ICLR 2025 Papers — Page 23
International Conference on Learning Representations · 3704 papers
Multi-Reward as Condition for Instruction-based Image Editing
Xin Gu (ByteDance Inc), Sijie Zhu (ByteDance Inc)
Image TranslationGenerationTransformerLarge Language ModelDiffusion modelImageTextBenchmark
🎯 What it does: This paper improves the training quality of instruction-based image editing models by introducing multi-view reward data as an additional condition and constructs the Real-Edit evaluation benchmark.
Multi-Robot Motion Planning with Diffusion Models
Yorai Shaoul (Carnegie Mellon University), Maxim Likhachev (Carnegie Mellon University)
OptimizationRobotic IntelligenceDiffusion modelMultimodality
🎯 What it does: Developed MMD, a multi-robot motion planner that combines diffusion models with MAPF search algorithms, utilizing single-robot data to generate collision-safe and data-compliant trajectories.
Multi-Scale Fusion for Object Representation
Rongzhen Zhao (Aalto University), Joni Pajarinen (Aalto University)
Object DetectionRepresentation LearningTransformerAuto EncoderImageVideo
🎯 What it does: Designed and implemented a multi-scale fusion (MSF) technique that utilizes image pyramids and cross-scale, intra-scale quantization fusion to enhance the quality of object representations in VAE-guided Object Center Learning (OCL).
Multi-session, multi-task neural decoding from distinct cell-types and brain regions
Mehdi Azabou (Georgia Institute of Technology), Blake Aaron Richards
TransformerBiomedical Data
🎯 What it does: Using a large-scale multi-task Transformer model (POYO+) to train on Allen Brain Observatory data, a neural decoder is constructed that can span across brain regions and cell subtypes;
Multi-Task Corrupted Prediction for Learning Robust Audio-Visual Speech Representation
Sungnyun Kim (KAIST), Se-Young Yun (KAIST)
Knowledge DistillationRepresentation LearningTransformerVideoMultimodalityAudio
🎯 What it does: A self-supervised multimodal speech representation learning framework called CAV2vec is proposed, specifically designed for robust training against audio-visual joint distortion.
Multi-Task Dense Predictions via Unleashing the Power of Diffusion
Yuqi Yang (Nankai University), Bo Li (vivo Mobile Communication Co. Ltd)
SegmentationDepth EstimationTransformerDiffusion modelImage
🎯 What it does: This paper proposes TaskDiffusion—a multi-task dense prediction framework based on a joint denoising diffusion process, which achieves multi-task collaborative inference through cross-task label encoding and a cross-task diffusion decoder.
Multiagent Finetuning: Self Improvement with Diverse Reasoning Chains
Vighnesh Subramaniam (Massachusetts Institute of Technology), Igor Mordatch (University of California Berkeley)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: A multi-agent fine-tuning framework is proposed, which allows a group of homogeneous language models to interact with each other, focusing on either generation or evaluation, to build a diverse dataset and perform iterative fine-tuning for self-improvement.
Multilevel Generative Samplers for Investigating Critical Phenomena
Ankur Singha (BIFOLD), Shinichi Nakajima (BIFOLD)
GenerationData SynthesisOptimizationSequentialPhysics Related
🎯 What it does: This paper proposes the Renormalization-informed Generative Critical Sampler (RiGCS), which introduces a conditionally generative model with adjustable receptive fields within the multi-layer Monte Carlo + Heat Bath (MLMC-HB) framework to efficiently sample critical two-dimensional Ising systems while capturing long-range and higher-order interactions under scale invariance (SIC).
Multimodal Large Language Models for Inverse Molecular Design with Retrosynthetic Planning
Gang Liu (University of Notre Dame), Jie Chen (IBM Research)
OptimizationDrug DiscoveryGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityGraph
🎯 What it does: A multi-modal large language model called Llamole is proposed, capable of alternately generating between text and molecular graphs, achieving controllable reverse molecular design and retrosynthetic planning.
Multimodal Lego: Model Merging and Fine-Tuning Across Topologies and Modalities in Biomedicine
Konstantin Hemker (University of Cambridge), Mateja Jamnik (University of Cambridge)
Anomaly DetectionDrug DiscoverySupervised Fine-TuningMultimodalityBiomedical Data
🎯 What it does: A multi-modal learning framework MM-Lego has been developed, which can package any single-modal encoder through LegoBlock to achieve model fusion and merging with little or no fine-tuning.
Multimodal Quantitative Language for Generative Recommendation
Jianyang Zhai (Sun Yat-sen University), Yonghong Tian (Peking University)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderTextMultimodality
🎯 What it does: This study proposes a generative recommendation framework MQL4GRec that unifies multimodal product information (text, images) into a quantitative language mapping, and achieves cross-domain multimodal knowledge transfer through pre-training and fine-tuning.
Multimodal Situational Safety
Kaiwen Zhou (University of California), Xin Eric Wang (University of California)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A Multimodal Situational Safety (MSS) assessment framework is proposed, and an MSSBench benchmark containing chat and embodied scenarios is constructed to systematically evaluate the ability of current multimodal large language models (MLLMs) to assess query safety based on visual context.
Multimodal Unsupervised Domain Generalization by Retrieving Across the Modality Gap
Christopher Liao (Boston University), Brian Kulis (Boston University)
ClassificationRetrievalDomain AdaptationContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a multimodal unsupervised domain generalization framework MUDG, which retrieves source data in the CLIP visual-text space and pseudo-labels it, followed by diversified sampling and fine-tuning of the retrieved data, thereby achieving better cross-domain performance on unlabeled tasks.
Multimodality Helps Few-shot 3D Point Cloud Semantic Segmentation
Zhaochong An (University of Copenhagen), Serge Belongie (ETH Zurich)
SegmentationTransformerVision Language ModelMultimodalityPoint Cloud
🎯 What it does: Under the condition of no additional labeling cost, the performance of few-shot 3D point cloud semantic segmentation is improved by utilizing text labels and implicit 2D image modalities.
Multiple Heads are Better than One: Mixture of Modality Knowledge Experts for Entity Representation Learning
Yichi Zhang (Zhejiang University), Huajun Chen (Zhejiang University)
Representation LearningMixture of ExpertsMultimodality
🎯 What it does: Proposes the MOMOK framework, which uses relation-guided multimodal knowledge experts (ReMoKE) to learn multi-perspective, relation-aware modality embeddings, and achieves better multimodal knowledge graph completion (MMKGC) through multimodal joint decision-making (MuJoD) and expert information decoupling (ExID).
Multiplicative Logit Adjustment Approximates Neural-Collapse-Aware Decision Boundary Adjustment
Naoya Hasegawa (University of Tokyo), Issei Sato (University of Tokyo)
ClassificationRecognitionImageTabular
🎯 What it does: Proves that the Multiple Logarithmic Adjustment (MLA) is approximately the optimal decision boundary through Neural Collapse (NC) theory, and validates its effectiveness in long-tail recognition.
Multiview Equivariance Improves 3D Correspondence Understanding with Minimal Feature Finetuning
Yang You (Stanford University), Leonidas Guibas (Stanford University)
Object TrackingSegmentationPose EstimationTransformerSupervised Fine-TuningImageVideo
🎯 What it does: Systematically evaluate the 3D equivariance of ViT and enhance 3D correspondence through fine-tuning with multi-view alignment loss, significantly improving the performance of tasks such as pose estimation, video tracking, and semantic correspondence.
MuPT: A Generative Symbolic Music Pretrained Transformer
Xingwei Qu (M-A-P), Ge Zhang
GenerationTransformerLarge Language ModelAudio
🎯 What it does: A long-context pre-trained music generation model MuPT based on ABC Notation is proposed, introducing the Synchronized Multi-Track ABC (SMT-ABC) representation and SMS scale law, providing a scalable LLM solution for symbolic music generation.
MUSE: Machine Unlearning Six-Way Evaluation for Language Models
Weijia Shi (University of Washington), Chiyuan Zhang (Google Research)
Large Language ModelTextBenchmark
🎯 What it does: MUSE is proposed, a six-dimensional benchmark for evaluating the machine unlearning effects of language models.
MuseGNN: Forming Scalable, Convergent GNN Layers that Minimize a Sampling-Based Energy
Haitian Jiang (New York University), David Wipf (Amazon Web Services)
Graph Neural NetworkGraph
🎯 What it does: This paper presents MuseGNN, a scalable unfolded graph neural network that achieves large-scale graph training by embedding offline subgraph sampling into the energy function.
Mutual Effort for Efficiency: A Similarity-based Token Pruning for Vision Transformers in Self-Supervised Learning
Sheng Li (University of Pittsburgh), Geng Yuan (University of Georgia)
Computational EfficiencyRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: This paper proposes a ViT token pruning method for self-supervised learning (SSL) called SimPrune;
Mutual Reasoning Makes Smaller LLMs Stronger Problem-Solver
Zhenting Qi (Microsoft Research Asia), Mao Yang (Microsoft Research Asia)
TransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper presents rStar, a generative-discriminative process that enhances reasoning performance in small language models (SLMs) through self-play mutual reasoning, without the need for fine-tuning or stronger models.
MVTokenFlow: High-quality 4D Content Generation using Multiview Token Flow
Hanzhuo Huang (ShanghaiTech University), Sibei Yang (Sun Yat-sen University)
GenerationData SynthesisDiffusion modelOptical FlowVideo
🎯 What it does: This paper proposes the MVTokenFlow framework, which generates high-quality 4D content from monocular videos.
N-ForGOT: Towards Not-forgetting and Generalization of Open Temporal Graph Learning
Liping Wang (Hong Kong University of Science and Technology), Lei Chen (Hong Kong University of Science and Technology)
OptimizationRepresentation LearningGraph Neural NetworkGraphTime Series
🎯 What it does: Proposes the N-ForGOT framework, which provides a pluggable two-module solution to address the issues of catastrophic forgetting and distribution drift in Open Temporal Graph Learning (OTGL);
NarrativeBridge: Enhancing Video Captioning with Causal-Temporal Narrative
Asmar Nadeem (University of Surrey), Armin Mustafa (University of Surrey)
GenerationTransformerLarge Language ModelMixture of ExpertsContrastive LearningVideoText
🎯 What it does: This paper presents a new video subtitle benchmark dataset CTN (Causal-Temporal Narrative) and a Cause-Effect Network (CEN) model specifically designed to capture causal-temporal narratives.
Narrowing Information Bottleneck Theory for Multimodal Image-Text Representations Interpretability
Zhiyu Zhu (University of Technology Sydney), Fang Chen (University of Technology Sydney)
Explainability and InterpretabilityTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: In response to the interpretability issues of multimodal image-text models like CLIP, a new Narrowing Information Bottleneck Theory (NIBT) is proposed, which achieves an explanation method that is free of randomness, hyperparameters, and can distinguish negative attributes.
Natural Language Inference Improves Compositionality in Vision-Language Models
Paola Cascante-Bonilla (University of Maryland), Rachel Rudinger (University of Maryland)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: This paper proposes a method for generating entailment and contradiction sentences using natural language inference to enhance the combinatorial reasoning and text diversity of visual language models (CECE).
NatureLM-audio: an Audio-Language Foundation Model for Bioacoustics
David Robinson (Earth Species Project), Olivier Pietquin (Earth Species Project)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningMultimodalityAudio
🎯 What it does: This paper proposes NatureLM-audio, an audio-language foundation model focused on bioacoustics, for detecting, classifying, labeling, and describing animal sounds.
Navigating Neural Space: Revisiting Concept Activation Vectors to Overcome Directional Divergence
Frederik Pahde (Fraunhofer Heinrich Hertz Institute), Sebastian Lapuschkin (Fraunhofer Heinrich Hertz Institute)
Explainability and InterpretabilityImageBiomedical Data
🎯 What it does: A concept activation vector method based on patterns (Pattern-CAV) is proposed to accurately capture the true direction of concepts in neural networks, addressing the directional bias issue of traditional filter-based methods (Filter-CAV).
Navigating the Digital World as Humans Do: Universal Visual Grounding for GUI Agents
Boyu Gou (Ohio State University), Yu Su (Ohio State University)
TransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Developed a GUI agent framework SeeAct-V based on visual perception and pixel-level interaction, and trained the UGround visual localization model;
Navigation-Guided Sparse Scene Representation for End-to-End Autonomous Driving
Peidong Li (Zhijia Technology), Dixiao Cui (Zhijia Technology)
Autonomous DrivingTransformerPoint Cloud
🎯 What it does: The SSR framework is proposed, utilizing only 16 navigation-guided sparse BEV tokens to achieve end-to-end autonomous driving perception and planning, eliminating the reliance on traditional perception tasks.
ND-SDF: Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction
Ziyu Tang (Zhejiang University), Guofeng Zhang (Zhejiang University)
RestorationNeural Radiance FieldPoint CloudMesh
🎯 What it does: A method based on an adaptive normal offset field, ND-SDF, is constructed for high-precision reconstruction of indoor 3D surfaces from multi-view images.
Near-Exact Privacy Amplification for Matrix Mechanisms
Christopher A. Choquette-Choo (Google DeepMind), Abhradeep Guha Thakurta
OptimizationSafty and PrivacyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a general framework for approximate differential privacy analysis that can utilize random batching (balls-in-bins) to achieve privacy amplification under any lower triangular non-negative correlation matrix C, and optimizes C through Monte Carlo accounting to minimize the RMSE of the prefix sum, thereby improving the accuracy of private machine learning.
Near-optimal Active Regression of Single-Index Models
Yi Li (Nanyang Technological University), Wai Ming Tai
OptimizationTabular
🎯 What it does: This paper studies the active regression problem of the single-index model and proposes a new algorithm that can achieve a (1 + ε)-approximate solution by querying the minimum number of b entries.
Near-Optimal Online Learning for Multi-Agent Submodular Coordination: Tight Approximation and Communication Efficiency
Qixin Zhang (City University of Hong Kong), Dacheng Tao (Nanyang Technological University)
OptimizationRobotic IntelligenceVideo
🎯 What it does: This paper proposes two online multi-agent submodular function maximization algorithms, MA-OSMA and MA-OSEA, which can achieve a tight (1−e^{−c}/c) approximation on sparse connected communication networks and provide a dynamic loss upper bound of O(√{C_T T/(1−β)}).
Near-Optimal Policy Identification in Robust Constrained Markov Decision Processes via Epigraph Form
Toshinori Kitamura (University of Tokyo), Yutaka Matsuo (University of Tokyo)
OptimizationReinforcement LearningTabular
🎯 What it does: The first theoretically guaranteed near-optimal policy identification algorithm in Robust Constrained Markov Decision Processes (RCMDP) is proposed, along with the corresponding double-loop solving framework EpiRC-PGS.
Near, far: Patch-ordering enhances vision foundation models' scene understanding
Valentinos Pariza (University of Amsterdam), Yuki M Asano
SegmentationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: A self-supervised training method based on differentiable nearest neighbor consistency, NeCo, has been developed for dense post-pretraining on existing Vision Transformers, significantly improving the quality of spatial features.
NEAR: A Training-Free Pre-Estimator of Machine Learning Model Performance
Raphael T. Husistein (ETH Zurich), Marco Eckhoff (ETH Zurich)
Hyperparameter SearchNeural Architecture SearchTabularBenchmark
🎯 What it does: This paper proposes a zero-cost estimator NEAR based on the effective rank of the activation matrix, which is used to evaluate network performance without training, and is applied in tasks such as neural architecture search, layer size estimation, and selection of activation functions and weight initialization.
Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs
Zijia Zhao (University of Chinese Academy of Sciences), Jing Liu
Data SynthesisRetrievalLarge Language ModelPrompt EngineeringVideoMultimodalityBenchmark
🎯 What it does: The VideoNIAH framework has been constructed, and based on it, a scalable synthetic video evaluation benchmark VNBench has been created, which can perform fine-grained assessments of video LLMs' temporal awareness, sequential reasoning, and spatiotemporal consistency.
Needle Threading: Can LLMs Follow Threads Through Near-Million-Scale Haystacks?
Jonathan Roberts (University of Cambridge), Samuel Albanie (University of Hong Kong)
RetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: The researchers designed a set of artificial synthetic 'needle online' and 'thread tracking' retrieval tasks to evaluate the retrieval and reasoning capabilities of 17 long-context LLMs at different context lengths.
NeRAF: 3D Scene Infused Neural Radiance and Acoustic Fields
Amandine Brunetto (Mines Paris PSL University), Fabien Moutarde (Mines Paris PSL University)
GenerationData SynthesisNeural Radiance FieldImageMultimodalityAudio
🎯 What it does: A cross-modal neural field model NeRAF has been developed that can simultaneously generate perspective images and room impulse responses (RIR), capable of synthesizing realistic audio and images from any camera and microphone position.
Nesterov acceleration in benignly non-convex landscapes
Kanan Gupta (University of Pittsburgh), Stephan Wojtowytsch
OptimizationTabularOrdinary Differential Equation
🎯 What it does: Under geometric assumptions that only require the combination of 'strong objective conditions' and 'quasi-strong convexity' without the need for strong convexity, it is proven that Nesterov's accelerated method can still achieve accelerated convergence rates in continuous time, discrete time, and in stochastic gradient scenarios with additive/multiplicative noise; new Lyapunov analysis and tangent drift control are also provided.
NeSyC: A Neuro-symbolic Continual Learner For Complex Embodied Tasks In Open Domains
Wonje Choi (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelContrastive LearningMultimodalityChain-of-Thought
🎯 What it does: A neural-symbolic continual learning framework named NESYC has been designed and implemented, which utilizes large language models and symbolic reasoning tools to collaboratively generate and continuously improve executable action rules, thereby achieving adaptive planning in open-domain embodied tasks.
NetFormer: An interpretable model for recovering dynamical connectivity in neuronal population dynamics
Ziyu Lu (University of Washington), Lu Mi (Georgia Institute of Technology)
Explainability and InterpretabilityTransformerMultimodalityTime Series
🎯 What it does: This paper proposes an interpretable model called NetFormer based on a linearized attention mechanism, aimed at recovering time-varying (non-stationary) connectivity structures from neuronal activity sequences.
NetMoE: Accelerating MoE Training through Dynamic Sample Placement
Xinyi Liu (Peking University), Bin CUI
OptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper proposes the NetMoE framework, which dynamically adjusts the distribution of training samples on devices at each MoE layer, transforming inter-node communication into intra-node communication, significantly improving All-to-All communication efficiency.
Neural Approximate Mirror Maps for Constrained Diffusion Models
Berthy Feng, Katherine Bouman
GenerationData SynthesisOptimizationDiffusion modelImage
🎯 What it does: This paper proposes a method for learning mirror mappings using neural networks (NAMM), enabling diffusion models to satisfy arbitrary differentiable constraints while maintaining generation quality.
Neural Causal Graph for Interpretable and Intervenable Classification
Jiawei Wang (Hunan University), Tat-Seng Chua (National University of Singapore)
ClassificationExplainability and InterpretabilityGraph Neural NetworkReinforcement LearningImage
🎯 What it does: Proposed a Neural Causal Graph (NCG) framework to achieve interpretable and intervenable classification models;
Neural Context Flows for Meta-Learning of Dynamical Systems
Roussel Desmond Nzoyem (University of Bristol), Tom Deakin (University of Bristol)
Meta LearningTime SeriesOrdinary Differential Equation
🎯 What it does: Researches meta-learning methods for neural ODEs in multiple environments, proposing Neural Context Flows (NCF) for rapid adaptation.
Neural Dueling Bandits: Preference-Based Optimization with Human Feedback
Arun Verma (Singapore MIT Alliance for Research and Technology), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: In response to the contextual dueling bandit problem, we utilize human preference feedback to train a neural network to estimate the reward function. Based on this, we propose two algorithms based on UCB and Thompson Sampling to achieve arm selection, providing a sublinear regret upper bound.
Neural Eulerian Scene Flow Fields
Kyle Vedder (University of Pennsylvania), Joachim Pehserl (NVIDIA)
Autonomous DrivingOptimizationOptical FlowPoint CloudOrdinary Differential Equation
🎯 What it does: Proposes to estimate scene flow as fitting an ODE for the entire observation sequence and implements an unsupervised method called EulerFlow;
Neural Exploratory Landscape Analysis for Meta-Black-Box-Optimization
Zeyuan Ma (South China University of Technology), Yue-Jiao Gong (South China University of Technology)
OptimizationMeta LearningTransformerNeural Radiance Field
🎯 What it does: This paper studies a learnable exploration landscape analyzer, NeurELA, which dynamically extracts low-level optimization states using an end-to-end two-stage attention network, replacing traditional handcrafted features to enhance the performance of Meta-black-box optimization.
Neural Fluid Simulation on Geometric Surfaces
Haoxiang Wang (Tsinghua University), Qionghai Dai (Tsinghua University)
GenerationData SynthesisComputational EfficiencyAuto EncoderMesh
🎯 What it does: A surface incompressible fluid simulation framework based on implicit neural representation (NFFS) is proposed, which can directly construct divergence-free fields on different surface representations and achieve energy-preserving time integration.
Neural Functions for Learning Periodic Signal
Woojin Cho (TelePIX), Noseong Park (KAIST)
Time Series
🎯 What it does: A periodic signal learning framework called NeRT based on implicit neural representation is proposed, which can perform both interpolation and extrapolation simultaneously.
Neural Interactive Proofs
Lewis Hammond (University of Oxford), Sam Adam-Day (University of Oxford)
Graph Neural NetworkTransformerReinforcement LearningTextGraph
🎯 What it does: This study proposes a unified framework for Neural Interactive Proofs, designs various new protocols (nip, mnip, zk-nip, zk-mnip), provides theoretical equivalence proofs, and conducts experimental evaluations on two major tasks: graph isomorphism and code verification.
Neural Multi-Objective Combinatorial Optimization via Graph-Image Multimodal Fusion
Jinbiao Chen (Sun Yat-sen University), Yaoxin Wu (Eindhoven University of Technology)
OptimizationGraph Neural NetworkTransformerMultimodalityGraph
🎯 What it does: A Graph-Image Multimodal Fusion (GIMF) framework is proposed to enhance the performance of neural multi-objective combinatorial optimization by constructing instance images and jointly learning with graph structures.
Neural networks on Symmetric Spaces of Noncompact Type
Xuan Son Nguyen (CY Cergy Paris University), Aymeric Histace (CY Cergy Paris University)
ClassificationImageTime Series
🎯 What it does: A unified framework is proposed for constructing the distance from a point to a hyperplane in non-compact symmetric spaces (including hyperplane spaces and SPD manifolds), and based on this, a fully connected layer and attention mechanism are designed to build a novel neural network.
Neural ODE Transformers: Analyzing Internal Dynamics and Adaptive Fine-tuning
Anh Tong (Korea University), Jaesik Choi (KAIST)
TransformerSupervised Fine-TuningTextOrdinary Differential Equation
🎯 What it does: Proposes to express the Transformer architecture as a non-autonomous neural ODE (DiffEqTransformer), generating time-dependent weights for attention and feedforward layers, and using an ODE solver to achieve adaptive layer counts;
Neural Phylogeny: Fine-Tuning Relationship Detection among Neural Networks
Runpeng Yu (National University of Singapore), Xinchao Wang (National University of Singapore)
Convolutional Neural NetworkTransformerSupervised Fine-TuningDiffusion modelImage
🎯 What it does: This paper proposes the task of 'neural system evolutionary tree detection', which involves identifying parent-child model pairs within a set of neural networks and determining the fine-tuning direction.
Neural Sampling from Boltzmann Densities: Fisher-Rao Curves in the Wasserstein Geometry
Jannis Chemseddine (Institute of Mathematics), Gabriele Steidl (Institute of Mathematics)
Stochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The research addresses the problem of sampling from unnormalized Boltzmann densities, proposing a neural sampling method based on gradient flows, and conducting theoretical and experimental analyses on different interpolation paths (linear, learned interpolation, gradient flow interpolation).
Neural Spacetimes for DAG Representation Learning
Haitz Sáez de Ocáriz Borde (University of Oxford), Michael M. Bronstein (University of Oxford)
Representation LearningGraph Neural NetworkGraph
🎯 What it does: This paper proposes a learnable Neural SpaceTimes (NST) that maps weighted directed acyclic graph (DAG) nodes to spacetime events, encoding the graph's distance and causal relationships through trainable spatial metrics and temporal orders.
Neural Stochastic Differential Equations for Uncertainty-Aware Offline RL
Cevahir Koprulu (University of Texas at Austin), ufuk topcu
Reinforcement LearningTabularBenchmarkStochastic Differential Equation
🎯 What it does: A neural stochastic differential equation (Neural SDE) based offline model learning framework called NUNO is proposed for achieving robust offline reinforcement learning under low-quality data or incomplete demonstrations.
Neural Wave Equation for Irregularly Sampled Sequence Data
Arkaprava Majumdar (Indian Institute of Technology), P. K. Srijith (Indian Institute of Technology)
ClassificationRecognitionRecurrent Neural NetworkTime SeriesSequentialBiomedical DataOrdinary Differential Equation
🎯 What it does: A neural wave equation model is proposed, which implements the continuous evolution of hidden states over time and depth through a parameterized source term of a non-homogeneous wave equation to address the issue of irregularly sampled sequence labeling.
Neuralized Markov Random Field for Interaction-Aware Stochastic Human Trajectory Prediction
Zilin Fang (National University of Singapore), Gim Hee Lee (National University of Singapore)
GenerationData SynthesisRobotic IntelligenceRecurrent Neural NetworkGraph Neural NetworkAuto EncoderTime SeriesSequential
🎯 What it does: A Neuralized Markov Random Field (Neuralized MRF) model is proposed to simultaneously capture the Markov evolution of individual movements and the effects of group interactions, thereby achieving stochastic human trajectory prediction with interaction awareness.
NeuralPlane: Structured 3D Reconstruction in Planar Primitives with Neural Fields
Hanqiao Ye (University of Chinese Academy of Sciences), Shuhan Shen (Chinese Academy of Sciences)
SegmentationGenerationDepth EstimationNeural Radiance FieldContrastive LearningImage
🎯 What it does: This paper proposes NeuralPlane, a planar primitive that utilizes neural fields to reconstruct indoor scenes from multi-view images without annotations, achieving joint modeling of geometry and semantics.
NeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions
Tue Minh Cao, My T. Thai (University of Florida)
Explainability and InterpretabilityConvolutional Neural NetworkLarge Language ModelImage
🎯 What it does: An automated framework called NeurFlow is proposed, which uses neuron groups instead of individual neurons to explain the internal mechanisms of CNNs and constructs inter-layer interaction circuits.
NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals
Weibang Jiang, Dongsheng Li (Microsoft Research Asia)
RecognitionAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderMultimodalityTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: This study proposes NeuroLM, a multi-task foundational model that treats EEG signals as a 'foreign language', utilizing large language models (LLM) to achieve a unified EEG processing framework capable of completing six different EEG tasks at once.
Neuron based Personality Trait Induction in Large Language Models
Jia Deng (Renmin University of China), Ji-Rong Wen (Renmin University of China)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a neuron-level personality trait induction method called NPTI.
Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning
Wei Wu (Peking University), Jinzhuo Wang (Peking University)
Representation LearningSpiking Neural NetworkContrastive LearningTime SeriesBiomedical Data
🎯 What it does: Proposes the NeurPIR framework, which learns time-invariant intrinsic representations of neurons from dynamic data of neuronal populations through contrastive learning.
Neuron-based Multifractal Analysis of Neuron Interaction Dynamics in Large Models
Xiongye Xiao (University of Southern California), Paul Bogdan (University of Southern California)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes transforming large models into Neuron Interaction Networks (NIN) and evaluating their internal structural dynamics based on multifractal analysis (NeuroMFA), thereby quantifying the 'self-organization' and 'emergent' capabilities of large models.
Neuroplastic Expansion in Deep Reinforcement Learning
Jiashun Liu (Hong Kong University of Science and Technology), Ling Pan (Hong Kong University of Science and Technology)
Reinforcement LearningImage
🎯 What it does: A dynamic network growth method called Neuroplastic Expansion (NE) is proposed, which maintains plasticity and enhances learning performance in deep reinforcement learning through gradient-driven network expansion, lazy neuron pruning, and experience replay.
New Algorithms for the Learning-Augmented k-means Problem
Junyu Huang (Central South University), Jianxin Wang (Central South University)
OptimizationImage
🎯 What it does: This study enhances the k-means clustering problem by proposing three linear-time algorithms based on sampling: Fast-Sampling, Fast-Estimation, and Fast-Filtering, achieving significant approximation guarantees.
Newton Meets Marchenko-Pastur: Massively Parallel Second-Order Optimization with Hessian Sketching and Debiasing
Elad Romanov (Stanford University), Mert Pilanci (Stanford University)
OptimizationTabular
🎯 What it does: In a large-scale parallel, communication-constrained serverless computing environment, an adaptive random sampling and debiased Hessian matrix estimation method is proposed, combined with multi-worker averaging to approximate Newton steps, thereby achieving a scalable second-order optimization algorithm.
NExT-Mol: 3D Diffusion Meets 1D Language Modeling for 3D Molecule Generation
Zhiyuan Liu (National University of Singapore), Tat-Seng Chua (National University of Singapore)
GenerationData SynthesisDrug DiscoveryTransformerLarge Language ModelDiffusion modelGraph
🎯 What it does: This paper proposes a two-step 3D molecular generation framework NExT-Mol based on a large 1D SELFIES language model MoLlama and a 3D diffusion model DMT, utilizing pre-trained 1D representations to enhance 3D predictions through cross-modal projection, and improving generation diversity through random SELFIES data augmentation.
NextBestPath: Efficient 3D Mapping of Unseen Environments
Shiyao Li (Ecole Nationale des Ponts et Chaussees), Vincent Lepetit (Inria)
OptimizationRobotic IntelligenceConvolutional Neural NetworkSimultaneous Localization and MappingPoint Cloud
🎯 What it does: The Next Best Path (NBP) method is proposed for active 3D mapping, and a multi-difficulty indoor environment dataset AiMDoom is constructed.
NExUME: Adaptive Training and Inference for DNNs under Intermittent Power Environments
Cyan Subhra Mishra (Pennsylvania State University), Chita R. Das
OptimizationComputational EfficiencyNeural Architecture SearchTabularTime Series
🎯 What it does: A complete framework called NExUME has been designed and implemented, capable of training and inferring deep neural networks in energy-harvesting environments during power outages by dynamically adjusting dropout, quantization, and task scheduling.
nGPT: Normalized Transformer with Representation Learning on the Hypersphere
Ilya Loshchilov (NVIDIA), Boris Ginsburg (NVIDIA)
OptimizationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes the Normalized Transformer (nGPT), which normalizes all vectors (embeddings, attention matrices, MLP weights, and hidden states) of the Transformer to the unit sphere, achieving faster convergence.
NL-Eye: Abductive NLI For Images
Mor Ventura (Technion), Roi Reichart (Technion)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Construct the NL-EYE benchmark to evaluate the causal reasoning and interpretability of visual language models (VLM) in multi-image scenarios.
NNsight and NDIF: Democratizing Access to Open-Weight Foundation Model Internals
Jaden Fried Fiotto-Kaufman, David Bau
TransformerLarge Language ModelTextBenchmark
🎯 What it does: The NNsight and NDIF frameworks are proposed, enabling researchers to experiment with ultra-large-scale open-source weight models without self-hosting the models.
No Equations Needed: Learning System Dynamics Without Relying on Closed-Form ODEs
Krzysztof Kacprzyk (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
Drug DiscoveryTime SeriesBiomedical DataOrdinary Differential Equation
🎯 What it does: A Direct Semantic Modeling framework is proposed and implemented to directly predict the semantic representation of dynamic systems (trajectory shapes, transition points, attributes) from data, generating trajectories that satisfy this semantics through a semantic predictor and a trajectory predictor, bypassing the traditional two-step process of closed-form ODE discovery and subsequent analysis.
No Free Lunch: Fundamental Limits of Learning Non-Hallucinating Generative Models
Changlong Wu (Purdue University), Wojciech Szpankowski (Jagiellonian University)
Generation
🎯 What it does: A learnability framework for hallucination-free generative models is constructed from the perspective of learning theory, clarifying the fundamental limitations that prevent the model from avoiding hallucinations based solely on training data;
No Location Left Behind: Measuring and Improving the Fairness of Implicit Representations for Earth Data
Daniel Cai (Brown University), Randall Balestriero (Brown University)
Representation LearningData-Centric LearningMultimodalityBenchmark
🎯 What it does: The FAIR-EARTH dataset is proposed to evaluate the fairness of implicit neural representations (INR) in Earth data, and a spherical wavelet-based encoding (SPHERICAL WAVELET) is developed and validated on this dataset to mitigate local signal bias.
No Need to Talk: Asynchronous Mixture of Language Models
Anastasiia Filippova (École polytechnique fédérale de Lausanne), Ronan Collobert (Apple)
TransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: We propose SMALLTALK LM, an asynchronous multi-expert language model training and inference framework that utilizes lightweight routers to allocate sequences to independent experts based on short prefixes.
No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images
Botao Ye (ETH Zurich), Songyou Peng (Google DeepMind)
Data SynthesisPose EstimationDepth EstimationTransformerGaussian SplattingSimultaneous Localization and MappingImageVideo
🎯 What it does: A feedforward network directly predicts a 3D Gaussian model on sparse uncalibrated images without camera pose information, achieving scalable 3D reconstruction and novel view synthesis.
No Preference Left Behind: Group Distributional Preference Optimization
Binwei Yao (Stanford University), Junjie Hu (University of Wisconsin Madison)
Recommendation SystemOptimizationLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies a new framework called GDPO, which allows large language models to generate diverse responses according to group preference distributions.
No Training, No Problem: Rethinking Classifier-Free Guidance for Diffusion Models
Seyedmorteza Sadat (ETH Zurich), Romann M. Weber (Disney Research Studios)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper re-examines unconditional guidance (CFG) and proposes two new untrained guidance methods: Independent Conditional Guidance (ICG) and Time Step Guidance (TSG) to enhance the generation quality of diffusion models.
Node Identifiers: Compact, Discrete Representations for Efficient Graph Learning
Yuankai Luo (Beihang University), Xiao-Ming Wu (Hong Kong Polytechnic University)
ClassificationComputational EfficiencyRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: An end-to-end NID framework is proposed, which uses vector quantization to compress the continuous node embeddings of multi-layer GNNs into discrete int4 node IDs of 6–15 dimensions, applicable for both supervised and unsupervised learning.
Node Similarities under Random Projections: Limits and Pathological Cases
Tvrtko Tadić (Microsoft), Jennifer Neville (Microsoft Research)
Graph Neural NetworkGaussian SplattingGraph
🎯 What it does: This study investigates the preservation effects of random projections on dot product and cosine similarity in graph learning tasks, proposing new asymptotic and finite sample analyses, and discovering pathological cases of dot product at low/high degree nodes. The theory is then applied to node embedding-based ranking tasks, with performance evaluated using the NDCG metric.
Node-Time Conditional Prompt Learning in Dynamic Graphs
Xingtong Yu (Singapore Management University), Yuan Fang (Singapore Management University)
ClassificationRecommendation SystemGraph Neural NetworkPrompt EngineeringGraph
🎯 What it does: Designed DYGPROMPT, a dual-prompt learning framework for dynamic graphs, for efficient transfer in pre-training and downstream tasks.
Noise Separation guided Candidate Label Reconstruction for Noisy Partial Label Learning
Xiaorui Peng (Southeast University), Min-Ling Zhang (Southeast University)
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A framework for Noise Partial Label Learning (NPLL) is proposed, which reduces the noise rate and shortens the candidate label set length through sample separation and candidate label set reconstruction, thereby enhancing the generalization performance of the classifier.
Noise-conditioned Energy-based Annealed Rewards (NEAR): A Generative Framework for Imitation Learning from Observation
Anish Abhijit Diwan (TU Delft), Jan Peters (TU Darmstadt)
Robotic IntelligenceReinforcement LearningScore-based ModelVideo
🎯 What it does: This paper proposes a noise-modulated energy function-based annealing reward framework (NEAR), which uses the energy function directly as a reward to learn observation-based imitation learning policies, gradually enhancing the reward precision through annealing.
Noisy Test-Time Adaptation in Vision-Language Models
Chentao Cao (Hong Kong Baptist University), Bo Han (Hong Kong Baptist University)
Domain AdaptationAnomaly DetectionTransformerVision Language ModelImageMultimodalityBenchmark
🎯 What it does: A zero-shot noise testing adaptation (ZS-NTTA) framework is proposed, and the AdaND method is designed for this task, utilizing a pre-trained vision-language model (CLIP) to freeze the classifier, a single-layer adaptive noise detector with dynamic thresholds, and injecting Gaussian noise into clean data streams to avoid misjudgments.
Non-Adversarial Inverse Reinforcement Learning via Successor Feature Matching
Arnav Kumar Jain (Mila Quebec AI Institute), Sanjiban Choudhury (Mila Quebec AI Institute)
Reinforcement LearningSequential
🎯 What it does: Inverse reinforcement learning is achieved by directly matching the expert's Successor Features, without the need for adversarial reward learning, and can use only the expert's state sequences for imitation.
Non-Equilibrium Dynamics of Hybrid Continuous-Discrete Ground-State Sampling
Timothee Leleu, Sam Reifenstein (NTT Research)
OptimizationGraphPhysics RelatedStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A hybrid continuous-discrete algorithm MHCACm is proposed, combining continuous dynamics (CAC + momentum) with Metropolis-Hastings for sampling degenerate ground states of the Ising Hamiltonian.
Non-myopic Generation of Language Models for Reasoning and Planning
Chang Ma (University of Hong Kong), Lingpeng Kong (Hong Kong University of Science and Technology)
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This paper proposes a Predictive-Decoding method that enhances non-myopic performance in reasoning and planning tasks by re-weighting the distribution generated by LLM.
Nonasymptotic Analysis of Stochastic Gradient Descent with the Richardson–Romberg Extrapolation
Marina Sheshukova (HSE University), Sergey Samsonov (HSE University)
OptimizationTabular
🎯 What it does: This paper studies the non-asymptotic error upper bound of the iterator obtained by combining constant step size SGD with Polyak-Ruppert averaging and Richardson-Romberg extrapolation under strongly convex smooth objectives, providing an exact expansion that includes both main and secondary terms.
Nonconvex Stochastic Optimization under Heavy-Tailed Noises: Optimal Convergence without Gradient Clipping
Zijian Liu (Stern School of Business New York University), Zhengyuan Zhou (Stern School of Business New York University)
Optimization
🎯 What it does: This paper proposes an algorithm for non-convex stochastic optimization in the presence of heavy-tailed noise that achieves convergence without the need for gradient clipping.
Nonlinear multiregion neural dynamics with parametric impulse response communication channels
Matthew Dowling (New York University), Cristina Savin (New York University)
Recurrent Neural NetworkAuto EncoderTime Series
🎯 What it does: A multi-region neural dynamic model (MRDS-IR) is proposed, which combines the nonlinear local dynamics of each region with a linear communication channel parameterized by impulse response, and presents a variational filtering and learning algorithm that does not invert state noise.
Nonlinear Sequence Embedding by Monotone Variational Inequality
Jonathan Yuyang Zhou, Yao Xie (Georgia Institute of Technology)
ClassificationOptimizationRepresentation LearningTextTime SeriesSequential
🎯 What it does: This paper proposes an unsupervised low-dimensional sequence embedding method that utilizes low-rank matrix recovery and a monotonic variational inequality (VI) framework to estimate parameters of nonlinear autoregressive sequences, thereby generating sequence representations that can be used for downstream tasks such as clustering and classification.
Not All Heads Matter: A Head-Level KV Cache Compression Method with Integrated Retrieval and Reasoning
Yu Fu (University of California), Wen Xiao (Microsoft)
RetrievalCompressionTransformerLarge Language ModelText
🎯 What it does: This paper proposes a KV cache compression method for attention heads, HeadKV-R2, which significantly reduces KV cache usage while maintaining the long text reasoning capabilities of LLMs.
Not All Language Model Features Are One-Dimensionally Linear
Joshua Engels (Massachusetts Institute of Technology), Max Tegmark (Massachusetts Institute of Technology)
Explainability and InterpretabilityTransformerLarge Language ModelAuto EncoderText
🎯 What it does: This paper first provides a strict definition of multidimensional features and uses Sparse Autoencoders (SAE) to automatically retrieve interpretable multidimensional features from the hidden layers of GPT-2, Mistral-7B, and Llama-3-8B. Through clustering and visualization, they discovered features with a circular distribution (such as the seven days of the week and twelve months). Subsequently, the authors designed circular subspace intervention experiments based on activation patches, demonstrating that these circular features play a causal role in modular arithmetic tasks (such as 'What day is it seven days after Monday?') and compared different models and layers.