep Regularized ResNets and Transformers
Peter Súkeník (Institute of Science and Technology Austria), Marco Mondelli
OptimizationTransformerImageText
🎯 What it does: The paper proves that in sufficiently deep residual networks (ResNet) and transformers, when using cross-entropy or mean squared error loss along with weight regularization, the global optimal solution will approximately exhibit the geometric structure of neural collapse.
Neural Collapse under Gradient Flow on Shallow ReLU Networks for Orthogonally Separable Data
Hancheng Min (Shanghai Jiao Tong University), Rene Vidal
ClassificationOptimizationConvolutional Neural NetworkImageOrdinary Differential Equation
🎯 What it does: This paper studies the gradient flow training process of two-layer ReLU networks on orthogonally separable data and proves that it ultimately converges to a Neural Collapse solution.
Neural Combinatorial Optimization for Time Dependent Traveling Salesman Problem
Ruixiao Yang (Massachusetts Institute of Technology), Chuchu Fan (Massachusetts Institute of Technology)
OptimizationGraph Neural NetworkReinforcement LearningMixture of ExpertsTime Series
🎯 What it does: This study proposes a neural network solution framework for the Time-Dependent Traveling Salesman Problem (TDTSP), capable of simultaneously learning the dynamic structures of space and time, and improving solution quality through post-processing.
Neural Correlates of Serial Dependence: Synaptic Short-term Plasticity Orchestrates Repulsion and Attraction
Xiuning Zhang (Tsinghua University), Yuanyuan Mi (Tsinghua University)
Sequential
🎯 What it does: By constructing a dual-layer continuous attractor network and introducing synaptic short-term plasticity (inhibitory STD in the lower layer and facilitatory STF in the upper layer), this study explains perceptual repulsion and post-perceptual attraction in visual sequence dependence from a neural perspective.
Neural Emulator Superiority: When Machine Learning for PDEs Surpasses its Training Data
Felix Koehler (Technical University of Munich), Nils Thuerey (Technical University of Munich)
Convolutional Neural NetworkTransformerTime SeriesPhysics Related
🎯 What it does: The researchers proposed and validated the phenomenon of 'emulator superiority', which indicates that neural network simulators trained on data generated by low-fidelity numerical solutions can outperform their training sources under higher fidelity references in multi-step predictions.
Neural Entropy
Akhil Premkumar (Yale University)
GenerationData SynthesisCompressionDiffusion modelImageStochastic Differential Equation
🎯 What it does: This paper proposes and quantifies Neural Entropy to measure the amount of effective information stored by diffusion models during training. Experiments reveal its logarithmic growth characteristic with respect to the number of samples and further explore the thermodynamic uncertainty of the diffusion process and information compression efficiency.
Neural Evolution Strategy for Black-box Pareto Set Learning
Chengyu LU, Qingfu Zhang (City University of Hong Kong)
OptimizationBenchmark
🎯 What it does: This paper proposes a new black-box Pareto set learning (BPSL) framework that combines Pareto set learning with evolutionary strategies (ES), utilizing neural networks to model the distribution of decision variables, enabling the learning of Pareto sets for high-dimensional, multi-objective problems without the need for analytical expressions or gradients.
Neural Fractional Attention Differential Equations
Qiyu Kang (Anhui University), Zheng-Jun Zha (Anhui University)
Graph Neural NetworkTransformerGraphTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A framework called Fractional Attention Differential Equation (FADE) is proposed, which combines fractional calculus with learnable attention kernels to enhance memory effects.
Neural Green’s Functions
Seungwoo Yoo (Korea Advanced Institute of Science and Technology), Minhyuk Sung (Korea Advanced Institute of Science and Technology)
MeshGraphBenchmarkPhysics Related
🎯 What it does: Proposed Neural Green's Function, a neural solver for diagonalizable linear partial differential equations;
Neural Hamiltonian Diffusions for Modeling Structured Geometric Dynamics
Sungwoo Park (Korea University)
OptimizationProtein Structure PredictionTransformerTime SeriesSequentialPhysics RelatedStochastic Differential Equation
🎯 What it does: The Neural Hamiltonian Diffusion (NHD) framework is proposed, which uses neural networks to learn Hamiltonian dynamics with noise on curvature differences, periodic or relativistic geometric structures, and ensures coordinate invariance through horizontal lifts of the frame bundle.
Neural MJD: Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction
Yuanpei Gao (University of British Columbia), Renjie Liao (University of British Columbia)
Graph Neural NetworkTransformerTime SeriesFinance RelatedStochastic Differential Equation
🎯 What it does: This paper proposes Neural MJD, a parameterized non-stationary Merton jump-diffusion model using neural networks to predict time series with abrupt jumps.
Neural Mutual Information Estimation with Vector Copulas
Yanzhi Chen (University of Cambridge), Michael U. Gutmann (University of Edinburgh)
Flow-based ModelImageTextMultimodality
🎯 What it does: This paper proposes a mutual information estimator based on vector Copula, achieving more robust estimation by separating the marginal distribution and the dependence structure.
Neural Networks for Learnable and Scalable Influence Estimation of Instruction Fine-Tuning Data
Ishika Agarwal (University of Illinois Urbana-Champaign), Dilek Hakkani-Tür (University of Illinois Urbana-Champaign)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By training a small neural network (InfluenceNetwork) that accounts for only 0.0007% of the original large model's parameters, we efficiently estimate the influence values of the training data, enabling low-cost data subset selection in instruction fine-tuning.
Neural Rule Lists: Learning Discretizations, Rules, and Order in One Go
Sascha Xu (CISPA Helmholtz Center for Information Security), Jilles Vreeken (CISPA Helmholtz Center for Information Security)
ClassificationOptimizationSupervised Fine-TuningTabular
🎯 What it does: An end-to-end differentiable rule list learning framework called NEURULES is proposed, which can simultaneously learn feature thresholds, rule combinations, and rule rankings, ultimately resulting in a sparse and accurate rule list.
Neural Stochastic Flows: Solver-Free Modelling and Inference for SDE Solutions
Naoki Kiyohara (Imperial College London), Yingzhen Li (Imperial College London)
Data SynthesisOptimizationRecurrent Neural NetworkFlow-based ModelVideoTime SeriesSequentialStochastic Differential Equation
🎯 What it does: Proposes Neural Stochastic Flows (NSFs) and their extension in implicit state space models, utilizing conditional normal flows to directly learn the weak solution transition distribution of SDEs, eliminating the need for traditional numerical integration.
Neural Tangent Knowledge Distillation for Optical Convolutional Networks
Jinlin Xiang (University of Washington), Eli Shlizerman (University of Washington)
SegmentationKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: A complete pipeline for the design, training, and error compensation of an optical convolutional network (ONN) that is task-independent and hardware-independent is proposed.
Neural Thermodynamics: Entropic Forces in Deep and Universal Representation Learning
Liu Ziyin (Massachusetts Institute of Technology), Isaac L. Chuang
OptimizationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: A neural thermodynamics framework is proposed and validated, using entropy loss and symmetry analysis to explain the entropy force and symmetry breaking in the dynamics of SGD learning, unifying phenomena such as gradient balance, universal representation, and sharpening/flattening.
Neural-Driven Image Editing
Pengfei Zhou (National University of Singapore), Yang You (National University of Singapore)
GenerationData SynthesisTransformerDiffusion modelContrastive LearningImageMultimodality
🎯 What it does: LoongX is proposed, a hands-free image editing framework driven by multimodal neural signals.
NeuralPLexer3: Accurate Biomolecular Complex Structure Prediction with Flow Models
Zhuoran Qiao (Iambic Therapeutics), Matthew Welborn (Iambic Therapeutics)
Protein Structure PredictionTransformerFlow-based ModelBiomedical Data
🎯 What it does: Developed NeuralPLexer3, a method for predicting the structures of biomolecular complexes based on flow models.
NeuralSurv: Deep Survival Analysis with Bayesian Uncertainty Quantification
Mélodie Monod (Imperial College London), Samir Bhatt (Imperial College London)
Biomedical Data
🎯 What it does: We propose NeuralSurv, a framework capable of deep Bayesian modeling of survival data over continuous time;
NeurIPT: Foundation Model for Neural Interfaces
Zitao Fang (Xiamen University Malaysia), Sim Kuan Goh (Xiamen University Malaysia)
TransformerMixture of ExpertsTime SeriesBiomedical Data
🎯 What it does: NEURIPT is proposed, a foundational model for diversifying EEG-based neural interfaces, aimed at learning robust and transferable representations across different settings.
Neuro-Spectral Architectures for Causal Physics-Informed Networks
Arthur Bizzi (École Polytechnique Fédérale de Lausanne), Lucas Nissenbaum (Instituto de Matemática Pura e Aplicada)
OptimizationTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper presents NeuSA—a physics-informed network based on spectral decomposition and neural ordinary differential equations (ODEs) for solving time-varying nonlinear partial differential equations (PDEs), particularly wave equations, Burgers' equations, and the sine-Gordon equation.
NeuroGenPoisoning: Neuron-Guided Attacks on Retrieval-Augmented Generation of LLM via Genetic Optimization of External Knowledge
Hanyu Zhu (University of Massachusetts Dartmouth), Long Jiao (University of Massachusetts Dartmouth)
RetrievalOptimizationAdversarial AttackTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Proposes NeuroGenPoisoning, an external knowledge attack framework for RAG systems generated through neuron attribution and genetic algorithms;
NeuroH-TGL: Neuro-Heterogeneity Guided Temporal Graph Learning Strategy for Brain Disease Diagnosis
Shengrong Li (Nanjing University of Aeronautics and Astronautics), Daoqiang Zhang (Nanjing University of Aeronautics and Astronautics)
ClassificationGraph Neural NetworkTime SeriesBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: This study separates dynamic functional brain networks into two categories: topological consistency and temporal trends, exploring spatial and temporal heterogeneous nodes, and utilizes temporal propagation graph convolutional networks for brain disease diagnosis.
Neurons as Detectors of Coherent Sets in Sensory Dynamics
Joshua L. Pughe-Sanford (Flatiron Institute), Dmitri Chklovskii
Time SeriesStochastic Differential Equation
🎯 What it does: This paper proposes viewing sensory neurons as self-supervised learners, utilizing the subdominant singular functions of the Stochastic Koopman Operator (SKO) to identify coherent sets in input time series, and assigning attribution indices to new stimuli by projecting onto the corresponding singular vectors' signs or rectified values, thereby explaining the temporal receptive fields of neurons, ON/OFF differentiation, and the functions of predictive and retrospective cells.
NeuroPath: Neurobiology-Inspired Path Tracking and Reflection for Semantically Coherent Retrieval
Junchen Li (University of Electronic Science and Technology of China), Shuang Liang (University of Electronic Science and Technology of China)
RetrievalLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: NeuroPath is proposed, a retrieval-augmented generation framework based on the mechanism of place cells in neurobiology, utilizing dynamic semantic path tracking and post-retrieval to achieve multi-hop question answering.
Neurosymbolic Diffusion Models
Emile van Krieken (University of Edinburgh), Antonio Vergari (University of Edinburgh)
GenerationOptimizationReinforcement LearningDiffusion modelImage
🎯 What it does: This paper proposes the Neural Symbolic Diffusion Model (NESYDMS), which combines neural symbolic predictors with symbolic programs through a discrete masking diffusion process to address the issues of dependency and uncertainty between concepts.
NeuSymEA: Neuro-symbolic Entity Alignment via Variational Inference
Shengyuan Chen (Hong Kong Polytechnic University), Xiao Huang (Hong Kong Polytechnic University)
OptimizationExplainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: This paper proposes NeuSymEA, a unified neural-symbolic entity alignment framework designed to integrate symbolic reasoning with neural embeddings for cross-knowledge graph entity alignment.
New Parallel and Streaming Algorithms for Directed Densest Subgraph
Slobodan Mitrović (University of California), Mohammad Amin Raeisi (Yale University)
OptimizationGraph Neural NetworkGraphTime Series
🎯 What it does: This paper proposes parallel and streaming algorithms for approximating the directed densest subgraph under the MPC (sublinear memory) and single-channel semi-streaming models, achieving 2+ε approximation and O(log n) approximation, respectively, and provides a dynamic update scheme.
New Perspectives on the Polyak Stepsize: Surrogate Functions and Negative Results
Francesco Orabona (King Abdullah University of Science and Technology), Ryan D'Orazio (Mila Québec AI Institute)
Optimization
🎯 What it does: This paper presents a unified perspective, viewing the Polyak step size as gradient descent on a surrogate objective function, and explains its adaptability and convergence properties; it further generalizes to a more general surrogate ψ(x)=½h(x)², providing theoretical analysis of convergence and negative results in both deterministic and stochastic settings; it also summarizes and improves upon various existing Polyak variants (such as SPS max, SPS+, SPS ℓ max), and provides corresponding convergence bounds and examples of non-convergence.
Next Semantic Scale Prediction via Hierarchical Diffusion Language Models
Cai Zhou (Massachusetts Institute of Technology), Tommi Jaakkola (Massachusetts Institute of Technology)
GenerationDiffusion modelText
🎯 What it does: A Hierarchical Diffusion Language Model (HDLM) is proposed, which improves the performance of discrete diffusion models in text generation by constructing a hierarchical vocabulary to achieve gradual predictions from coarse to fine semantic scales.
NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering
Zhihao Huang (Northwest Polytechnical University), Xuelong Li (TeleAI, China Telecom)
GenerationData SynthesisTransformerAuto EncoderImage
🎯 What it does: A frequency decomposition-based autoregressive image generation framework NFIG is proposed, which first generates low-frequency global structures and then gradually adds high-frequency details, implementing frequency-guided residual quantization VAE as an image tokenizer.
NFL-BA: Near-Field Light Bundle Adjustment for SLAM in Dynamic Lighting
Andrea Dunn Beltran, Roni Sengupta
Pose EstimationDepth EstimationOptimizationNeural Radiance FieldGaussian SplattingSimultaneous Localization and MappingVideo
🎯 What it does: Proposed the Near-Field Light Bundle Adjustment (NFL-BA) loss to improve camera tracking and reconstruction of neural rendering-based SLAM in dynamic co-located light source environments (such as endoscopy).
No Experts, No Problem: Avoidance Learning from Bad Demonstrations
Huy Hoang (Singapore Management University), Pradeep Varakantham (Singapore Management University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This study investigates an offline imitation learning method that avoids undesirable behaviors by training with bad examples and unlabeled data in the absence of expert demonstrations.
No Loss, No Gain: Gated Refinement and Adaptive Compression for Prompt Optimization
Wenhang Shi (Renmin University of China), Xiaoyong Du (Renmin University of China)
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an automated prompt optimization framework named GRACE, which utilizes two information loss strategies: gated refinement and adaptive compression, to iteratively update prompts and enhance the performance of large language models in downstream tasks.
No Object Is an Island: Enhancing 3D Semantic Segmentation Generalization with Diffusion Models
Fan Li (Northwestern Polytechnical University), Yuelei Xu (Northwestern Polytechnical University)
SegmentationDomain AdaptationAutonomous DrivingDiffusion modelPoint Cloud
🎯 What it does: Proposes the XDiff3D framework, which utilizes diffusion models to generate object proxy queries, enhancing the cross-domain generalization performance of 3D semantic segmentation.
No-Regret Learning Under Adversarial Resource Constraints: A Spending Plan Is All You Need!
Francesco Emanuele Stradi (Politecnico di Milano), Christian Kroer (Columbia University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: In budget-constrained online decision-making problems, the author proposes achieving no-regret learning using a given consumption plan, even when the reward and cost distributions change arbitrarily over time.
No-Regret Online Autobidding Algorithms in First-price Auctions
Yilin LI, Hanrui Zhang (Chinese University of Hong Kong)
Optimization
🎯 What it does: A no-regret online algorithm for the autobidding problem with ROI constraints in first-price auctions has been designed and implemented, providing theoretically optimal or approximately optimal scheduling strategies under both full information and Bandit information feedback models.
No-Regret Thompson Sampling for Finite-Horizon Markov Decision Processes with Gaussian Processes
Jasmine Bayrooti (University of Cambridge), Carl Henrik Ek (University of Cambridge)
Reinforcement Learning
🎯 What it does: A reinforcement learning algorithm RL-GPS based on multi-output Gaussian processes (GP) is proposed. The algorithm samples the reward and transition function from the GP posterior at the beginning of each round, then derives the agent's value function through value iteration and executes a greedy policy.
NOBLE - Neural Operator with Biologically-informed Latent Embeddings to Capture Experimental Variability in Biological Neuron Models
Luca Ghafourpour (ETH Zurich), Anima Anandkumar (California Institute of Technology)
GenerationData SynthesisOptimizationSpiking Neural NetworkNeural Radiance FieldBiomedical Data
🎯 What it does: Proposed the NOBLE framework, which uses neural operators combined with interpretable latent embeddings to predict neuronal membrane potential responses and generate synthetic neurons that can capture experimental variability.
Noise Consistency Training: A Native Approach for One-step Generator in Learning Additional Controls
Yihong Luo (Hong Kong University of Science and Technology), Jing Tang (Hong Kong University of Science and Technology)
GenerationData SynthesisOptimizationDiffusion modelImage
🎯 What it does: This paper proposes Noise Consistency Training (NCT), which injects new control signals in a lightweight manner through noise space consistency loss and boundary loss on first-order diffusion generators, without the need to retrain the diffusion model or additional diffusion distillation.
Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models
Luca Eyring (Technical University of Munich), Zeynep Akata (Technical University of Munich)
GenerationOptimizationComputational EfficiencyKnowledge DistillationReinforcement LearningDiffusion modelImage
🎯 What it does: This paper proposes the Noise Hypernetworks (HyperNoise) method, which learns a lightweight noise modulation network to directly generate optimized initial noise during inference, thereby transferring the computational load of traditional noise optimization at test time to the training phase, achieving 'one-time training and multiple high-quality generations.'
Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models
Cameron Tice (Geodesic Research), Teun van der Weij (Independent)
TransformerLarge Language ModelText
🎯 What it does: By injecting Gaussian noise of varying amplitudes into the language model weights, this study investigates noise injection methods to detect and recover the model's strategic underperformance behavior, and validates its effectiveness across various models and benchmarks.
Noise Matters: Optimizing Matching Noise for Diffusion Classifiers
Yanghao Wang (Hong Kong University of Science and Technology), Long Chen (Hong Kong University of Science and Technology)
ClassificationOptimizationDiffusion modelImage
🎯 What it does: This paper studies the noise instability of the Diffusion Classifier (DC) and proposes improving classification performance through learning noise matching (Noise Optimization, NoOp);
Noise-Robustness Through Noise: A Framework combining Asymmetric LoRA with Poisoning MoE
Zhaokun Wang (University of Electronic Science and Technology of China), Wenhong Tian (University of Electronic Science and Technology of China)
TransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: This paper proposes a noise injection method that introduces a specialized 'poisoned expert' in a mixture of experts network to achieve noise-robust and parameter-efficient fine-tuning method called LoPE.
Noisy Multi-Label Learning through Co-Occurrence-Aware Diffusion
Senyu Hou (Shanxi University), Wenjian Wang (Shanxi University)
ClassificationData-Centric LearningDiffusion modelImage
🎯 What it does: The Co-Occurrence-Aware Diffusion (CAD) model is proposed, treating noisy multi-label learning as a generative task from features to true multi-labels. It achieves label denoising through a diffusion model and guides learning using neighborhood label estimation and label co-occurrence constraints.
NoisyGRPO: Incentivizing Multimodal CoT Reasoning via Noise Injection and Bayesian Estimation
Longtian Qiu (ShanghaiTech University), Xuming He (ShanghaiTech University)
Reinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelDiffusion modelMultimodalityChain-of-Thought
🎯 What it does: In the multimodal reinforcement learning framework, controllable noise is injected into visual inputs to enhance exploration, using noise levels as priors and trajectory rewards as likelihoods, and Bayesian advantage estimation is employed to optimize the chain-of-thought reasoning capabilities of multimodal large language models.
NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation
Xiangyan Liu (National University of Singapore), Michael Qizhe Shieh (National University of Singapore)
Reinforcement LearningVision Language ModelImage
🎯 What it does: By incorporating NoisyRollout during the reinforcement learning fine-tuning process of the visual language model (VLM), the model's exploration capability and robustness are enhanced by mixing clear images and moderately distorted images in the reasoning trajectories while keeping the original RL objective unchanged.
Non-Adaptive Adversarial Face Generation
Sunpill Kim (Hanyang University), Jae Hong Seo (Hanyang University)
GenerationAdversarial AttackImage
🎯 What it does: A non-iterative, non-adaptive adversarial face generation method is proposed, achieving one-time attacks by utilizing attribute sub-balls in the feature space.
Non-Asymptotic Analysis Of Data Augmentation For Precision Matrix Estimation
Lucas Morisset (École Polytechnique), Alain Oliviero Durmus
OptimizationHyperparameter SearchImage
🎯 What it does: This paper studies the estimation of the inverse of the covariance matrix (precision matrix) using data augmentation (DA) and linear shrinkage methods in high-dimensional data scenarios. It provides a non-asymptotic concentration bound based on random matrix theory, which allows for direct estimation of quadratic errors from the data and is used for hyperparameter tuning.
Non-Asymptotic Guarantees for Average-Reward Q-Learning with Adaptive Stepsizes
Zaiwei Chen (Purdue University)
Reinforcement Learning
🎯 What it does: This paper presents the first finite-time analysis of average reward Q-learning with asynchronous implementation, studying an algorithm that uses adaptive step sizes, and proving that the mean square error of the algorithm converges at a specific rate.
Non-Clairvoyant Scheduling with Progress Bars
Ziyad Benomar (CREST), Jens Schlöter (Centrum Wiskunde en Informatica)
OptimizationTabular
🎯 What it does: A progress bar model is proposed to study non-prior scheduling problems, designing online algorithms suitable for aggressive and stochastic progress bars and providing competitive ratio analysis.
Non-convex entropic mean-field optimization via Best Response flow
Razvan-Andrei Lascu (RIKEN AIP), Mateusz B. Majka (Heriot-Watt University)
Optimization
🎯 What it does: The study addresses the problem of minimizing non-convex functions (with relative entropy regularization) in probability measure spaces and employs Best Response flows for the solution; it also extends to non-convex non-concave max-min problems.
Non-Convex Tensor Recovery from Tube-Wise Sensing
Tongle Wu (Pennsylvania State University), Ying Sun (Pennsylvania State University)
CompressionOptimizationTabular
🎯 What it does: A tube-based local tensor compression sensing model based on t-SVD is proposed, and the exact recovery of low tensor rank tensors is achieved through Burer–Monteiro decomposition and gradient descent.
Non-equilibrium Annealed Adjoint Sampler
Jaemoo Choi (Georgia Institute of Technology), Guan-Horng Liu (Meta)
OptimizationTabularBenchmarkStochastic Differential Equation
🎯 What it does: This paper proposes a Non-Equilibrium Annealing Adjoint Sampler (NAAS), which achieves unbiased and efficient sampling by introducing an annealing reference SDE within the SOC framework and employing dual matching and reverse dual matching.
Non-exchangeable Conformal Prediction with Optimal Transport: Tackling Distribution Shift with Unlabeled Data
Alvaro Correia, Christos Louizos (Qualcomm AI Research)
Domain AdaptationOptimizationImageTabular
🎯 What it does: Using optimal transport theory, a method is proposed to define and compensate for the coverage gap in synthetic predictions under distribution shift, and to reduce the coverage gap by learning weighted calibration samples.
Non-Line-of-Sight 3D Reconstruction with Radar
Haowen Lai (University of Pennsylvania), Mingmin Zhao (University of Pennsylvania)
Object DetectionSegmentationDepth EstimationConvolutional Neural NetworkSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper studies a system called HoloRadar that uses millimeter-wave radar to achieve full Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) 3D scene reconstruction, capable of reconstructing hidden structures and people in indoor corner scenarios.
Non-Markovian Discrete Diffusion with Causal Language Models
Yangtian Zhang (Yale University), David van Dijk (Yale University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelTextSequential
🎯 What it does: This paper proposes a non-Markovian discrete diffusion model called CaDDi, which eliminates the error accumulation problem caused by a single state in traditional discrete diffusion models by conditioning each denoising step on the entire generation trajectory.
Non-monotone Submodular Optimization: $p$-Matchoid Constraints and Fully Dynamic Setting
Kiarash Banihashem (University of Maryland), Morteza Monemizadeh (TU Eindhoven)
Optimization
🎯 What it does: This paper proposes a dynamic algorithm for the non-monotonic submodular maximization problem under p-matchoid constraints in a fully dynamic environment, which can efficiently maintain an approximately optimal solution after each element insertion or deletion.
Non-rectangular Robust MDPs with Normed Uncertainty Sets
Navdeep Kumar (Technion), Shie Mannor (Technion)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a robust MDP under non-rectangular Lp-bounded uncertainty sets and provides an equivalence relation that can be decomposed into infinitely many sa-rectangular sets, utilizing binary search to achieve polynomial-time robust policy evaluation.
Non-Singularity of the Gradient Descent Map for Neural Networks with Piecewise Analytic Activations
Alexandru Crăciun (Technical University of Munich), Debarghya Ghoshdastidar (Technical University of Munich)
Optimization
🎯 What it does: It is proven that neural networks using piecewise analytic activation functions have non-singular gradient descent (GD) and stochastic gradient descent (SGD) mappings for almost all step sizes (the inverse image of a zero-measure set is also a zero-measure set).
Non-stationary Bandit Convex Optimization: A Comprehensive Study
Xiaoqi Liu (University of Oxford), Arya Akhavan (École Polytechnique de Paris)
OptimizationHyperparameter Search
🎯 What it does: The paper studies bandit convex optimization problems in non-stationary environments and proposes two algorithms, TEWA-SE and cExO, aimed at minimizing regret under different non-stationary metrics.
Non-stationary Equivariant Graph Neural Networks for Physical Dynamics Simulation
Chaohao Yuan (Tsinghua University), Yu Rong (Alibaba Group)
Graph Neural NetworkGraphTime SeriesSequentialPhysics Related
🎯 What it does: This paper proposes a non-stationary equivariant graph neural network, NS-EGNN, which can capture the spectral features of time-varying dynamics while maintaining E(3) equivariance and performing multi-step predictions on sequences.
Non-Stationary Lipschitz Bandits
Nicolas Nguyen (University of Tübingen), Claire Vernade (University of Tübingen)
Reinforcement LearningTime Series
🎯 What it does: This paper studies the non-stationary bandit problem with continuous action spaces and rewards that satisfy the Lipschitz condition, and proposes an adaptive algorithm called MDBE.
Non-Stationary Structural Causal Bandits
Yeahoon Kwon (Seoul National University), Sanghack Lee (Seoul National University)
Reinforcement LearningTime Series
🎯 What it does: A new framework that integrates non-stationary multi-armed bandits with structural causal models (NS-SCM-MAB) is proposed, along with a non-greedy intervention set POMIS+ designed to capture long-term causal effects and its solving algorithm.
Non-Uniform Multiclass Learning with Bandit Feedback
Steve Hanneke (Purdue University), Hongao Wang (Purdue University)
Classification
🎯 What it does: This paper studies a non-uniform learning framework in multi-class learning with bandit feedback and an infinitely large label space. It provides necessary and sufficient combinatorial characterizations of PAC and online learnability, and constructs a hypothesis class that is learnable under full supervision but not learnable under bandit feedback, thereby revealing the fundamental differences between non-uniform and general learning frameworks.
Nonlinear Laplacians: Tunable principal component analysis under directional prior information
Yuxin Ma (Johns Hopkins University), Dmitriy Kunisky (Johns Hopkins University)
Optimization
🎯 What it does: A class of nonlinear Laplacian spectral algorithms is proposed for detecting and recovering low-rank signals in sparse first-order PCA models with directional priors.
Nonlinearly Preconditioned Gradient Methods: Momentum and Stochastic Analysis
Konstantinos Oikonomidis (KU Leuven), Panagiotis Patrinos (KU Leuven)
Recommendation SystemOptimizationConvolutional Neural NetworkImageTabular
🎯 What it does: This paper proposes a gradient descent method based on nonlinear preprocessing, and introduces heavy-ball momentum and its stochastic variant, providing convergence analysis for general non-convex smooth optimization problems.
Nonparametric Quantile Regression with ReLU-Activated Recurrent Neural Networks
Hang Yu (Nanjing University), Zhao Ren (University of Pittsburgh)
Recurrent Neural NetworkTabularTime SeriesFinance Related
🎯 What it does: This study investigates the theoretical convergence properties of ReLU-activated Recurrent Neural Networks (RNN) and Sparse RNN (SRNN) in nonparametric quantile regression.
NopeRoomGS: Indoor 3D Gaussian Splatting Optimization without Camera Pose Input
Wenbo Li (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
Pose EstimationDepth EstimationOptimizationGaussian SplattingSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: A 3D Gaussian Splatting framework without camera pose priors, called NopeRoomGS, is proposed. It employs a local-to-global optimization process, first jointly optimizing depth and camera pose using neural geometric representation within short video clips, and then fusing the obtained local Gaussian point clouds into a global 3DGS. High-quality camera pose estimation and view synthesis are achieved through alternating optimization and planar constraints.
NoPo-Avatar: Generalizable and Animatable Avatars from Sparse Inputs without Human Poses
Jing Wen (University of Illinois Urbana-Champaign), Shenlong Wang (University of Illinois Urbana-Champaign)
Image TranslationGenerationTransformerGaussian SplattingImage
🎯 What it does: Reconstructing animatable 3D avatars from sparse input images without using any pose or camera pose information.
Normal-Abnormal Guided Generalist Anomaly Detection
Yuexin Wang (Xi'an Jiaotong-Liverpool University), Jimin XIAO
Anomaly DetectionTransformerContrastive LearningImageBenchmark
🎯 What it does: A general anomaly detection framework NAGL based on normal-anomalous sample guidance is proposed, which can detect unknown anomalies across domains after training in the original domain.
Normalization in Attention Dynamics
Nikita Karagodin (Massachusetts Institute of Technology), Philippe Rigollet (Massachusetts Institute of Technology)
TransformerOrdinary Differential Equation
🎯 What it does: This paper studies the impact of different normalization schemes (Post‑LN, Pre‑LN, Mix‑LN, Peri‑LN, nGPT, LN‑Scaling) on deep representations by viewing the evolution of token representations in the Transformer as a particle system on a sphere, and interprets normalization as a velocity regulation mechanism.
Normalize Filters! Classical Wisdom for Deep Vision
Gustavo Perez (University of California), Stella X. Yu (University of Michigan)
ClassificationRecognitionConvolutional Neural NetworkTransformerImage
🎯 What it does: A Filter Normalization method is proposed, which normalizes the convolution kernel weights and then adds learnable scale and offset, achieving equivariance under intensity variations such as lighting and atmospheric transmission, thereby enhancing the robustness of deep visual models.
Normalized Attention Guidance: Universal Negative Guidance for Diffusion Models
Dar-Yen Chen (University of Surrey), Yi-Zhe Song (University of Surrey)
GenerationData SynthesisDiffusion modelImageVideoText
🎯 What it does: A training-free Normalized Attention Guidance (NAG) method is proposed, achieving robust negative guidance in diffusion models, especially for few-step sampling.
Normalizing Flows are Capable Models for Continuous Control
Raj Ghugare (Princeton University), Benjamin Eysenbach (Princeton University)
Robotic IntelligenceReinforcement LearningFlow-based Model
🎯 What it does: A simple and efficient Normalizing Flow (NF) architecture is proposed as a universal probabilistic model for various algorithms in reinforcement learning (behavior cloning, offline RL, goal-conditioned RL, unsupervised RL), and its feasibility is validated on 82 tasks.
NormFit: A Lightweight Solution for Few-Shot Federated Learning with Non-IID Data
Azadeh Motamedi (Queen's University), Il-Min Kim (Queen's University)
ClassificationFederated LearningPrompt EngineeringVision Language ModelImage
🎯 What it does: Proposes NormFit, which fine-tunes only the Pre-LayerNorm of the VLM visual encoder in federated learning for few-shot adaptation.
Not All Data are Good Labels: On the Self-supervised Labeling for Time Series Forecasting
Yuxuan Yang (Zhejiang University), Huan Li (Zhejiang University)
TransformerSupervised Fine-TuningTime Series
🎯 What it does: A self-supervised label reconstruction method SCAM is proposed, which enhances time series prediction performance by replacing overfitted labels with pseudo-labels through adaptive masking.
Novel Class Discovery for Point Cloud Segmentation via Joint Learning of Causal Representation and Reasoning
Yang Li (Tianjin University), Yahong Han (Tianjin University)
SegmentationRepresentation LearningGraph Neural NetworkGenerative Adversarial NetworkImagePoint Cloud
🎯 What it does: A method for joint learning of causal representation and reasoning is proposed for new category discovery in point cloud semantic segmentation.
Novel Exploration via Orthogonality
Andreas Theophilou (University of Bath), Özgür Şimşek (University of Bath)
Robotic IntelligenceGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A novel exploration method based on the source Laplacian matrix (NEO) is proposed, utilizing its smoothest eigenvector to construct multi-step exploration options that guide agents to under-explored states.
Novel View Synthesis from A Few Glimpses via Test-Time Natural Video Completion
Yan Xu (University of Michigan), Stella X. Yu (University of Michigan)
RestorationGenerationData SynthesisDiffusion modelGaussian SplattingVideoBenchmark
🎯 What it does: A zero-shot, generation-guided 3D Gaussian Splatting training framework is proposed, utilizing a pre-trained video diffusion model to generate intermediate views under sparse input to enhance sparse scene reconstruction.
NPN: Non-Linear Projections of the Null-Space for Imaging Inverse Problems
Roman Jacome (Universidad Industrial de Santander), Henry Arguello (Universidad Industrial de Santander)
RestorationSuper ResolutionCompressionConvolutional Neural NetworkDiffusion modelImageMagnetic Resonance ImagingComputed Tomography
🎯 What it does: The paper proposes the NPN (Non-Linear Projections of the Null-Space) regularization method, which utilizes low-dimensional nonlinear projections of the null space of the sensing matrix H to constrain the solution of inverse problems.
NSNQuant: A Double Normalization Approach for Calibration-Free Low-Bit Vector Quantization of KV Cache
Donghyun Son (Seoul National University), Sungjoo Yoo (Seoul National University)
CompressionTransformerLarge Language ModelText
🎯 What it does: NSNQuant is proposed, a calibration-free low-bit vector quantization method for compressing the KV cache of LLMs.
NTKMTL: Mitigating Task Imbalance in Multi-Task Learning from Neural Tangent Kernel Perspective
Xiaohan Qin (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Reinforcement LearningImage
🎯 What it does: This paper proposes a multi-task learning method based on Neural Tangent Kernel (NTK) theory, called NTKMTL, to alleviate task imbalance.
NUTS: Eddy-Robust Reconstruction of Surface Ocean Nutrients via Two-Scale Modeling
Hao Zheng (Shanghai Jiaotong University), Enhui Liao (Shanghai Jiaotong University)
Transformer
🎯 What it does: A two-scale model NUTS has been developed to reconstruct sea surface nutrient concentrations from sparse observations and noisy currents.