arXivSub Start free trial

ICML 2025 Papers — Page 20

International Conference on Machine Learning · 3257 papers

Multi-band Frequency Reconstruction for Neural Psychoacoustic Coding

Dianwen Ng (Nanyang Technological University), EngSiong Chng

CompressionConvolutional Neural NetworkAuto EncoderAudio

🎯 What it does: This paper proposes and implements MUFFIN, a fully convolutional neural psychoacoustic encoder based on multi-band residual vector quantization, achieving high-fidelity compression across different audio types.

Multi-Domain Graph Foundation Models: Robust Knowledge Transfer via Topology Alignment

Shuo Wang (University of Electronic Science and Technology of China), zhao kang

Domain AdaptationGraph Neural NetworkPrompt EngineeringContrastive LearningGraph

🎯 What it does: A multi-domain graph foundation model (MDGFM) is proposed, which achieves cross-domain knowledge transfer through topology alignment and graph structure learning, suitable for sparse noisy graphs.

Multi-Marginal Stochastic Flow Matching for High-Dimensional Snapshot Data at Irregular Time Points

Justin Lee (University of Virginia), Heman Shakeri (University of Virginia)

OptimizationFlow-based ModelTime SeriesBiomedical Data

🎯 What it does: This paper proposes a Multi-Boundary Stochastic Flow Matching (MMSFM) method for trajectory reconstruction of high-dimensional snapshot data at irregular time points.

Multi-Modal Object Re-identification via Sparse Mixture-of-Experts

Yingying Feng (Northeastern University), Jiayi Ji (National University of Singapore)

RecognitionRetrievalTransformerMixture of ExpertsImageMultimodality

🎯 What it does: This paper proposes the MFRNet network, which achieves object re-identification through multi-modal feature fusion and a sparse expert model.

Multi-Objective Causal Bayesian Optimization

Shriya Bhatija (Technical University of Munich), Thomas Bohné (University of Cambridge)

OptimizationGraphTabularBiomedical DataFinance Related

🎯 What it does: Proposes Multi-Objective Causal Bayesian Optimization (MO-CBO) to find Pareto optimal intervention combinations under known causal graphs.

Multi-objective Linear Reinforcement Learning with Lexicographic Rewards

Bo Xue (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)

OptimizationReinforcement Learning

🎯 What it does: A multi-objective linear reinforcement learning algorithm LLRL is proposed and implemented for hierarchical (lexicographic) rewards, providing theoretical convergence (return) guarantees.

Multi-Session Budget Optimization for Forward Auction-based Federated Learning

Xiaoli Tang (Nanyang Technological University), Xiaoxiao Li (University of British Columbia)

OptimizationFederated LearningReinforcement LearningImage

🎯 What it does: This paper proposes a Multi-Session Budget Optimization Strategy (MBOS-AFL) that helps data consumers in federated learning dynamically allocate the total budget in multi-round auctions and bid reasonably in each round to maximize overall utility.

Multi-Stage Manipulation with Demonstration-Augmented Reward, Policy, and World Model Learning

Adrià López Escoriza (ETH Zurich), Hao Su (University of California San Diego)

Robotic IntelligenceReinforcement LearningWorld ModelImage

🎯 What it does: Combining demonstration data, the DEMO framework proposes simultaneous online learning of sparse rewards, staged dense rewards, policies, and world models to achieve long-term multi-stage manipulation tasks from visual input.

Multi-Timescale Dynamics Model Bayesian Optimization for Plasma Stabilization in Tokamaks

Rohit Sonker (Carnegie Mellon University), Jeff Schneider (Carnegie Mellon University)

OptimizationRecurrent Neural NetworkReinforcement LearningTime SeriesPhysics Related

🎯 What it does: A multi-time-domain Bayesian optimization method (DynaBO) is proposed, combining high-frequency data-driven dynamic models and low-frequency Gaussian processes to quickly and experimentally find excitation schemes that suppress crack instabilities in complex systems such as tokamaks.

Multi-Turn Code Generation Through Single-Step Rewards

Arnav Kumar Jain (Mila Quebec AI Institute), Sanjiban Choudhury (Cornell University)

GenerationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText

🎯 What it does: A multi-round code generation framework named µ CODE has been developed, utilizing stepwise rewards to train the generator and validator, improving code quality in multi-round iterations.

Multi-View Graph Clustering via Node-Guided Contrastive Encoding

Yazhou Ren (University of Electronic Science and Technology of China), Lifang He (Lehigh University)

Representation LearningGraph Neural NetworkAuto EncoderContrastive LearningGraph

🎯 What it does: A multi-view graph clustering method based on node-guided contrastive encoding (NGCE) is proposed, which can simultaneously handle homophilic and heterophilic graph data.

Multiaccuracy and Multicalibration via Proxy Groups

Beepul Bharti (Johns Hopkins University), Jeremias Sulam (Johns Hopkins University)

TabularBiomedical DataMagnetic Resonance Imaging

🎯 What it does: This paper studies how to use proxy attributes to assess and improve the multiaccuracy and multicalibration fairness of models when the true sensitive group information is missing.

Multidimensional Adaptive Coefficient for Inference Trajectory Optimization in Flow and Diffusion

Dohoon Lee (Seoul National University), Kyogu Lee (Seoul National University)

GenerationOptimizationDiffusion modelFlow-based ModelGenerative Adversarial NetworkImage

🎯 What it does: A MAC module is proposed to make the inference trajectory of flow/diffusion models dimensionally adaptable, thereby improving generation quality.

Multilayer Matrix Factorization via Dimension-Reducing Diffusion Variational Inference

Junbin Liu (Chinese University of Hong Kong), Wing-Kin Ma (Chinese University of Hong Kong)

Recommendation SystemOptimizationRepresentation LearningDiffusion modelImage

🎯 What it does: This paper proposes a Dimension Reduction Diffusion Variational Inference (DRD-VI) method for probabilistic inference in Multi-layer Matrix Factorization (MMF).

Multimodal Medical Code Tokenizer

Xiaorui Su (Harvard University), Marinka Zitnik (Harvard University)

Recommendation SystemDrug DiscoveryGraph Neural NetworkSupervised Fine-TuningMultimodalityBiomedical DataElectronic Health Records

🎯 What it does: This paper proposes a multimodal medical tokenization model called MEDTOK, which generates a unified discrete vocabulary by jointly utilizing text descriptions and relationships from a medical knowledge graph.

Multinoulli Extension: A Lossless Yet Effective Probabilistic Framework for Subset Selection over Partition Constraints

Qixin Zhang (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)

OptimizationVideo

🎯 What it does: The Multinoulli Extension (ME) is proposed, and a non-parametric continuous greedy/stochastic gradient algorithm, Multinoulli-SCG, is developed for the subset selection problem of near-submodular (α-weak DR submodular or (γ,β)-weak submodular) set functions under partition constraints, with a query complexity of only O(1/ε²).

Multiobjective distribution matching

Xiaoyuan Zhang (City University of Hong Kong), Qingfu Zhang (City University of Hong Kong)

GenerationData SynthesisAuto EncoderImage

🎯 What it does: This study investigates the multi-objective distribution matching problem, deriving the Pareto set and frontier of the exponential family under KL and reverse KL divergence within the framework of information geometry, and based on this, designs a multi-objective variational autoencoder (MOVAE) to achieve smooth interpolation between distributions.

MultiPDENet: PDE-embedded Learning with Multi-time-stepping for Accelerated Flow Simulation

Qi Wang (Renmin University of China), Hao Sun (Renmin University of China)

OptimizationComputational EfficiencyConvolutional Neural NetworkTime SeriesPhysics RelatedOrdinary Differential Equation

🎯 What it does: This paper studies a PDE embedding network called MultiPDENet that integrates numerical methods with machine learning to accelerate fluid simulations using multi-scale time steps.

Multiple-policy Evaluation via Density Estimation

Yilei Chen (Boston University), Ioannis Paschalidis

Reinforcement Learning

🎯 What it does: A multi-strategy evaluation algorithm CAESAR based on density estimation is proposed, which can simultaneously estimate the total reward of multiple strategies in an offline environment.

Multivariate Conformal Selection

Tian Bai (McGill University), Archer Y. Yang (McGill University)

Drug DiscoveryTabularBiomedical DataAlzheimer's Disease

🎯 What it does: A multivariate Conformal Selection (mCS) method is proposed, extending the original CS to the case of multidimensional responses, utilizing regional monotonicity to ensure finite sample FDR control, and achieving efficient candidate selection through two types of inconsistency scores.

MuseControlLite: Multifunctional Music Generation with Lightweight Conditioners

Fang-Duo Tsai (National Taiwan University), Yi-Hsuan Yang (National Taiwan University)

GenerationData SynthesisTransformerDiffusion modelAudio

🎯 What it does: This paper proposes MuseControlLite, a lightweight fine-tuning framework that enables precise control over musical attributes and audio references in text-to-music generation models, and supports audio interpolation and extension.

Mutual Learning for SAM Adaptation: A Dual Collaborative Network Framework for Source-Free Domain Transfer

Yabo Liu (Harbin Institute of Technology), Jinghua Wang

Object DetectionSegmentationDomain AdaptationKnowledge DistillationSupervised Fine-TuningImageMultimodality

🎯 What it does: In the source unsupervised domain transfer scenario, a dual-network collaborative adaptive framework is designed for the Segment Anything Model (SAM) to achieve zero-label transfer to the target domain.

MVA: Linear Attention with High-order Query-Keys Integration and Multi-level Vocabulary Decomposition

wang ning, Guoqi Li (Institute of Automation Chinese Academy of Sciences)

TransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: This paper transfers the pre-trained Softmax-Attention language model to a linear attention model and designs the MVA mechanism based on this.

MxMoE: Mixed-precision Quantization for MoE with Accuracy and Performance Co-Design

Haojie Duanmu (Shanghai Jiao Tong University), Dahua Lin (The Chinese University of Hong Kong)

Mixture of ExpertsText

🎯 What it does: This paper proposes and implements MxMoE, a mixed-precision quantization framework for Mixture-of-Experts models, aimed at balancing model accuracy and inference performance.

N2GON: Neural Networks for Graph-of-Net with Position Awareness

Yejiang Wang (Northeastern University), Xingwei Wang

ClassificationRepresentation LearningGraph Neural NetworkGraphBiomedical DataBenchmark

🎯 What it does: This paper proposes a new graph structure called Graph-of-Net (GON) and designs a position-aware neural network N2GON to learn node representations within GON, handling connections both within and between nodes.

Natural Perturbations for Black-box Training of Neural Networks by Zeroth-Order Optimization

Hiroshi Sawada (NTT Corporation), Yuya Hikima (NTT Corporation)

OptimizationConvolutional Neural NetworkImage

🎯 What it does: This paper proposes a strategy for sampling natural perturbations in black-box zero-order optimization and provides a computable covariance matrix. It then combines this with a block coordinate method to achieve efficient training of large-scale hardware-implemented neural networks.

Navigating Conflicting Views: Harnessing Trust for Learning

Jueqing Lu (Monash University), Lan Du (Monash University)

ClassificationImage

🎯 What it does: A discount mechanism based on computational trust is proposed to resolve conflicts between different views in multi-view classification.

Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning

Fangwen Wu (Zhejiang Lab), Meng Wang (Hefei University of Technology)

ClassificationKnowledge DistillationTransformerSupervised Fine-TuningImage

🎯 What it does: In category incremental learning unrelated to research tasks, semantic drift caused by low-rank adaptation is addressed by proposing two calibration methods: mean shift compensation and covariance calibration, combined with feature-level self-distillation to improve model stability.

Navigating the Social Welfare Frontier: Portfolios for Multi-objective Reinforcement Learning

Cheol Woo Kim (Harvard University), Swati Gupta (Massachusetts Institute of Technology)

OptimizationReinforcement LearningTabular

🎯 What it does: The study generates α-approximate portfolios for all social welfare functions p-mean with p≤1 in multi-objective reinforcement learning, proposes the p-MEANPORTFOLIO algorithm based on linear search with theoretical guarantees; also introduces the heuristic BUDGETCONSTRAINEDPORTFOLIO with budget constraints and warm-start acceleration techniques.

NBDI: A Simple and Effective Termination Condition for Skill Extraction from Task-Agnostic Demonstrations

Myunsoo Kim (Korea University), Byung-Jun Lee (Gauss Labs Inc.)

Robotic IntelligenceReinforcement LearningSequential

🎯 What it does: This paper proposes a decision point identification method based on state-action novelty, NBDI, which can learn terminable skills from task-agnostic demonstrations.

Near Optimal Best Arm Identification for Clustered Bandits

Yash (Indian Institute of Technology Bombay), Nikhil Karamchandani (Indian Institute of Technology Bombay)

Recommendation SystemOptimizationTabular

🎯 What it does: This paper studies the optimal arm identification problem in the multi-agent multi-armed bandit (F-MAB) setting and designs a two-stage clustering-identification algorithm under unknown clustering.

Near Optimal Non-asymptotic Sample Complexity of 1-Identification

ZITIAN LI, Wang Chi Cheung (National University of Singapore)

Reinforcement Learning

🎯 What it does: A new 1-recognition algorithm SEE is proposed and analyzed, providing matching upper and lower bounds for non-asymptotic sample complexity;

Near-Optimal Consistency-Robustness Trade-Offs for Learning-Augmented Online Knapsack Problems

Mohammadreza Daneshvaramoli (University of Massachusetts), Mohammad Hajiesmaili (University of Massachusetts)

OptimizationReinforcement LearningTabularFinance Related

🎯 What it does: A series of learning-augmented algorithms are proposed for the online knapsack problem, aiming to achieve a near-optimal trade-off between consistency and robustness.

Near-Optimal Decision Trees in a SPLIT Second

Varun Babbar (Duke University), Margo Seltzer (University of British Columbia)

ClassificationOptimizationComputational EfficiencyTabular

🎯 What it does: The SPLIT series algorithms (SPLIT, LicketySPLIT, and RESPLIT) are proposed, which achieve approximately optimal decision trees by using dynamic programming + branch-and-bound search in the shallow layers of the tree and greedy splitting in the deeper layers, allowing for the rapid generation of the Rashomon set.

Near-optimal Regret Using Policy Optimization in Online MDPs with Aggregate Bandit Feedback

Tal Lancewicki (Tel Aviv University), Yishay Mansour (Google Research)

OptimizationReinforcement Learning

🎯 What it does: This paper studies online finite-horizon Markov decision processes (MDPs) with adversarial loss and feedback from aggregation bands, proposing the first policy optimization algorithm suitable for this setting.

Near-Optimal Sample Complexity for MDPs via Anchoring

Jongmin Lee (Seoul National University), Roberto Cominetti (Pontificia Universidad Catolica de Chile)

OptimizationReinforcement Learning

🎯 What it does: A new model-free algorithm is proposed for computing ε-optimal policies in average reward Markov decision processes (MDPs), particularly in weak communication settings.

Near-optimal Sketchy Natural Gradients for Physics-Informed Neural Networks

Maricela Best Mckay, Brian Wetton (University of British Columbia)

OptimizationComputational EfficiencyTabularPhysics Related

🎯 What it does: A randomized natural gradient algorithm (SNGD) is proposed, which significantly reduces the computational cost and memory usage during training by compressing the Gram matrix of PINN.

NEAR: Neural Electromagnetic Array Response

Yinyan Bu (University of California San Diego), Piya Pal (University of California San Diego)

Super ResolutionOptimizationTime SeriesPhysics Related

🎯 What it does: The NEAR framework is proposed, utilizing untrained implicit neural representations (INR) to predict the complex array response at arbitrary locations under sparse radar array measurements, achieving angle super-resolution.

Nearly Optimal Algorithms for Contextual Dueling Bandits from Adversarial Feedback

Qiwei Di (University of California), Quanquan Gu (University of California)

OptimizationAdversarial AttackReinforcement LearningTabular

🎯 What it does: This paper studies and proposes the Robust Contextual Dual-Handed Bow (RCDB) algorithm, which addresses the contextual adversarial sampling problem where strong adversaries can flip labels on comparable results, and designs an improved version RCDB-S for the Sigmoid link function.

Nearly Optimal Sample Complexity for Learning with Label Proportions

Robert Istvan Busa-Fekete (Google Research), Uri Stemmer (Tel Aviv University)

ClassificationOptimizationConvolutional Neural NetworkImageTabular

🎯 What it does: This paper proposes a learning from label proportions (LLP) method using variance reduction techniques and provides almost optimal upper and lower bounds on sample complexity under squared loss.

NegMerge: Sign-Consensual Weight Merging for Machine Unlearning

Hyo Seo Kim (Sogang University), Junsuk Choe (Sogang University)

ClassificationData-Centric LearningTransformerVision Language ModelImage

🎯 What it does: A novel machine forgetting method called NegMerge is proposed, which effectively deletes specified knowledge in the model by merging task vectors from multiple models.

Neighbour-Driven Gaussian Process Variational Autoencoders for Scalable Structured Latent Modelling

Xinxing Shi (University of Manchester), Mauricio A Álvarez

Auto EncoderVideoSequential

🎯 What it does: In structured data, two approximate methods for Gaussian Process Variational Autoencoder (GPVAE) are proposed, which efficiently compute in mini-batch training by expanding the Gaussian process (GP) prior only among the nearest neighbors of each sample: Hierarchical Prior Approximation (HPA) and Sparse Precision Approximation (SPA);

Nemotron-CORTEXA: Enhancing LLM Agents for Software Engineering Tasks via Improved Localization and Solution Diversity

Atefeh Sohrabizadeh (NVIDIA), Bryan Catanzaro (NVIDIA)

AI Code AssistantTransformerLarge Language ModelContrastive LearningTextBenchmark

🎯 What it does: The NEMOTRON-CORTEXA system is proposed, utilizing a customized code embedding model, location agents based on AST and LSP, and diversified patch generation to enhance the efficiency of locating and fixing LLM software engineering tasks.

Nested Expectations with Kernel Quadrature

Zonghao Chen (University College London), Francois-Xavier Briol (University College London)

OptimizationComputational EfficiencyTabularFinance Related

🎯 What it does: This paper proposes a Nested Kernel Quadrature (NKQ) estimator based on kernel quadrature to effectively solve nested expectations.

Nesterov Method for Asynchronous Pipeline Parallel Optimization

Thalaiyasingam Ajanthan (Pluralis Research), Alexander Long (Pluralis Research)

OptimizationTransformerLarge Language ModelText

🎯 What it does: This paper proposes an asynchronous pipeline parallel optimization method based on Nesterov Accelerated Gradient (NAG) to alleviate convergence issues caused by stale gradients.

NestQuant: nested lattice quantization for matrix products and LLMs

Semyon Savkin (Massachusetts Institute of Technology), Yury Polyanskiy (Massachusetts Institute of Technology)

TransformerLarge Language ModelText

🎯 What it does: Introducing NestQuant - a post-training quantization scheme based on the Gosset 8-dimensional nested lattice (Voronoi code) and random rotation, capable of low-bit (4-bit) quantization for LLM weights, KV cache, and activations simultaneously;

NETS: A Non-equilibrium Transport Sampler

Michael Samuel Albergo, Eric Vanden-Eijnden (New York University)

Physics RelatedStochastic Differential Equation

🎯 What it does: A new non-equilibrium transport sampler (NETS) is proposed, which incorporates a learned drift term into Langevin dynamics to smoothly transfer the base distribution to the target distribution within a finite time, achieving unbiased estimation through the Jarzynski equality.

Network Sparsity Unlocks the Scaling Potential of Deep Reinforcement Learning

Guozheng Ma (Nanyang Technical University), Dacheng Tao (Nanyang Technical University)

Reinforcement LearningSequential

🎯 What it does: In deep reinforcement learning, static sparse networks are achieved through one-time random pruning, investigating whether they can break the scale limits of existing advanced architectures (such as SimBa).

Neural Collapse Beyond the Unconstrained Features Model: Landscape, Dynamics, and Generalization in the Mean-Field Regime

Diyuan Wu (Institute of Science and Technology Austria), Marco Mondelli

OptimizationConvolutional Neural NetworkImage

🎯 What it does: This paper studies the training dynamics and loss landscape of three-layer fully connected neural networks under the mean-field limit, proving that the intra-class variance of the network converges to zero (NC1) during the gradient flow training process, and further demonstrating that NC1 and zero test error can coexist under well-separated data distributions.

Neural Discovery in Mathematics: Do Machines Dream of Colored Planes?

Konrad Mundinger (Zuse Institute Berlin), Sebastian Pokutta (Zuse Institute Berlin)

Optimization

🎯 What it does: This paper proposes a method to transform the Hadwiger-Nelson problem into a continuous optimization task, utilizing neural networks for gradient search on a differentiable loss function, ultimately discovering a new six-color plane coloring scheme.

Neural Encoding and Decoding at Scale

Yizi Zhang (Columbia University), Cole Lincoln Hurwitz

TransformerMultimodalityBiomedical Data

🎯 What it does: A multimodal multitask Transformer model named NEDS is proposed, capable of simultaneously performing neural encoding and behavioral decoding within the same framework.

Neural Event-Triggered Control with Optimal Scheduling

Luan Yang (Fudan University), Wei Lin (Fudan University)

OptimizationTime SeriesOrdinary Differential Equation

🎯 What it does: A neural network event-triggered control framework, Neural ETC, is proposed to achieve stable control of the system under limited communication resources.

Neural Genetic Search in Discrete Spaces

Hyeonah Kim (Quebec Artificial Intelligence Institute), Changhyun Kwon (KAIST)

GenerationOptimizationLarge Language ModelTextGraph

🎯 What it does: This paper proposes a search method that integrates a genetic algorithm mechanism during the testing phase of deep generative models—Neural Genetic Search (NGS).

Neural Graph Matching Improves Retrieval Augmented Generation in Molecular Machine Learning

Runzhong Wang (Massachusetts Institute of Technology), Connor W. Coley (Massachusetts Institute of Technology)

GenerationRetrievalDrug DiscoveryGraph Neural NetworkTransformerGraphRetrieval-Augmented Generation

🎯 What it does: This paper proposes and implements MARASON, a retrieval-augmented generation framework that integrates neural graph matching into mass spectrometry simulation tasks.

Neural Guided Diffusion Bridges

Gefan Yang (University of Copenhagen), Stefan Sommer (University of Copenhagen)

Diffusion model

🎯 What it does: This paper proposes a neural guided diffusion bridge method for efficiently simulating conditional diffusion processes.

Neural Interpretable PDEs: Harmonizing Fourier Insights with Attention for Scalable and Interpretable Physics Discovery

Ning Liu (Lehigh University), Yue Yu (Lehigh University)

OptimizationExplainability and InterpretabilityComputational EfficiencyTabularPhysics Related

🎯 What it does: This study proposes a novel neural operator architecture, NIPS, which can simultaneously perform forward PDE solving and inverse physical mechanism discovery, suitable for multi-system and multi-scale physical problems.

Neural Representational Consistency Emerges from Probabilistic Neural-Behavioral Representation Alignment

Yu Zhu (Chinese Academy of Sciences), Tiejun Huang (Beijing Academy of Artificial Intelligence)

Representation LearningRecurrent Neural NetworkAuto EncoderTime SeriesSequential

🎯 What it does: A probabilistic neural-behavior representation alignment framework named PNBA is proposed and implemented, achieving cross-individual and cross-species zero-shot alignment and validation under multi-scale neural variations through joint encoding, decoding, and probabilistic matching.

Neural Solver Selection for Combinatorial Optimization

Chengrui Gao (Nanjing University), Chao Qian (Nanjing University)

OptimizationGraph Neural NetworkGraph

🎯 What it does: A general framework is proposed to select the most suitable neural solver at the instance level to solve combinatorial optimization problems.

NeuralCohort: Cohort-aware Neural Representation Learning for Healthcare Analytics

Changshuo Liu (National University of Singapore), James Wei Luen Yip (National University Hospital)

Representation LearningRecurrent Neural NetworkGraph Neural NetworkContrastive LearningTabularBiomedical DataElectronic Health Records

🎯 What it does: A clustering-based NeuralCohort method is proposed to construct fine-grained patient cohorts and enhance representation learning in electronic health record (EHR) data.

NeuronTune: Towards Self-Guided Spurious Bias Mitigation

Guangtao Zheng (University of Virginia), Aidong Zhang (University of Virginia)

ClassificationOptimizationImageTextMultimodality

🎯 What it does: A post-hoc unsupervised method called NeuronTune is proposed, which alleviates the model's dependence on spurious correlations by identifying and suppressing the neurons that produce bias within the embedded space of a trained model.

Neurosymbolic World Models for Sequential Decision Making

Leonardo Hernandez Cano, Armando Solar-Lezama (Massachusetts Institute of Technology)

OptimizationRobotic IntelligenceReinforcement LearningSequential

🎯 What it does: With the help of offline data, unsupervised methods are used to learn neural symbolic finite state machine (FSM) primitives, which are then combined into FSM world models tailored for specific environments, ultimately achieving policy optimization in a model-driven reinforcement learning loop.

NeuroTree: Hierarchical Functional Brain Pathway Decoding for Mental Health Disorders

Jun-En Ding (Stevens Institute of Technology), Feng Liu (Stevens Institute of Technology)

Graph Neural NetworkContrastive LearningGraphBiomedical DataMagnetic Resonance ImagingOrdinary Differential Equation

🎯 What it does: This paper proposes the NEUROTREE framework, which combines k-hop AGE-GCN, neural ODE, and Contrastive Masked Functional Connectivity (CMFC) to decode fMRI functional networks in a tree structure, enabling hierarchical analysis and interpretation of brain pathways in mental disorders.

Neutral residues: revisiting adapters for model extension

Franck SIGNE TALLA (Kyutai), Herve Jegou (Kyutai)

Domain AdaptationTransformerLarge Language ModelMixture of ExpertsTextBiomedical Data

🎯 What it does: This study explores how to extend existing large language models (LLMs) to new domains (such as additional languages) without retraining, by introducing improved adapters (neutral residues), while enhancing performance in the new domain while maintaining original performance.

New Bounds for Sparse Variational Gaussian Processes

Michalis Titsias

Tabular

🎯 What it does: A new Sparse Variational Gaussian Process (SVGP) method is proposed, which derives a tighter evidence lower bound by relaxing the assumptions on the variational distribution, thereby improving the model's predictive performance.

NextCoder: Robust Adaptation of Code LMs to Diverse Code Edits

Tushar Aggarwal (Microsoft Research India), Nagarajan Natarajan (Microsoft Research India)

Data SynthesisDomain AdaptationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText

🎯 What it does: A synthetic data pipeline has been developed to generate diverse code editing examples, and a robust model adaptation algorithm called SeleKT has been proposed to transfer large code language models (such as QwenCoder-2.5) to code editing tasks, resulting in a series of models named NextCoder.

NExtLong: Toward Effective Long-Context Training without Long Documents

Chaochen Gao (Institute of Information Engineering, Chinese Academy of Sciences), Songlin Hu (Institute of Information Engineering, Chinese Academy of Sciences)

TransformerLarge Language ModelTextBenchmarkRetrieval-Augmented Generation

🎯 What it does: The NExtLong framework is proposed, which extends long-context data through negative documents and trains LLMs;

NICE Data Selection for Instruction Tuning in LLMs with Non-differentiable Evaluation Metric

Jingtan Wang (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)

TransformerLarge Language ModelReinforcement LearningText

🎯 What it does: A method for impact estimation based on non-differentiable evaluation metrics, called NICE, is proposed to select the most valuable samples for specific downstream tasks from large-scale instruction tuning datasets.

NMA-tune: Generating Highly Designable and Dynamics Aware Protein Backbones

Urszula Julia Komorowska (University of Cambridge), Mateja Jamnik (University of Cambridge)

Protein Structure PredictionGraph Neural NetworkDiffusion modelBiomedical Data

🎯 What it does: A dynamic conditional protein scaffold generation method based on diffusion models, called NMA-tune, is proposed.

No Free Lunch from Random Feature Ensembles: Scaling Laws and Near-Optimality Conditions

Benjamin Samuel Ruben, Cengiz Pehlevan (Harvard University)

Image

🎯 What it does: Under a fixed total model size (total number of features), this paper compares the generalization performance of a single large random feature ridge regression model with an ensemble model composed of multiple smaller models, exploring when near-optimal performance can be achieved.

No Metric to Rule Them All: Toward Principled Evaluations of Graph-Learning Datasets

Corinna Coupette (Aalto University), Bastian Rieck (University of Fribourg)

Data-Centric LearningGraph Neural NetworkGraph

🎯 What it does: Proposes the RINGS framework, which evaluates the quality of graph learning datasets through pattern perturbation (graph structure or node features); introduces two metrics: performance separability and pattern complementarity.

No Soundness in the Real World: On the Challenges of the Verification of Deployed Neural Networks

Attila Szász (University of Szeged), Márk Jelasity (University of Szeged)

Adversarial AttackConvolutional Neural NetworkImage

🎯 What it does: This paper proves that theoretically claimable verifiable neural networks are not necessarily secure in real deployment environments, and demonstrates through the construction of backdoored networks that can be triggered in specific deployed environments that existing state-of-the-art verifiers cannot guarantee actual deployment security.

No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspaces

Daniel Marczak (Warsaw University of Technology), Joost van de Weijer (Universitat Autonoma de Barcelona)

OptimizationRepresentation LearningTransformerSupervised Fine-TuningImageText

🎯 What it does: This paper proposes two model merging methods based on singular value decomposition—Iso-C and Iso-CTS. These methods utilize isotropic scaling to uniformize the singular value spectrum of the task matrix and merge the common subspace with task-specific subspaces to enhance subspace alignment and overall multi-task performance.

No-Regret is not enough! Bandits with General Constraints through Adaptive Regret Minimization

Martino Bernasconi (Bocconi University), Andrea Celli (Bocconi University)

OptimizationReinforcement Learning

🎯 What it does: This paper studies the multi-armed bandit problem with general constraints and proposes a new algorithmic framework aimed at maximizing cumulative rewards while minimizing constraint violations.

Noise Conditional Variational Score Distillation

Xinyu Peng (Shanghai Jiao Tong University), Hongkai Xiong (Shanghai Jiao Tong University)

RestorationGenerationDiffusion modelScore-based ModelGenerative Adversarial NetworkImage

🎯 What it does: Proposes Noise Conditional Variational Score Distillation (NCVSD), which distills a pre-trained diffusion model into a generative denoiser capable of multi-step sampling and supports zero-shot inference;

Noise-Guided Predicate Representation Extraction and Diffusion-Enhanced Discretization for Scene Graph Generation

Guoqing Zhang (Beijing Jiaotong University), Yigang Cen (Beijing Jiaotong University)

GenerationData SynthesisGraph Neural NetworkTransformerDiffusion modelImage

🎯 What it does: By utilizing three main modules: noise-guided predicate representation extraction, conditional diffusion enhancement, and discretization mapping, the diversity and homogeneity of predicate representations in scene graph generation are improved, effectively alleviating the long-tail bias problem.

Noisy SIGNSGD Is More Differentially Private Than You (Might) Think

Richeng Jin (Zhejiang University), Huaiyu Dai (North Carolina State University)

OptimizationFederated LearningSafty and PrivacyImage

🎯 What it does: This paper studies the amplification effect of symbolic compression on differential privacy and proves that in distributed learning, Noisy SIGNSGD can maintain 32× communication compression while achieving comparable privacy and convergence performance to traditional DP-SGD.

NoLiMa: Long-Context Evaluation Beyond Literal Matching

Ali Modarressi (Ludwig Maximilian University of Munich), Hinrich Schuetze

TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought

🎯 What it does: Proposes the NOLIMA benchmark for evaluating the associative reasoning ability of LLMs beyond literal matching in long contexts.

Non-Asymptotic and Non-Lipschitzian Bounds on Optimal Values in Stochastic Optimization Under Heavy Tails

Jindong Tong (University of Florida), Johannes O. Royset (University of Southern California)

OptimizationTabularTime SeriesFinance Related

🎯 What it does: Constructed computable confidence intervals under non-asymptotic, non-Lipschitz, and heavy-tailed random distributions using Sample Average Approximation (SAA) and its improved form Diameter Risk Minimization (DRM).

Non-asymptotic Error Bounds in $\mathcal{W}_2$-Distance with Sqrt(d) Dimension Dependence and First Order Convergence for Langevin Monte Carlo beyond Log-Concavity

Bin Yang (Central South University), Xiaojie Wang (Central South University)

OptimizationStochastic Differential Equation

🎯 What it does: This paper re-examines the Langevin Monte Carlo (LMC) sampling algorithm and provides a non-asymptotic error analysis of the W2 distance in non-convex settings, proving that under specific logarithmic smoothness conditions, the error bound is O(√dh), and guarantees the mixing time is O(√dϵ^-1).

Non-Asymptotic Length Generalization

Thomas Chen (Stanford University), Zhiyuan Li (Toyota Technological Institute at Chicago)

🎯 What it does: This paper provides provable guarantees for non-asymptotic length generalization, defines length complexity, and demonstrates the application of the minimum complexity interpolation learning algorithm across different function classes.

Non-stationary Diffusion For Probabilistic Time Series Forecasting

Weiwei Ye (Central South University), Ning Gui (Central South University)

Diffusion modelTime Series

🎯 What it does: This paper proposes a non-stationary diffusion model NsDiff, which redesigns the diffusion process of DDPM using the Location-Scale Noise Model (LSNM) to better capture the variable uncertainty of time series.

Non-stationary Online Learning for Curved Losses: Improved Dynamic Regret via Mixability

Yu-Jie Zhang (RIKEN AIP), Masashi Sugiyama (University of Tokyo)

Optimization

🎯 What it does: This paper proposes an online learning framework that utilizes mixability to achieve dynamic regret improvements for convex loss in non-stationary environments.

Non-Stationary Predictions May Be More Informative: Exploring Pseudo-Labels with a Two-Phase Pattern of Training Dynamics

Hongbin Pei (Xi'an Jiaotong University), Xiaohong Guan (Xi'an Jiaotong University)

ClassificationRecognitionContrastive LearningImageGraph

🎯 What it does: Proposed a 2-phasic metric to identify two-stage pseudo-labels and enhance pseudo-label learning with specialized loss.

Nonconvex Theory of $M$-estimators with Decomposable Regularizers

Weiwei Liu (Wuhan University)

Optimization

🎯 What it does: This paper studies the M-estimator theory of non-convex loss functions in high-dimensional settings, particularly whether the estimation error remains within a bounded set when the loss function is non-convex, and whether the convergence rate can still be recovered.

Nonlinear transformers can perform inference-time feature learning

Naoki Nishikawa (University of Tokyo), Taiji Suzuki (University of Tokyo)

TransformerTabular

🎯 What it does: This study explores how pre-trained transformers perform feature learning during inference, particularly focusing on the contextual learning ability of single-index models.

Nonlinearly Preconditioned Gradient Methods under Generalized Smoothness

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

OptimizationTabular

🎯 What it does: Analyzed the nonlinear preconditioned gradient method to solve smooth minimization problems, proposing a generalized smoothness property based on abstract convexity that transcends Lipschitz smoothness, and providing sufficient first and second-order conditions.

Nonparametric Identification of Latent Concepts

Yujia Zheng (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)

Flow-based ModelImage

🎯 What it does: A non-parametric theoretical framework is proposed, demonstrating that when observations from different categories are sufficiently diverse, hidden concepts and their connection structures can be theoretically identified from the observational data.

Nonparametric Modern Hopfield Models

Jerry Yao-Chieh Hu (Northwestern University), Han Liu (Northwestern University)

RetrievalComputational EfficiencyTransformerImageText

🎯 What it does: A unified framework is proposed to view the modern Hopfield model as a non-parametric regression problem, and based on this framework, an efficient variant with a sparse structure is derived.

Nonparametric Teaching for Graph Property Learners

Chen Zhang (University of Hong Kong), Ngai Wong (University of Hong Kong)

Graph Neural NetworkGraph

🎯 What it does: From the perspective of non-parametric teaching, the GraNT method is proposed to enhance the training efficiency of graph attribute learners (GCN).

Normalizing Flows are Capable Generative Models

Shuangfei Zhai (Apple), Joshua M. Susskind (Apple)

GenerationData SynthesisTransformerScore-based ModelFlow-based ModelImage

🎯 What it does: This paper proposes a Transformer-based Block Autoregressive Flow architecture called TARFLOW, which constructs a powerful Normalizing Flow generative model through Gaussian noise augmentation, score-based denoising, and guidance techniques.

Not all solutions are created equal: An analytical dissociation of functional and representational similarity in deep linear neural networks

Lukas Braun (University of Oxford), Andrew M Saxe

Image

🎯 What it does: This paper analyzes deep linear networks and their corresponding nonlinear networks, studying whether the similarity of functions and representations can be decoupled, and proposes a solution space structure where the functions are the same but the representations can be completely different.

Not All Tokens Matter All The Time: Dynamic Token Aggregation Towards Efficient Detection Transformers

Jiacheng Cheng (Northwestern Polytechnical University), Junwei Han (Northwestern Polytechnical University)

Object DetectionComputational EfficiencyTransformerImage

🎯 What it does: This paper proposes Dynamic DETR, which utilizes dynamic token aggregation to achieve multi-layer token sparsification at each stage of the DETR encoder, significantly reducing computational load.

Not All Wrong is Bad: Using Adversarial Examples for Unlearning

Ali Ebrahimpour-Boroojeny (University of Illinois), Varun Chandrasekaran (University of Illinois)

Adversarial AttackData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImage

🎯 What it does: This paper proposes an approximate machine unlearning method called AMUN, which can reduce the prediction confidence of forgotten samples while maintaining the overall accuracy of the model.

Novelty Detection in Reinforcement Learning with World Models

Geigh Zollicoffer (Georgia Institute of Technology), Mark Riedl (Georgia Institute of Technology)

Anomaly DetectionReinforcement LearningWorld ModelSequential

🎯 What it does: This paper proposes a KL divergence upper bound based on a world model to automatically detect novelty in reinforcement learning environments without the need for manual thresholds.

NTK-DFL: Enhancing Decentralized Federated Learning in Heterogeneous Settings via Neural Tangent Kernel

Gabriel Thompson (North Carolina State University), Huaiyu Dai (North Carolina State University)

Federated LearningImage

🎯 What it does: This paper proposes a decentralized federated learning framework called NTK-DFL, which drives model weight evolution through the Neural Tangent Kernel (NTK) and achieves global model learning by averaging parameters among neighbors.

NTPP: Generative Speech Language Modeling for Dual-Channel Spoken Dialogue via Next-Token-Pair Prediction

Qichao Wang (Tencent), Peilin Zhao (Tencent)

GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningAudio

🎯 What it does: A dual-channel speech generation language model based on next-token-pair prediction (NTPP) is proposed, utilizing a decoder-only Transformer to directly learn the joint distribution of two speakers, achieving speaker-independent real-time bidirectional dialogue.

O-MAPL: Offline Multi-agent Preference Learning

The Viet Bui (Singapore Management University), Thanh Hong Nguyen (University of Oregon)

Reinforcement LearningTabularBenchmark

🎯 What it does: An offline multi-agent preference learning framework named O-MAPL is proposed, which directly trains policies end-to-end on preference data, eliminating the explicit reward modeling step.

Objective drives the consistency of representational similarity across datasets

Laure Ciernik (Technische Universität Berlin), Lukas Muttenthaler (Technische Universität Berlin)

Representation LearningConvolutional Neural NetworkContrastive LearningImage

🎯 What it does: A comparison of the internal representations of 64 visual models across 23 different datasets is conducted, proposing a framework based on similarity vectors to measure the consistency of model representation similarity across different datasets.

Observation Interference in Partially Observable Assistance Games

Scott Emmons (University of California), Stuart Russell (University of California)

OptimizationReinforcement Learning from Human Feedback

🎯 What it does: This paper studies partially observable assistive games (POAG) and analyzes why AI assistants may have the motivation to interfere with human observation information when both share rewards but only have partial observability. It theoretically proves three main motivations: transmitting private information, querying human preferences, and alleviating human bounded rationality; and experimentally validates the trade-offs and benefits of interfering with observations.

Occult: Optimizing Collaborative Communications across Experts for Accelerated Parallel MoE Training and Inference

Shuqing Luo (University of North Carolina), Tianlong Chen (University of North Carolina)

OptimizationComputational EfficiencyLarge Language ModelMixture of ExpertsText

🎯 What it does: An algorithm named Occult is proposed - a system-coordinated design scheme to optimize the all-to-all communication of experts in the Mixture-of-Experts (MoE) model, thereby accelerating the training and inference of MoE LLMs.