arXivSub Start free trial

NeurIPS 2025 Papers with Code β€” Page 14

Conference on Neural Information Processing Systems Β· 2283 papers

Multilevel neural simulation-based inference

Yuga Hikida (Aalto University), Francois-Xavier Briol

Code

🎯 What it does: This paper proposes Multi-Layer Neural Likelihood Estimation/Posterior Estimation (ML-NLE/ML-NPE), which accelerates neural SBI by combining multi-layer Monte Carlo with low and high precision simulators.

Multimodal 3D Genome Pre-training

Minghao Yang (Hong Kong University of Science and Technology), Hui Xiong (Hong Kong University of Science and Technology)

CodeTransformerContrastive LearningMultimodalityBiomedical Data

🎯 What it does: We propose the MIX-HIC multimodal foundational model, which jointly utilizes Hi-C contact maps and epigenomic trajectories for large-scale self-supervised pre-training, and applies it to various downstream tasks such as Hi-C prediction, chromosome loop detection, and CAGE expression.

Multimodal Causal Reasoning for UAV Object Detection

Nianxin Li (University of Electronic Science and Technology of China), Ce Zhu

CodeObject DetectionVision Language ModelContrastive LearningImageMultimodality

🎯 What it does: In drone target detection, a framework called MCR-UOD based on multimodal causal reasoning is proposed, which enhances detection robustness through visual-language models and causal interventions.

Multimodal Disease Progression Modeling via Spatiotemporal Disentanglement and Multiscale Alignment

Chen Liu (Hong Kong Polytechnic University), Jing Qin (Hong Kong Baptist University)

CodeClassificationRecognitionTransformerImageMultimodalityBiomedical DataElectronic Health Records

🎯 What it does: The DiPro framework is proposed, which uses regional-level spatiotemporal separation and multi-scale alignment to integrate long sequences of chest X-rays and electronic medical records for disease progression modeling and ICU prediction.

MultiNet: Adaptive Multi-Viewed Subgraph Convolutional Networks for Graph Classification

Xinya Qin (Beijing Normal University), Edwin Hancock

CodeClassificationGraph Neural NetworkGraphBiomedical Data

🎯 What it does: A MultiNet is designed and implemented, utilizing adaptive multi-view subgraph convolution and an alignment readout mechanism to address the issue of over-smoothing at the graph level and improve graph classification performance.

Multipole Attention for Efficient Long Context Reasoning

Coleman Richard Charles Hooper (University of California), Amir Gholami (University of California)

CodeComputational EfficiencyTransformerLarge Language ModelText

🎯 What it does: The Multipole Attention method is proposed, which uses clustered key centroids to apply precise attention to important keys, while the remaining keys are represented approximately to accelerate long-chain inference.

Multiresolution Analysis and Statistical Thresholding on Dynamic Networks

Raphael Romero, Alexander Modell (Imperial College London)

CodeAnomaly DetectionGraph Neural NetworkGraphTime Series

🎯 What it does: This paper proposes the ANIE method for adaptively identifying structural changes at different time scales in continuous-time dynamic networks.

Multivariate Latent Recalibration for Conditional Normalizing Flows

Victor Dheur (University of Mons), Souhaib Ben Taieb (Mohamed bin Zayed University of Artificial Intelligence)

CodeGenerationFlow-based ModelImageTabular

🎯 What it does: A Latent Recalibration (LR) method is proposed for post-calibration in the latent space of conditional reversible generative models (such as normalizing flow and flow matching);

MuRating: A High Quality Data Selecting Approach to Multilingual Large Language Model Pretraining

Zhixun Chen (Hong Kong University of Science and Technology), Meng Fang (University of Liverpool)

CodeTransformerLarge Language ModelContrastive LearningText

🎯 What it does: Proposes the MuRating framework, which aggregates various English automatic evaluators and transfers them to multilingual data quality assessment.

MUSTAFAR: Promoting Unstructured Sparsity for KV Cache Pruning in LLM Inference

Donghyeon Joo (University of Maryland), Bahar Asgari (d-Matrix)

CodeTransformerLarge Language ModelTextBenchmark

🎯 What it does: Pruning the Key/Value cache of LLM using unstructured sparsification, combined with bitmap compression format and custom sparse attention kernels, significantly reduces KV cache usage and improves inference throughput.

MutualVPR: A Mutual Learning Framework for Resolving Supervision Inconsistencies via Adaptive Clustering

Qiwen Gu (Tongji University), Guang Chen (Tongji University)

CodeRetrievalRepresentation LearningTransformerContrastive LearningImage

🎯 What it does: This paper proposes MutualVPR, a visual localization framework that achieves mutual learning through adaptive clustering to address the issue of supervision inconsistency caused by viewpoint changes and occlusions.

MVSMamba: Multi-View Stereo with State Space Model

Jianfei Jiang (University of Science and Technology Beijing), Huimin Ma (University of Science and Technology Beijing)

CodeDepth EstimationConvolutional Neural NetworkImage

🎯 What it does: MVSMamba is proposed, a multi-view stereo reconstruction network based on the Mamba architecture;

NaDRO: Leveraging Dual-Reward Strategies for LLMs Training on Noisy Data

Haolong Qian (Tsinghua University), Chun Yuan (Microsoft)

CodeTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: The research trains large language models for long-sequence decision-making under noisy data and proposes the NaDRO framework.

NAUTILUS: A Large Multimodal Model for Underwater Scene Understanding

Wei Xu (Huazhong University of Science and Technology), Xiang Bai (Huazhong University of Science and Technology)

CodeClassificationObject DetectionDepth EstimationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality

🎯 What it does: The NAUTILUS large multimodal model (LMM) is proposed for underwater scene understanding, and the first underwater instruction-following dataset NautData is constructed, covering three levels of granularity and eight tasks, including coarse and fine classification, counting, VQA, detection, localization, area description, and image description.

β€ŒNavigating the MIL Trade-Off: Flexible Pooling for Whole Slide Image Classification

Hossein Jafarinia (Sharif University of Technology), Mohammad Hossein Rohban (Sharif University of Technology)

CodeClassificationTransformerImageBenchmark

🎯 What it does: This paper addresses WSI classification under low data conditions, re-examining and improving traditional MIL aggregation methods, proposing Maxsoft pooling and PerPatch augmentation;

NaViL: Rethinking Scaling Properties of Native Multimodal Large Language Models under Data Constraints

Changyao Tian (Shanghai AI Laboratory), Jifeng Dai (Tsinghua University)

CodeTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodality

🎯 What it does: The research focuses on end-to-end training of multimodal large language models under data-constrained conditions, systematically exploring the design space and scaling characteristics, and proposing the NaViL model.

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)

CodeData 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.

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)

CodeObject 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.

NestedFP: High-Performance, Memory-Efficient Dual-Precision Floating Point Support for LLMs

Haeun Lee (Seoul National University), Jae W. Lee (Seoul National University)

CodeLarge 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.

Network two-sample test for block models

Chung Kyong Nguen (University of California), OSCAR HERNAN MADRID PADILLA (University of California)

CodeGraph 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 Attention Search

Difan Deng (Leibniz University Hannover), Marius Lindauer (Leibniz University Hannover)

CodeOptimizationComputational 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)

CodeCompressionOptical 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 Entropy

Akhil Premkumar (Yale University)

CodeGenerationData 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)

CodeOptimizationBenchmark

🎯 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)

CodeGraph 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 MJD: Neural Non-Stationary Merton Jump Diffusion for Time Series Prediction

Yuanpei Gao (University of British Columbia), Renjie Liao (University of British Columbia)

CodeGraph 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 Networks for Learnable and Scalable Influence Estimation of Instruction Fine-Tuning Data

Ishika Agarwal (University of Illinois Urbana-Champaign), Dilek Hakkani-TΓΌr (University of Illinois Urbana-Champaign)

CodeComputational 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.

NeuralPLexer3: Accurate Biomolecular Complex Structure Prediction with Flow Models

Zhuoran Qiao (Iambic Therapeutics), Matthew Welborn (Iambic Therapeutics)

CodeProtein 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

MΓ©lodie Monod (Imperial College London), Samir Bhatt (Imperial College London)

CodeBiomedical Data

🎯 What it does: We propose NeuralSurv, a framework capable of deep Bayesian modeling of survival data over continuous time;

Neuro-Spectral Architectures for Causal Physics-Informed Networks

Arthur Bizzi (Γ‰cole Polytechnique FΓ©dΓ©rale de Lausanne), Lucas Nissenbaum (Instituto de MatemΓ‘tica Pura e Aplicada)

CodeOptimizationTime SeriesPhysics RelatedOrdinary Differential Equation

🎯 What it does: This paper presents NeuSAβ€”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.

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)

CodeRetrievalLarge 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)

CodeGenerationOptimizationReinforcement 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)

CodeOptimizationExplainability 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.

Next Semantic Scale Prediction via Hierarchical Diffusion Language Models

Cai Zhou (Massachusetts Institute of Technology), Tommi Jaakkola (Massachusetts Institute of Technology)

CodeGenerationDiffusion 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)

CodeGenerationData 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.

No Loss, No Gain: Gated Refinement and Adaptive Compression for Prompt Optimization

Wenhang Shi (Renmin University of China), Xiaoyong Du (Renmin University of China)

CodeOptimizationTransformerLarge 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)

CodeSegmentationDomain 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 Thompson Sampling for Finite-Horizon Markov Decision Processes with Gaussian Processes

Jasmine Bayrooti (University of Cambridge), Carl Henrik Ek (University of Cambridge)

CodeReinforcement 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.

Noise Hypernetworks: Amortizing Test-Time Compute in Diffusion Models

Luca Eyring (Technical University of Munich), Zeynep Akata (Technical University of Munich)

CodeGenerationOptimizationComputational 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)

CodeTransformerLarge 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)

CodeClassificationOptimizationDiffusion 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);

Non-stationary Equivariant Graph Neural Networks for Physical Dynamics Simulation

Chaohao Yuan (Tsinghua University), Yu Rong (Alibaba Group)

CodeGraph 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 Structural Causal Bandits

Yeahoon Kwon (Seoul National University), Sanghack Lee (Seoul National University)

CodeReinforcement 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.

Nonlinearly Preconditioned Gradient Methods: Momentum and Stochastic Analysis

Konstantinos Oikonomidis (KU Leuven), Panagiotis Patrinos (KU Leuven)

CodeRecommendation 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.

Normal-Abnormal Guided Generalist Anomaly Detection

Yuexin Wang (Xi'an Jiaotong-Liverpool University), Jimin XIAO

CodeAnomaly 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.

Normalizing Flows are Capable Models for Continuous Control

Raj Ghugare (Princeton University), Benjamin Eysenbach (Princeton University)

CodeRobotic 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)

CodeClassificationFederated 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)

CodeTransformerSupervised 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 Exploration via Orthogonality

Andreas Theophilou (University of Bath), Γ–zgΓΌr Şimşek (University of Bath)

CodeRobotic 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.

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)

CodeReinforcement 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.

NystrΓΆm-Accelerated Primal LS-SVMs: Breaking the $O(an^3)$ Complexity Bottleneck for Scalable ODEs Learning

Weikuo Wang (China Three Gorges University), Huan Luo (China Three Gorges University)

CodeOptimizationComputational EfficiencyTime SeriesOrdinary Differential Equation

🎯 What it does: A Nystrâm accelerated LS-SVM framework is proposed for efficiently solving linear and nonlinear ordinary differential equations (ODEs).

OASIS: One-Shot Federated Graph Learning via Wasserstein Assisted Knowledge Integration

Guancheng Wan (Wuhan University), Mang Ye (Nanyang Technological University)

CodeFederated LearningKnowledge DistillationGraph Neural NetworkGenerative Adversarial NetworkGraph

🎯 What it does: The OASIS framework is proposed to achieve federated graph learning with a single communication round, capturing fine-grained structural information of local graphs through a generator and structural codebook, and performing knowledge distillation on the server side.

Object Concepts Emerge from Motion

Haoqian Liang (Beijing University of Posts and Telecommunications), Naiyan Wang (Beijing University of Posts and Telecommunications)

CodeObject DetectionSegmentationDepth EstimationAutonomous DrivingConvolutional Neural NetworkContrastive LearningOptical FlowVideo

🎯 What it does: Generate unsupervised instance masks using motion boundaries in videos, and train a visual encoder through contrastive learning to learn object-centered visual representations.

Object-Centric Concept-Bottlenecks

David Steinmann (TU Darmstadt), Kristian Kersting (TU Darmstadt)

CodeObject DetectionExplainability and InterpretabilityContrastive LearningImageBenchmark

🎯 What it does: This paper proposes and implements Object-Centric Concept Bottlenecks (OCB), which combines object detection with concept extraction to achieve stronger interpretability and performance improvements.

Object-Centric Representation Learning for Enhanced 3D Semantic Scene Graph Prediction

KunHo Heo (Kyung Hee University), MyeongAh Cho (Kyung Hee University)

CodeObject DetectionRepresentation LearningGraph Neural NetworkContrastive LearningPoint Cloud

🎯 What it does: This paper proposes a 3D semantic scene graph prediction framework based on object feature pre-training and relationship feature encoding, significantly improving object recognition and relationship inference performance.

Object-X: Learning to Reconstruct Multi-Modal 3D Object Representations

Gaia Di Lorenzo, Daniel Barath

CodeGenerationCompressionDomain AdaptationRepresentation LearningConvolutional Neural NetworkGaussian SplattingSimultaneous Localization and MappingMultimodalityPoint Cloud

🎯 What it does: For multimodal 3D scenes, we propose the Object-X framework, which learns decodable object-level embeddings that can generate high-quality 3D Gaussian spot reconstructions and can be directly used for downstream tasks such as visual localization, single-image reconstruction, and scene alignment.

Objective Soups: Multilingual Multi-Task Modeling for Speech Processing

A F M Saif (Rensselaer Polytechnic Institute), Tianyi Chen (Cornell Tech)

CodeOptimizationTransformerContrastive LearningAudio

🎯 What it does: A multi-level multi-objective optimization framework (VS-MSP, VC-MSP, VM-MSP) is proposed to unify the training of multilingual speech recognition and translation tasks, addressing the issue of conflicting objectives.

OCN: Effectively Utilizing Higher-Order Common Neighbors for Better Link Prediction

Juntong Wang (Peking University), Muhan Zhang (Peking University)

CodeRecommendation SystemOptimizationGraph Neural NetworkGraph

🎯 What it does: The study focuses on link prediction and proposes the OCN method, which effectively utilizes high-order common neighbors and eliminates redundancy and over-smoothing issues through orthogonalization and path normalization.

OCTDiff: Bridged Diffusion Model for Portable OCT Super-Resolution and Enhancement

Ye Tian (Columbia University), Kaveri A. Thakoor (Columbia University)

CodeRestorationSuper ResolutionConvolutional Neural NetworkDiffusion modelImage

🎯 What it does: Proposes a bridge diffusion model OCTDiff to enhance the resolution and quality of portable OCT images.

Offline Guarded Safe Reinforcement Learning for Medical Treatment Optimization Strategies

Runze Yan (Emory University), Xiao Hu

CodeOptimizationReinforcement LearningBiomedical DataElectronic Health Records

🎯 What it does: This paper proposes a model-based offline safe reinforcement learning framework, OGSRL, aimed at learning treatment strategies from clinical data that are both safe and can surpass those of clinical doctors.

Offline imitation learning in $Q^\pi$-realizable MDPs without expert realizability

Antoine Moulin, Luca Viano

CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabularSequential

🎯 What it does: This paper proposes an offline imitation learning method called SPOIL, which utilizes the linear QΟ€ realizability assumption of MDPs to learn an approximate expert policy from a dataset of expert state-action pairs without requiring the expert to be realizable.

OLinear: A Linear Model for Time Series Forecasting in Orthogonally Transformed Domain

Wenzhen Yue (Peking University), Ji Shi (Peking University)

CodeTransformerTime SeriesFinance Related

🎯 What it does: This paper proposes OLinear, a linear multivariate time series forecasting model that performs encoding and decoding in the orthogonal transformation domain. It utilizes OrthoTrans to adaptively transform the original time series data and then employs the NormLin linear layer to model cross-variable relationships.

Omni-R1: Reinforcement Learning for Omnimodal Reasoning via Two-System Collaboration

Hao Zhong (Zhejiang University), Chunhua Shen (Zhejiang University)

CodeSegmentationReinforcement LearningVideoTextMultimodalityAudio

🎯 What it does: This paper proposes a reinforcement learning-based two-system architecture, Omni-R1, capable of achieving long-term reasoning and fine-grained pixel-level understanding in video-audio-text tasks simultaneously.

OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales

Tung Nguyen (University of California, Los Angeles), Aditya Grover (University of California, Los Angeles)

CodeTransformerDiffusion modelAuto EncoderTime SeriesSequential

🎯 What it does: A model called OmniCast is proposed for probabilistic weather forecasting from short-term to seasonal scales.

OmniConsistency: Learning Style-Agnostic Consistency from Paired Stylization Data

Yiren Song (Show Lab National University of Singapore), Mike Zheng Shou (Show Lab National University of Singapore)

CodeImage TranslationGenerationTransformerDiffusion modelImage

🎯 What it does: A pluggable 'OmniConsistency' plugin is designed to achieve structural and detail consistency in image stylization through Diffusion Transformer, separating style learning and consistency learning in a two-stage training process to achieve high-fidelity consistency across styles.

OmniGaze: Reward-inspired Generalizable Gaze Estimation in the Wild

Hongyu Qu (Nanjing University of Science and Technology), Jinhui Tang (Nanjing Forestry University)

CodePose EstimationDomain AdaptationTransformerReinforcement LearningImage

🎯 What it does: A semi-supervised framework named OMNIGAZE is proposed for estimating 3D gaze direction from facial images in the real world, utilizing a large number of unlabeled facial images and filtering and weighting pseudo-labels through a reward model.

Omnipresent Yet Overlooked: Heat Kernels in Combinatorial Bayesian Optimization

Colin Doumont (ETH Zurich), Henry Moss (University of Cambridge)

CodeOptimizationGraphBenchmark

🎯 What it does: This paper constructs a unified framework through the heat kernel, mapping various combinatorial Bayesian optimization (BO) methods (such as CASMOPOLITAN, COMBO, Bounce, etc.) to the same class of kernels, and proves their equivalence both theoretically and practically.

OmniSegmentor: A Flexible Multi-Modal Learning Framework for Semantic Segmentation

Bo-Wen Yin (Nankai University), Qibin Hou (Nankai University)

CodeSegmentationTransformerImageMultimodality

🎯 What it does: We propose OmniSegmentor, a flexible and configurable semantic segmentation framework suitable for various visual modalities (RGB, depth, thermal imaging, LiDAR, events); simultaneously, we construct the ImageNeXt dataset to provide data support for large-scale multimodal pre-training.

OmniSVG: A Unified Scalable Vector Graphics Generation Model

Yiying Yang, Yu-Gang Jiang

CodeGenerationTransformerVision Language ModelMultimodalityBenchmark

🎯 What it does: We propose OmniSVG, a unified framework for generating complex SVGs using pre-trained vision-language models (VLM);

On Efficiency-Effectiveness Trade-off of Diffusion-based Recommenders

Wenyu Mao (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)

CodeRecommendation SystemTransformerDiffusion modelSequential

🎯 What it does: A two-stage TA-Rec framework is proposed, where in the pre-training stage, time consistency regularization (TCR) is used to smooth the denoising function of the diffusion model, enabling recommendations to be made in a single step; then in the fine-tuning stage, adaptive preference alignment (APA) dynamically adjusts the optimization strength based on time steps and the similarity of positive and negative samples, thereby enhancing the recommendation effectiveness.

On Evaluating LLM Alignment by Evaluating LLMs as Judges

Yixin Liu (Yale University), Arman Cohan (Yale University)

CodeGenerationTransformerLarge Language ModelText

🎯 What it does: This paper studies the consistency (GE-consistency) of large language models (LLMs) in the roles of generation and evaluation, and based on this, proposes an alignment evaluation framework ALIGNEVAL that directly assesses generated results without relying on LLM judges.

On Geometry-Enhanced Parameter-Efficient Fine-Tuning for 3D Scene Segmentation

Liyao Tang (University of Sydney), Dacheng Tao (Nanyang Technological University)

CodeSegmentationConvolutional Neural NetworkSupervised Fine-TuningPoint Cloud

🎯 What it does: Proposed the Geometry Encoding Mixer (GEM) module, achieving geometry-aware parameter-efficient fine-tuning.

On Group Sufficiency Under Label Bias

Haoran Zhang (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)

CodeImageTextTabular

🎯 What it does: Research on learning methods to achieve group sufficiency under label bias conditions.

On Linear Mode Connectivity of Mixture-of-Experts Architectures

Viet-Hoang Tran (National University of Singapore), Tan Minh Nguyen

CodeTransformerMixture of ExpertsImageText

🎯 What it does: This paper studies the linear mode connectivity (LMC) in the mixture of experts (MoE) architecture and provides a matching algorithm to find low-loss linear paths between independently trained models.

On Logic-based Self-Explainable Graph Neural Networks

Alessio Ragno (INSA Lyon), CΓ©line Robardet (INSA Lyon)

CodeExplainability and InterpretabilityKnowledge DistillationGraph Neural NetworkGraph

🎯 What it does: A self-explanatory graph neural network architecture named LogiX-GIN is proposed, which can directly generate interpretable logical rules during the learning process, balancing graphical reasoning and interpretability.

On Reasoning Strength Planning in Large Reasoning Models

Leheng Sheng (National University of Singapore), Tat-Seng Chua (National University of Singapore)

CodeLarge Language ModelText

🎯 What it does: This study explores whether large reasoning models (LRM) plan the reasoning intensity (i.e., the length of reasoning steps) before generating answers. By predicting reasoning length using linear probes, it was found that there exists a single pre-allocated directional vector in the activation space, which encodes reasoning intensity through its magnitude and can causally control the logits of the /think end token, thus achieving controllable reasoning length. Furthermore, two potential applications are proposed: excessive reasoning detection and efficient reasoning.

On scalable and efficient training of diffusion samplers

Minkyu Kim (Korea Advanced Institute of Science and Technology), Minsu Kim (Korea Advanced Institute of Science and Technology)

CodeGenerationData SynthesisOptimizationReinforcement LearningDiffusion modelSequentialStochastic Differential Equation

🎯 What it does: A scalable and efficient diffusion sampler training framework SGDS is proposed, which combines gradient-guided MCMC Searcher with neural diffusion Learner. It achieves a hybrid learning of offline and online sampling through trajectory balancing objectives and utilizes Random Network Distillation (RND) to generate exploration rewards, periodically resetting the model to eliminate primacy bias.

On the Bias of Next-Token Predictors Toward Systematically Inefficient Reasoning: A Shortest-Path Case Study

Riccardo Alberghi (Bocconi University), Luca Saglietti (Bocconi University)

CodeOptimizationComputational EfficiencyTransformerGraphChain-of-Thought

🎯 What it does: In a custom hierarchical directed acyclic graph shortest path task, the author trained a Transformer model based on next word prediction, exploring the impact of different lengths and structures of reasoning trajectories (Chain-of-Thought) on model generalization.

On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity

Quentin Bertrand (University Jean Monnet Saint-Γ‰tienne), RΓ©mi Emonet

CodeFlow-based ModelImage

🎯 What it does: This paper studies the generalization mechanism of flow matching models, exploring whether target noise is the main factor leading to generalization, and proposes a method for training using a closed-form optimal velocity field.

On the creation of narrow AI: hierarchy and nonlocality of neural network skills

Eric J Michaud, Max Tegmark (Massachusetts Institute of Technology)

CodeCompressionOptimizationKnowledge DistillationTransformerSupervised Fine-TuningTabularSequential

🎯 What it does: This study investigates how to derive high-performance, resource-efficient narrow AI models from large general models, explores the impact of task hierarchy on learning, and validates the effects of sparse regularization and pruning in compressing and removing unnecessary skills.

On the Edge of Memorization in Diffusion Models

Sam Buchanan, Valentin De Bortoli

CodeGenerationData SynthesisDiffusion modelImage

🎯 What it does: This paper constructs a theoretical laboratory for systematically analyzing the balance between memorization and generalization in diffusion models. It derives a closed-form approximation of training loss through high-dimensional asymptotic analysis, predicting and validating the linear relationship between model size and the memorization critical point.

On the Empirical Power of Goodness-of-Fit Tests in Watermark Detection

Weiqing He (University of Pennsylvania), Qi Long (University of Pennsylvania)

CodeLarge Language ModelText

🎯 What it does: This paper applies traditional Goodness-of-Fit (GoF) tests to LLM text watermark detection and systematically evaluates the performance of eight GoF methods under three watermark schemes.

On the Entropy Calibration of Language Models

Steven Cao (Stanford University), Percy Liang (Stanford University)

CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: This paper studies the entropy calibration problem of language models, exploring whether the entropy of generated text is consistent with the logarithmic loss of real text;

On the Loss of Context Awareness in General Instruction Fine-tuning

Yihan Wang (University of California Los Angeles), Cho-Jui Hsieh (University of California Los Angeles)

CodeTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: This study investigates the decline in contextual awareness of large language models after supervised instruction fine-tuning and proposes a method to restore this capability through conditional fine-tuning by inserting context-dependent indicators into the instructions.

On the rankability of visual embeddings

Ankit Sonthalia (University of Tuebingen), Seong Joon Oh (University of Tuebingen)

CodeRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage

🎯 What it does: This study explores whether visual embedding models can capture continuous ordinal attributes in a linear manner, specifically whether they possess the property of rankability, and evaluates this across various encoders and attributes.

On the Stability of Graph Convolutional Neural Networks: A Probabilistic Perspective

Ning Zhang (University of Oxford), Xiaowen Dong (University of Oxford)

CodeClassificationAdversarial AttackGraph Neural NetworkGraph

🎯 What it does: This paper proposes a stability analysis framework for Graph Convolutional Neural Networks (GCNN) from a probabilistic perspective, deriving the expected embedding perturbation formulas for graph filters and multi-layer GCNNs under edge perturbations, and based on this, designs a task-independent Prob-PGD adversarial attack method.

On topological descriptors for graph products

Mattie Ji (University of Pennsylvania), Vikas K Garg

CodeGraph Neural NetworkGraph

🎯 What it does: This paper studies the Euler angles and persistent homology of graphs under color-based filtering, and presents an efficient algorithm for computing these topological descriptors on the Cartesian product of graphs.

On Transferring Transferability: Towards a Theory for Size Generalization

Eitan Levin (California Institute of Technology), Soledad Villar (Johns Hopkins University)

CodeDomain AdaptationRepresentation LearningGraph Neural NetworkPoint CloudGraph

🎯 What it does: A unified theoretical framework is proposed to characterize and verify the transferability of models under different input sizes (graphs, sets, point clouds) and the corresponding size generalization.

Once Upon an Input: Reasoning via Per-Instance Program Synthesis

Adam Stein (University of Pennsylvania), Eric Wong (University of Pennsylvania)

CodeOptimizationAI Code AssistantTransformerLarge Language ModelTextBenchmarkChain-of-Thought

🎯 What it does: A Per-Instance Program Synthesis (PIPS) framework is proposed, which can automatically generate and iteratively refine programs at the instance level to complete reasoning tasks.

One Filters All: A Generalist Filter For State Estimation

Shiqi Liu (Tsinghua University), Shengbo Eben Li (Tsinghua University)

CodeTransformerLarge Language ModelPrompt EngineeringTime Series

🎯 What it does: A general state estimation framework called LLM-Filter is proposed, which utilizes pre-trained large language models to embed noisy observations into text and achieves cross-system estimation through contextual prompts.

One for All: Universal Topological Primitive Transfer for Graph Structure Learning

Yide Qiu (Nanjing University of Science and Technology), Zhen Cui (Beijing Normal University)

CodeDomain AdaptationRepresentation LearningGraph Neural NetworkLarge Language ModelContrastive LearningTextGraph

🎯 What it does: This paper proposes the G²SN-Transfer framework, which utilizes dual-stream sequence alignment of topological primitives and text descriptions to achieve knowledge transfer across graph structures.

One Head to Rule Them All: Amplifying LVLM Safety through a Single Critical Attention Head

Junhao Xia (Nanjing University of Science and Technology), Jason Xue

CodeSafty and PrivacyTransformerVision Language ModelMultimodality

🎯 What it does: A defense framework based on a single key attention head (Oh Defense) is proposed, which enhances the security of large visual language models (LVLMs) without requiring additional training.

One Prompt Fits All: Universal Graph Adaptation for Pretrained Models

Yongqi Huang (Tianjin University), Zhiyong Feng (Tianjin University)

CodeDomain AdaptationRepresentation LearningGraph Neural NetworkPrompt EngineeringGraph

🎯 What it does: This paper studies a universal graph prompt learning method called UniPrompt, which can adapt to different downstream tasks through learnable topological prompts while keeping the pre-trained model frozen, making it particularly suitable for few-shot and cross-domain scenarios.

One Stone with Two Birds: A Null-Text-Null Frequency-Aware Diffusion Models for Text-Guided Image Inpainting

Haipeng Liu (Hefei University of Technology), Meng Wang (Hefei University of Technology)

CodeRestorationGenerationDiffusion modelImageText

🎯 What it does: A null-text-null frequency domain aware diffusion model NTN-Diff is proposed for text-guided image inpainting, which can simultaneously preserve the unmasked areas from being damaged and ensure that the inpainted areas are semantically consistent with the text.

One Subgoal at a Time: Zero-Shot Generalization to Arbitrary Linear Temporal Logic Requirements in Multi-Task Reinforcement Learning

Zijian Guo (Boston University), Wenchao Li (Boston University)

CodeReinforcement LearningSequential

🎯 What it does: Proposes the GenZ-LTL method, achieving zero-shot satisfaction of any linear temporal logic (LTL) specifications;

One Token Embedding Is Enough to Deadlock Your Large Reasoning Model

Mohan Zhang (University of North Carolina), Tianlong Chen

CodeComputational EfficiencyAdversarial AttackLarge Language ModelPrompt EngineeringText

🎯 What it does: This paper studies an attack method called Deadlock Attack, which uses a single malicious embedding token to induce large-scale reasoning models (LRMs) into an infinite chain reasoning loop (i.e., 'deadlock'), leading to resource exhaustion.

One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding

Zheyu Aqa Zhang (University of Illinois Urbana-Champaign), Yu-Xiong Wang (University of Illinois Urbana-Champaign)

CodeRecognitionCompressionTransformerLarge Language ModelVideo

🎯 What it does: This paper proposes an extreme video token compression framework called XComp, which retains only one efficient token per frame in the final layer of the LLM for long videos, and further improves information utilization through frame-level filtering.

One-Step is Enough: Sparse Autoencoders for Text-to-Image Diffusion Models

Viacheslav Surkov (Γ‰cole Polytechnique FΓ©dΓ©rale de Lausanne), David Bau (Northeastern University)

CodeGenerationData SynthesisExplainability and InterpretabilityTransformerDiffusion modelAuto EncoderImageTextBenchmark

🎯 What it does: Deconstructs the intermediate representation of SDXL Turbo (a few-step text-to-image diffusion model) using Sparse Autoencoders (SAE) and performs image editing through feature activation; also proposes the RIEBench evaluation benchmark.

One-Step Offline Distillation of Diffusion-based Models via Koopman Modeling

Nimrod Berman (Ben-Gurion University of the Negev), Omri Azencot (Ben-Gurion University of the Negev)

CodeGenerationKnowledge DistillationDiffusion modelImage

🎯 What it does: A method for offline single-step diffusion model distillation based on Koopman theory, KDM, is proposed, which can achieve high-quality image generation with a single forward pass.

Online Optimization for Offline Safe Reinforcement Learning

Yassine Chemingui (Washington State University), Jana Doppa

CodeOptimizationSafty and PrivacyReinforcement LearningTabular

🎯 What it does: This paper proposes an offline safe reinforcement learning framework O3SRL based on online optimization, which iteratively updates the Lagrange multipliers using offline RL or its stochastic approximation operators along with multi-armed bandit algorithms, ultimately obtaining a policy that maximizes rewards while satisfying cost constraints.