ICLR 2026 Papers — Page 32
International Conference on Learning Representations · 5356 papers
NeuCLIP: Efficient Large-Scale CLIP Training with Neural Normalizer Optimization
Xiyuan Wei (Texas A&M University), Tianbao Yang (Texas A&M University)
OptimizationComputational EfficiencyRepresentation LearningVision Language ModelContrastive LearningMultimodality
🎯 What it does: Develop NeuCLIP, a CLIP training framework that uses neural networks to predict the normalization term of the contrastive loss;
Neural Collapse in Multi-Task Learning
Youjun Wang (Wuhan University), Weiwei Liu (Wuhan University)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: Investigate the neural collapse (NC) phenomenon in neural networks for multi-task learning (MTL), observing new NC structures in two scenarios: single-source multi-task classification (SSMTC) and multi-source multi-task classification (MSMTC), with theoretical proofs provided.
Neural Compression of 3D Meshes using Sparse Implicit Representation
Jianqiang Wang (City University of Hong Kong), Junhui Hou (City University of Hong Kong)
CompressionConvolutional Neural NetworkAuto EncoderMesh
🎯 What it does: This paper proposes a 3D mesh compression framework based on sparse implicit representation (SIR) and sparse neural compression (SNC), converting mesh surfaces into sparse SDFs retaining only the regions near the surface and compressing them into low-bitstreams via sparse convolutional autoencoders.
Neural Dynamics Self-Attention for Spiking Transformers
Dehao Zhang (University of Electronic Science and Technology of China), Haizhou Li (Shenzhen Loop Area Institute)
ClassificationSegmentationSpiking Neural NetworkTransformerImage
🎯 What it does: Proposed self-attention mechanisms based on local receptive fields and neuronal dynamics (LRF-SSA and LRF-Dyn) to enhance the local modeling capability of spiking Transformers and reduce memory consumption during inference.
Neural Force Field: Few-shot Learning of Generalized Physical Reasoning
Shiqian Li (Peking University), Yixin Zhu (Peking University)
Meta LearningGraph Neural NetworkPhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposes the Neural Force Field (NFF) framework, which learns explicit force fields using neural operators and performs continuous-time prediction via ODE solvers, enabling physics reasoning and planning with few samples.
Neural Graduated Assignment for Maximum Common Edge Subgraphs
Chaolong Ying (Chinese University of Hong Kong), Tianshu Yu (Chinese University of Hong Kong)
OptimizationGraph Neural NetworkGraphBiomedical Data
🎯 What it does: Propose a trainable temperature gradient assignment (Neural Graduated Assignment, NGA) framework based on neural networks for approximately solving the maximum common edge subgraph (MCES) problem.
Neural Hamilton--Jacobi Characteristic Flows for Optimal Transport
Yesom Park (University of California, Los Angeles), Stanley Osher (University of California, Los Angeles)
OptimizationComputational EfficiencyFlow-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Propose a single-network neural feature flow (NCF) framework based on the Hamilton-Jacobi (HJ) equation to learn optimal transport (OT) mappings under arbitrary costs, including both forward and backward directions.
Neural Latent Arbitrary Lagrangian-Eulerian Grids for Fluid-Solid Interaction
Shilong Tao (Peking University), Yunhuai Liu (Peking University)
TransformerMeshTime SeriesSequentialPhysics Related
🎯 What it does: Designed and trained a deep learning framework called Fisale, based on multi-scale potential ALE grids and partitioned coupling modules, for predicting two-way fluid-structure interaction problems.
Neural Message-Passing on Attention Graphs for Hallucination Detection
Fabrizio Frasca (Technion), Haggai Maron (Technion)
Anomaly DetectionGraph Neural NetworkLarge Language ModelText
🎯 What it does: Construct an attributed graph using attention maps and activation information, perform graph neural network (GNN) information propagation on the internal computation trajectory of LLMs, and complete hallucination detection in generated text.
Neural Multi-Objective Combinatorial Optimization for Flexible Job Shop Scheduling Problems
Igor G. Smit (Eindhoven University of Technology), Wim P.M. Nuijten (Eindhoven University of Technology)
OptimizationGraph Neural NetworkReinforcement LearningGraphBenchmark
🎯 What it does: Proposed a decomposition-based reinforcement learning framework that generates Pareto fronts for multi-objective flexible job shop scheduling problems using a single neural network.
Neural Networks Learn Generic Multi-Index Models Near Information-Theoretic Limit
Bohan Zhang (Peking University), Jason D. Lee (UC Berkeley)
OptimizationRepresentation Learning
🎯 What it does: Investigates whether two-layer neural networks trained with hierarchical gradient descent can approximately learn multi-exponential models up to the information-theoretic limit under Gaussian inputs.
Neural Optimal Transport Meets Multivariate Conformal Prediction
Vladimir Kondratyev (MBZUAI), Eric Moulines (Ecole Polytechnique)
OptimizationComputational EfficiencyTabular
🎯 What it does: Combine neural optimal transport (OT) and vector quantile regression (VQR) to construct a multivariate conformal prediction model, generating distribution-agnostic, conditionally adaptive prediction sets by leveraging learned multivariate rank information.
Neural Posterior Estimation with Latent Basis Expansions
Declan McNamara (University of Michigan), Jeffrey Regier (University of Michigan)
Physics Related
🎯 What it does: Proposed a new variational inference method called Latent Basis Function Neural Posterior Estimation (LBF-NPE), which parameterizes the variational family through basis function expansion of latent variables to improve the flexibility of inference and the stability of optimization.
Neural Predictor-Corrector: Solving Homotopy Problems with Reinforcement Learning
Jiayao Mai (Hunan University), Peidong Liu (Westlake University)
OptimizationReinforcement LearningPoint CloudTabular
🎯 What it does: Propose a Neural Predictor-Corrector (NPC) framework that leverages reinforcement learning to automatically learn strategies for the predictor and corrector, unifying the solution of multiple classes of homotopy problems including robust optimization, global optimization, polynomial root-finding, and sampling.
Neural Sum-of-Squares: Certifying the Nonnegativity of Polynomials with Transformers
Nico Pelleriti (Zuse Institute Berlin), Sebastian Pokutta (Zuse Institute Berlin)
OptimizationComputational EfficiencyTransformer
🎯 What it does: Accelerate polynomial non-negativity determination by predicting compact bases via Transformer, combined with repair mechanisms and iterative SDP solving;
Neural Synchrony Between Socially Interacting Language Models
Zhining Zhang (Peking University), Heng Ji (University Of Illinois Urbana Champaign)
Representation LearningTransformerLarge Language ModelTextBenchmark
🎯 What it does: Studied the internal representation synchrony of large language models (LLMs) in social interactions, proposed and used the SyncR^2 metric to measure neural synchrony during LLM interactions;
Neural Theorem Proving for Verification Conditions: A Real-World Benchmark
Qiyuan Xu (Nanyang Technological University), Conrad Watt (Nanyang Technological University)
AI Code AssistantLarge Language ModelTextBenchmark
🎯 What it does: Proposes Neural Theorem Proving for Verification Conditions (NTP4VC) task and constructs the first multi-language (Isabelle, Lean, Rocq) real-world industrial VC proof benchmark;
Neural+Symbolic Approaches for Interpretable Actor-Critic Reinforcement Learning
Yue Yang (Monash University), Hao Wang (Monash University)
Explainability and InterpretabilityReinforcement Learning
🎯 What it does: Propose a Neural+Symbolic Actor-Critic framework (NSAC), using neural networks as critic for value estimation and rule sets (Additive Rule Ensemble) as actor for decision-making, achieving interpretable reinforcement learning.
NeuralOS: Towards Simulating Operating Systems via Neural Generative Models
Luke Rivard (University of Waterloo), Yuntian Deng (University of Waterloo)
GenerationData SynthesisConvolutional Neural NetworkRecurrent Neural NetworkDiffusion modelVideoSequential
🎯 What it does: Propose NeuralOS, an end-to-end model that generates operating system GUI frames through neural networks, capable of predicting screen images based on mouse and keyboard inputs;
Neuro-Symbolic Decoding of Neural Activity
Yanchen Wang (Columbia University), Jiajun Wu (Stanford University)
Explainability and InterpretabilityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: Proposes NEURONA, a neuro-symbolic framework that maps fMRI signals to interpretable conceptual representations and achieves precise decoding of multi-entity relationships in visual stimuli through symbolic execution.
Neuron-Aware Data Selection in Instruction Tuning for Large Language Models
Xin Chen (University of Macau), Derek F. Wong (Nanjing University)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Propose the NAIT framework, leveraging neuron activation features to efficiently screen instruction-tuning (IT) data, thereby enhancing the capabilities of large language models in specific domains.
Neuron-Level Analysis of Cultural Understanding in Large Language Models
Taisei Yamamoto (University of Tokyo), Hitomi Yanaka (University of Tokyo)
Explainability and InterpretabilityTransformerTextBenchmark
🎯 What it does: Identify and validate neurons affecting cultural understanding in large language models through gradient attribution and control dataset screening methods, and explore their distribution and functions.
Never Saddle for Reparameterized Steepest Descent as Mirror Flow
Tom Jacobs (CISPA Helmholtz Center for Information Security), Rebekka Burkholz (CISPA Helmholtz Center for Information Security)
OptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper investigates extending gradient flow to the Steepest Mirror Flow framework under deep diagonal reparameterization, and theoretically derives and experimentally verifies why Adam/AdamW outperforms SGD in fine-tuning.
New Hybrid Fine-Tuning Paradigm for LLMs: Algorithm Design and Convergence Analysis Framework
Shaocong Ma (University of Maryland), Heng Huang (University of Maryland)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a hybrid fine-tuning paradigm that applies zeroth-order optimization to the main parameters of large language models and first-order optimization to the PEFT module, achieving joint updates of both.
Newton Method Revisited: Global Convergence Rates up to $O(1/k^3)$ for Stepsize Schedules and Linesearch Procedures
Slavomir Hanzely, Martin Takáč (Mohammed bin Zayed University of Artificial Intelligence)
Optimization
🎯 What it does: This paper proposes a Newton method based on high-order regularization and provides a concise step size scheduling, achieving a global convergence rate of O(k⁻³) for convex functions with Hölder continuous second or third derivatives.
NewtonBench: Benchmarking Generalizable Scientific Law Discovery in LLM Agents
Tianshi Zheng (Hong Kong University of Science and Technology), Simon See (NVIDIA)
TransformerLarge Language ModelAgentic AIBenchmarkPhysics Related
🎯 What it does: Proposes NEWTONBENCH, an interactive benchmark for evaluating the capability of large language models (LLMs) in discovering scientific laws.
NewtonGen: Physics-consistent and Controllable Text-to-Video Generation via Neural Newtonian Dynamics
Yu Yuan (Purdue University), Stanley H. Chan (Purdue University)
GenerationVideoTextPhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposes the NewtonGen framework, achieving the generation of physically consistent and controllable dynamic images from text descriptions.
Next Visual Granularity Generation
Yikai Wang (Nanyang Technological University), Chen Change Loy (Nanyang Technological University)
GenerationFlow-based ModelAuto EncoderImage
🎯 What it does: Propose a framework called NVG for image generation through hierarchical visual granularity sequences, decomposing images into structural and content tags at different granularities and generating them progressively from coarse to fine;
NExT-OMNI: Towards Any-to-Any Omnimodal Foundation Models with Discrete Flow Matching
Run Luo (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences), Tat-Seng Chua (National University of Singapore)
GenerationRetrievalRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningFlow-based ModelAuto EncoderImageVideoTextMultimodalityAudio
🎯 What it does: Developed NExT-OMNI—a fully open-source omnimodal foundation model capable of cross-modal understanding, generation, and retrieval between any modalities;
Next-ToBE: Probabilistic Next Token-Bag Exploitation for Activating Anticipatory Capacity in LLMs
Yihe Liu (East China Normal University), Kai Zhang (Fudan University)
Large Language ModelTextBenchmark
🎯 What it does: Proposes Next-ToBE, which replaces the traditional one-hot target with a soft target distribution, using the probability distribution of LLMs at the current step to actively predict future k-step tokens, thereby activating and enhancing the anticipatory capability of LLMs and improving reasoning performance.
NextQuill: Causal Preference Modeling for Enhancing LLM Personalization
Xiaoyan Zhao (Chinese University of Hong Kong), Tat-Seng Chua (National University of Singapore)
Recommendation SystemExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the NextQuill framework, which personalizes and aligns large language models through causal preference modeling, identifying and aligning preference components driven by user history in model predictions and real responses.
NextStep-1: Toward Autoregressive Image Generation with Continuous Tokens at Scale
Chunrui Han (StepFun), Yibo Zhu (StepFun)
GenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningDiffusion modelFlow-based ModelAuto EncoderImageVideoTextMultimodalityBenchmark
🎯 What it does: Propose a fully autoregressive text-to-image generation model named NextStep-1, which concatenates text with continuous image tokens into a unified sequence and progressively predicts image patches using a lightweight flow matching head.
Neyman-Pearson Classification under Both Null and Alternative Distributions Shift
Mohammadreza Mousavi Kalan (University of Rennes), Samory Kpotufe (Columbia University)
ClassificationDomain AdaptationTabular
🎯 What it does: Propose an adaptive transfer learning method applicable to Neyman-Pearson classification when both class distribution shifts may occur; the method reduces the target Type-II error using source domain samples while ensuring the target Type-I error does not exceed a threshold, and automatically avoids negative transfer when the source domain contains no useful information.
NFT: Bridging Supervised Learning and Reinforcement Learning in Math Reasoning
Huayu Chen (Tsinghua), Haoxiang Wang (NVIDIA)
Reinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose Negative-aware Fine-Tuning (NFT), a self-reflective method based on supervised learning that enhances large models' performance in mathematical reasoning tasks by leveraging negative samples.
NGS-Marker: Robust Native Watermarking for 3D Gaussian Splatting
Hao Qin (Zhejiang University), Qiang Zhu (Zhejiang University)
TransformerGaussian SplattingPoint Cloud
🎯 What it does: Propose the NGS-Marker framework, achieving native watermark embedding and decoding in the local Gaussian set of 3D Gaussian Splatting assets, directly verifying copyright on the original geometry and resisting local infringement.
NI Sampling: Accelerating Discrete Diffusion Sampling by Token Order Optimization
Enshu Liu (Tsinghua University), Zinan Lin (Microsoft Research)
GenerationComputational EfficiencyDiffusion modelTextBenchmark
🎯 What it does: Propose and study a neural indicator sampling framework, NI Sampling, which uses a lightweight neural metric to determine which token should be decoded at each step in discrete diffusion language models, thus significantly accelerating sampling.
NIMO: a Nonlinear Interpretable MOdel
Shijian Xu (University of Basel), Volker Roth (University of Basel)
Explainability and InterpretabilityTabular
🎯 What it does: Propose a hybrid model NIMO that retains the interpretability of linear regression while achieving nonlinear correction for each feature through neural networks;
NLI : Non-uniform Linear Interpolation Approximation of Nonlinear Operations for Efficient LLMs Inference
Jiangyong Yu (Houmo AI), Dawei Yang (Houmo AI)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes a nonlinear operation approximation framework based on non-uniform linear interpolation (NLI) to replace high-precision floating-point operations in LLM inference, thereby improving computational efficiency.
No Caption, No Problem: Caption-Free Membership Inference via Model-Fitted Embeddings
Joonsung Jeon (Korea Advanced Institute of Science and Technology), Sung-eui Yoon
GenerationSafty and PrivacyAdversarial AttackVision Language ModelDiffusion modelImage
🎯 What it does: Developed MOFIT, a method for membership inference attacks on latent diffusion models in a caption-free environment;
No Labels, No Problem: Training Visual Reasoners with Multimodal Verifiers
Damiano Marsili (California Institute of Technology), Georgia Gkioxari (California Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Propose an unlabeled visual reasoning training framework VALOR, which enhances LLM reasoning and visual localization capabilities through a multimodal validator
No outlier channels but with outlier blocks
Shanwen Mao (Harbin Institute Of Technology), Jie Liu (State Key Laboratory Of Smart Farm Technologies And Systems)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose the NuBitQ framework and OCP plugin to achieve layer-wise adaptive non-uniform quantization, and suppress block-level anomaly errors through multi-level compensation.
No Pixel Left Behind: A Detail-Preserving Architecture for Robust High-Resolution AI-Generated Image Detection
Lianrui Mu (Zhejiang University), Jiaqi Hu (Zhejiang University)
Anomaly DetectionTransformerImageBenchmark
🎯 What it does: Propose HiDA-Net, a high-resolution AI-generated image detection framework that leverages full-resolution local patching and global view fusion;
No Prior, No Leakage: Revisiting Reconstruction Attacks in Trained Neural Networks
Yehonathan Refael, Itay Safran (Ben-Gurion University of the Negev)
Safty and PrivacyAdversarial AttackImage
🎯 What it does: Investigated an attack method that exploits implicit bias from trained neural networks to reconstruct training samples, and proved that such attacks are inherently unreliable without prior knowledge.
No Prompt Left Behind: Exploiting Zero-Variance Prompts in LLM Reinforcement Learning via Entropy-Guided Advantage Shaping
Thanh-Long V. Le (Korea Advanced Institute of Science and Technology), Eunho Yang (Korea Advanced Institute of Science and Technology)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose a reinforcement learning method RL-ZVP that utilizes zero-variance prompts (all samples have identical rewards) to enhance the reasoning capabilities of large language models.
No, of Course I Can! Deeper Fine-Tuning Attacks That Bypass Token-Level Safety Mechanisms
Joshua Kazdan (Stanford University), Krishnamurthy Dj Dvijotham
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Propose a new deep fine-tuning attack method called NOICE, which enables models to first refuse and then answer harmful requests, bypassing existing safety defenses based on the first few words.
Noise Stability of Transformer Models
Themistoklis Haris (Boston University), Yuichi Yoshida (National Institute of Informatics)
Explainability and InterpretabilityTransformerLarge Language ModelTextSequential
🎯 What it does: Proposed and studied noise stability as a measure of the simplicity bias in Transformer models, and designed a noise stability regularization method to accelerate the 'grokking' process of the model.
Noise Tolerance of Distributionally Robust Learning
Ramzi Dakhmouche (EPFL), Hossein Gorji (Empa)
Domain AdaptationOptimizationComputational EfficiencyTime SeriesPhysics Related
🎯 What it does: Propose a regression method based on Wasserstein batch matching, which trains directly on noisy data to enhance robustness to global noise.
Noise-Adaptive Diffusion Sampling for Inverse Problems Without Task-Specific Tuning
Yingzhi Xia (Institute of High Performance Computing, Agency for Science, Technology and Research), ZAIWANG GU
RestorationDiffusion modelImage
🎯 What it does: This paper proposes the HMC sampling method N-HMC and its adaptive variant NA-N-HMC, which perform backward diffusion in the noise space to solve inverse problems without requiring task-specific hyperparameters.
Noise-Aware Generalization: Robustness to In-Domain Noise and Out-of-Domain Generalization
Siqi Wang (Boston University), Bryan A. Plummer (Boston University)
Domain AdaptationImage
🎯 What it does: Studied the generalization problem in the presence of label noise and multi-domain data (Noise-Aware Generalization, NAG), and proposed an algorithm called DL4ND that uses cross-domain comparison for noise detection and label correction.
NoisePrints: Distortion-Free Watermarks for Authorship in Private Diffusion Models
Nir Goren (Tel Aviv University), Or Patashnik (University Of Michigan)
GenerationSafty and PrivacyDiffusion modelAuto EncoderImageVideo
🎯 What it does: Proposes a lightweight watermarking scheme called NoisePrints, based on the initialization noise of diffusion models, which associates a seed with the generated content to achieve author authentication.
Noisy but Valid: Robust Statistical Evaluation of LLMs with Imperfect Judges
Chen Feng (Queen's University Belfast), Miguel R. D. Rodrigues (University College London)
Large Language ModelText
🎯 What it does: Proposed the "Noise-but-Valid" hypothesis testing framework for certifying the safety of large language models in the presence of imperfect judges (LLM-as-Judge);
Noisy-Pair Robust Representation Alignment for Positive-Unlabeled Learning
Hengwei Zhao (Texas A&M University), Wenzhe Jiao (Texas A&M University)
Representation LearningImage
🎯 What it does: Proposes a positive-unlabeled (PU) learning framework named NcPU, combining a noise-robust non-contrastive learning loss (NoiSNCL) and phantom label disambiguation (PLD), to learn more discriminative representations without requiring auxiliary negative samples or prior class proportion assumptions;
Non-Asymptotic Analysis of (Sticky) Track-and-Stop
Riccardo Poiani (Bocconi University), Andrea Celli (Bocconi University)
OptimizationReinforcement Learning
🎯 What it does: Proposed the first non-asymptotic (finite confidence) upper bounds on sample complexity for Track-and-Stop (TAS) and Sticky Track-and-Stop (S-TAS) in pure exploration bandit tasks.
Non-Asymptotic Analysis of Efficiency in Conformalized Regression
Yunzhen Yao (EPFL), Michael Gastpar (EPFL)
🎯 What it does: Studies quantile regression in split conformal prediction (CQR and CMR), provides non-asymptotic upper bounds on their length deviation, and validates the theory through experiments.
Non-Clashing Teaching in Graphs: Algorithms, Complexity, and Bounds
Sujoy Bhore (Indian Institute of Technology), Fionn Mc Inerney (Telefónica Scientific Research)
Computational EfficiencyGraph
🎯 What it does: Theoretical research on a non-conflicting teaching model for closed neighborhoods in graphs, proposing new algorithms, complexity lower bounds, and combinatorial upper bounds.
Non-Collaborative User Simulators for Tool Agents
Jeonghoon Shim (Seoul National University), Yohan Jo (Seoul National University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: This paper develops a user simulation framework capable of mimicking four types of non-cooperative user behaviors (unavailable service, off-topic, impatience, incomplete expression) in multi-turn dialogues, and evaluates the robustness of tool agents on MultiWOZ and τ-bench.
Non-Convex Federated Optimization under Cost-Aware Client Selection
Xiaowen Jiang (Saarland University), Sebastian U Stich
OptimizationFederated LearningComputational EfficiencyTabular
🎯 What it does: This paper addresses non-convex federated optimization problems by constructing a model that associates different client selection strategies with communication/local computation costs. Based on this, a new algorithm called I-CGM-RG (primarily implemented as RG-SAGA and RG-SVRG) is proposed, combining the incomplete composite gradient method (I-CGM) and recursive gradient estimator (RG), achieving optimality in communication and local complexity.
Nonparametric Contextual Online Bilateral Trade
Emanuele Coccia (Bocconi University), Andrea Celli (Bocconi University)
Optimization
🎯 What it does: Proposes a non-parametric contextual online bilateral trade framework where the learner determines transaction prices under single-bit feedback and strong budget balance constraints, observing only the context.
Nonparametric Teaching of Attention Learners
Chen Zhang (University of Hong Kong), Ngai Wong (University of Hong Kong)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: Propose AtteNT, a paradigm that combines non-parametric teaching with gradient training of attention learners, utilizing greedy example selection to accelerate the training process.
Not All Bits Are Equal: Scale-Dependent Memory Optimization Strategies for Reasoning Models
Junhyuck Kim (Krafton), Dimitris Papailiopoulos (Microsoft Research)
CompressionOptimizationComputational EfficiencyKnowledge DistillationTransformerTextBenchmark
🎯 What it does: The study investigates how to balance memory usage and accuracy across dimensions such as weight precision, KV cache compression, generation length, and parallel sampling during inference under a fixed GPU memory budget, systematically evaluating the effectiveness of various compression and expansion strategies in inference tasks.
Not All Documents Are What You Need for Extracting Instruction Tuning Data
Chi Zhang (Beijing Institute of Technology), Lei Cao (University of Arizona)
Computational EfficiencyData-Centric LearningLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: Propose a framework called EQUAL that efficiently retrieves and extracts high-quality question-answer (QA) pairs from large-scale network corpora using an iterative multi-armed bandit (MAB) method to enhance the effectiveness of instruction tuning.
Not All Models Suit Expert Offloading: On Local Routing Consistency of Mixture-of-Expert Models
Jingcong Liang (Fudan University), zhongyu wei
Computational EfficiencyLarge Language ModelMixture of ExpertsText
🎯 What it does: This paper investigates the local routing consistency of expert routing in Mixture-of-Experts (MoE) large language models during inference, and proposes two quantitative metrics: Segment Routing Best Performance (SRP) and Segment Cache Best Hit Rate (SCH).
Not Search, But Scan: Benchmarking MLLMs on Scan-Oriented Academic Paper Reasoning
Rongjin Li (Beijing University of Posts and Telecommunications), Haihong E (Beijing University of Posts and Telecommunications)
TransformerLarge Language ModelTextMultimodalityBenchmarkRetrieval-Augmented Generation
🎯 What it does: Proposes ScholScan, a benchmark for academic paper reasoning oriented towards scanning, requiring models to fully scan papers without target prompts and identify errors.
Not-a-Bandit: Provably No-Regret Drafter Selection in Speculative Decoding for LLMs
Hongyi Liu (Rice University), Yu-Xiang Wang (University of California San Diego)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextSequential
🎯 What it does: Study online draft model selection, propose a regret-free full-information learning algorithm that can evaluate all drafters and real-time select the optimal drafter without additionally calling the target model
NOVA3R: Non-pixel-aligned Visual Transformer for Amodal 3D Reconstruction
Weirong Chen (Technical University Of Munich), Daniel Cremers (University Of Oxford)
GenerationTransformerDiffusion modelFlow-based ModelAuto EncoderImagePoint Cloud
🎯 What it does: Directly predict complete, non-pixel-aligned 3D point clouds from images with no pose information, supporting single-view or multi-view inputs.
NRGPT: An Energy-based Alternative for GPT
Nima Dehmamy (IBM Research), Dmitry Krotov (IBM Research)
GenerationTransformerLarge Language ModelText
🎯 What it does: Proposes NRGPT, unifying GPT's autoregressive language modeling with energy-based models, performing inference through energy gradient descent.
Nudging the Boundaries of LLM Reasoning
Justin Chen (Salesforce AI Research), Chien-Sheng Wu (Salesforce AI Research)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Designed an algorithm called NuRL that self-generates hints during reinforcement learning (RL) training to 'push the boundaries' of large language model (LLM) reasoning, helping the model learn difficult problems that it could not successfully explore otherwise.
Null-Space Filtering for Data-Free Continual Model Merging: Preserving Stability, Promoting Plasticity
Zihuan Qiu (University of Electronic Science and Technology of China), Hongliang Li (University of Electronic Science and Technology of China)
Representation LearningTransformerImageTextMultimodality
🎯 What it does: This paper proposes a data-agnostic continuous model fusion method called NUFILT, which can sequentially merge multiple fine-tuned models into a single high-performance model without accessing task data.
Numerion: A Multi-Hypercomplex Model for Time Series Forecasting
Hanzhong Cao (Peking University), Ying Tan (Peking University)
Time Series
🎯 What it does: Proposed a multi-hypercomplex space time series prediction model called Numerion, which utilizes Real-Hypercomplex-Real multi-layer perceptron (RHR-MLP) to naturally decompose, model, and fuse multi-frequency features in different dimensional hypercomplex spaces.
NurValues: Real-World Nursing Values Evaluation for Large Language Models in Clinical Context
Ben Yao (Hong Kong Polytechnic University), Jing Qin (Hong Kong Polytechnic University)
Large Language ModelPrompt EngineeringTextBiomedical DataBenchmarkChain-of-Thought
🎯 What it does: Constructed the NurValues benchmark using nursing behavior cases collected from five months of longitudinal interviews in real hospitals, generating two levels of difficulty datasets to evaluate the alignment capability of LLMs with the five major nursing values.
Nüwa: Mending the Spatial Integrity Torn by VLM Token Pruning
Yihong Huang (Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)), Qi Tian (Hong Kong Polytechnic University)
OptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Proposed the Nuwa two-stage visual token pruning framework, balancing the performance of vision-language models with efficient inference;
Obfuscated Activations Bypass LLM Latent-Space Defenses
Luke Bailey (Stanford University), Scott Emmons (UC Berkeley)
Adversarial AttackTransformerPrompt EngineeringAuto EncoderText
🎯 What it does: Investigated the robustness of LLM latent space monitors, proposed and evaluated obfuscated activations attacks capable of bypassing monitors;
Object Fidelity Diffusion for Remote Sensing Image Generation
Ziqi Ye (Fudan University), Haipeng Wang (Fudan University)
GenerationData SynthesisReinforcement LearningDiffusion modelAuto EncoderImage
🎯 What it does: Designed and implemented an online distillation and shape prior-based remote sensing image generation model (OF-Diff), achieving high-fidelity, controllable layout-to-image generation without relying on real image references, and further enhancing diversity and semantic consistency through DDPO.
Object-Centric Refinement for Enhanced Zero-Shot Segmentation
Srinivasa Rao Nandam (University of Surrey), Muhammad Awais (University of Surrey)
SegmentationTransformerPrompt EngineeringVision Language ModelContrastive LearningImageBenchmark
🎯 What it does: Proposed OC-ZSS, which improves CLIP's patch representations by utilizing self-supervised guided object prompts and a two-stage object refinement attention module to achieve zero-shot semantic segmentation.
Object-Centric World Models from Few-Shot Annotations for Sample-Efficient Reinforcement Learning
Weipu Zhang (National Key Lab of Autonomous Intelligent Unmanned Systems, Beijing Institute of Technology), Gang Wang (National Key Lab of Autonomous Intelligent Unmanned Systems, Beijing Institute of Technology)
SegmentationRecurrent Neural NetworkTransformerReinforcement LearningAuto EncoderWorld ModelVideo
🎯 What it does: Proposed and implemented OC-STORM, an object-centric model-based reinforcement learning framework that integrates a pre-trained video segmentation network with a small amount of annotated data to generate object features, enabling training of world models and improving policy learning efficiency in pixel-level environments through object information.
OBS-Diff: Accurate Pruning For Diffusion Models in One-Shot
Junhan Zhu (Westlake University), Huan Wang (Westlake University)
GenerationComputational EfficiencyDiffusion modelImageText
🎯 What it does: Propose a one-time, no-training pruning framework called OBS-Diff for large-scale text-to-image diffusion models, which can significantly reduce model computational and memory overhead while maintaining visual quality.
Obscure but Effective: Classical Chinese Jailbreak Prompt Optimization via Bio-Inspired Search
Xun Huang (Nanyang Technological University), Xiaojun Jia (Nanyang Technological University)
OptimizationAdversarial AttackLarge Language ModelPrompt EngineeringText
🎯 What it does: Develop a black-box jailbreak framework named CC-BOS that leverages classical Chinese context, utilizing a multi-dimensional strategy space and fruit fly optimization algorithm to automatically generate adversarial prompts in classical Chinese.
OccDriver: Future Occupancy Guided Dual-branch Trajectory Planner in Autonomous Driving
Zhao Huang (Tianjin University), Di Lin (SenseAuto)
Autonomous DrivingTransformerVideo
🎯 What it does: Proposes a future occupied space guided dual-branch trajectory planning framework (OccDriver), achieving interactive perception trajectory planning through coarse-to-fine collaborative prediction of vectorization and rasterization.
Occupancy Reward Shaping: Improving Credit Assignment for Offline Goal-Conditioned Reinforcement Learning
Aravind Venugopal, Jeff Schneider (Princeton University)
Reinforcement LearningFlow-based ModelTime SeriesPhysics Related
🎯 What it does: Through offline goal-conditioned reinforcement learning, the occupancy measure learned from the world model is converted into a global reward function, thereby addressing the credit assignment problem under sparse rewards.
OCR-Reasoning Benchmark: Unveiling the True Capabilities of MLLMs in Complex Text-Rich Image Reasoning
Mingxin Huang (South China University Of Technology), Lianwen Jin (Huawei Technologies Co Ltd)
Large Language ModelPrompt EngineeringVision Language ModelImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: This paper proposes the OCR-Reasoning benchmark to systematically evaluate the ability of multimodal large language models in text-rich image reasoning tasks.
Octax: Accelerated CHIP-8 Arcade Environments for Reinforcement Learning in JAX
Waris Radji (Univ. Lille), Hector Piteau (Independent Researcher)
Computational EfficiencyLarge Language ModelReinforcement LearningText
🎯 What it does: Developed OCTAX, a CHIP-8 simulation platform based on JAX, providing a GPU-accelerated classic arcade game environment for reinforcement learning experiments.
OD$^3$: Optimization-free Dataset Distillation for Object Detection
Salwa K. Al Khatib (Mohamed bin Zayed University of Artificial Intelligence), Zhiqiang Shen (Mohamed bin Zayed University of Artificial Intelligence)
Object DetectionData SynthesisKnowledge DistillationImage
🎯 What it does: Proposed the OD 3 framework, which performs data distillation on object detection datasets in a non-optimized manner, generating extremely small but efficient synthetic datasets.
ODE-GS: Latent ODEs for Dynamic Scene Extrapolation with 3D Gaussian Splatting
Daniel Wang (Yale University), Ganesh Sundaramoorthi (RTX)
GenerationTransformerGaussian SplattingVideoPoint CloudOrdinary Differential Equation
🎯 What it does: Propose a model called ODE-GS that combines 3D Gaussian projection with transformer-driven latent ordinary differential equations (Latent ODE) for time extrapolation prediction in dynamic 3D scenes
ODEBrain: Continuous-Time EEG Graph for Modeling Dynamic Brain Networks
Haohui Jia (Hokkaido University), Takashi Matsubara (Hokkaido University)
ClassificationGraph Neural NetworkBiomedical DataOrdinary Differential Equation
🎯 What it does: Proposed a continuous spatiotemporal graph model called ODEBRAIN based on neural ordinary differential equations (Neural ODE), for dynamic brain network modeling of EEG signals and achieving epilepsy seizure detection and abnormal EEG classification.
ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment
Hongjue Zhao (University of Illinois Urbana-Champaign), Huajie Shao (William & Mary)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextOrdinary Differential Equation
🎯 What it does: Proposed a unified activation scheduling framework based on ordinary differential equations (ODEs) and implemented a novel scheduling method called ODESTEER to enhance large language model (LLM) alignment effectiveness.
ODI-Bench: Can MLLMs Understand Immersive Omnidirectional Environments?
Liu Yang (Shanghai Jiao Tong University), Guangtao Zhai (Shanghai Jiao Tong University)
RecognitionVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: Proposed the ODI-Bench benchmark for panoramic image (ODI) understanding, containing 2000 high-quality panoramic images and over 4000 QA pairs, covering 10 fine-grained tasks, and benchmarked 20 multimodal large language models (MLLMs) under both closed-ended and open-ended evaluations;
Off-Policy Evaluation for Ranking Policies under Deterministic Logging Policies
Koichi Tanaka (Keio University), Yuta Saito (Hanuku-kaso, Co., Ltd.)
Recommendation SystemTabular
🎯 What it does: Proposed a new estimator for offline evaluation of ranking policies under a fully deterministic logging policy—Click-based Inverse Probability Scoring (CIPS) and its double robust extension (CDR);
Off-Policy Safe Reinforcement Learning with Cost-Constrained Optimistic Exploration
Guopeng Li (Delft University of Technology), Julian F. P. Kooij (Delft University of Technology)
Autonomous DrivingOptimizationReinforcement Learning
🎯 What it does: Propose an offline safe reinforcement learning algorithm COX-Q, combining cost-constrained optimistic exploration and distributed value learning to maintain safety during data collection and deployment while improving sample efficiency.
Off-Trajectory Reasoning: Can LLMs Collaborate on Reasoning Trajectories?
Aochong Oliver Li (Cornell University), Tanya Goyal (Cornell University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmark
🎯 What it does: This paper designs two tests, recovery and guidance, to evaluate the robustness of LLMs under off-trajectory reasoning, revealing the fragility of standard solo reasoning models when interacting with others' reasoning trajectories.
Offline Preference-Based Value Optimization
Hyungkyu Kang (Seoul National University), Min-hwan Oh (Seoul National University)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: Propose an offline preference-driven reinforcement learning algorithm, Preference-based Value Optimization (PVO), which directly optimizes the value function through a value alignment loss, making it consistent with human preference feedback.
Offline Reinforcement Learning with Adaptive Feature Fusion
Tieru Wang (Beijing University of Posts and Telecommunications), Guoshun Nan (Beijing University of Posts and Telecommunications)
TransformerReinforcement LearningImageSequential
🎯 What it does: Propose Q-Augmented Dual-Feature Fusion Decision Transformer (QDFFDT), integrating global sequence features and local Markovian features in offline reinforcement learning, and introducing Q-learning for value-guided decision-making.
OffTopicEval: When Large Language Models Enter the Wrong Chat, Almost Always!
Jingdi Lei (Nanyang Technological University), Soujanya Poria (Nanyang Technological University)
Safty and PrivacyLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: Evaluate the operational safety of large language models in specific agent tasks, propose a multilingual, multi-domain evaluation suite called OFFTOPICEVAL, and conduct large-scale benchmarking on 20 open-source/closed-source models.
OFMU: OPTIMIZATION-DRIVEN FRAMEWORK FOR MACHINE UNLEARNING
Sadia Asif (Rensselaer Polytechnic Institute), Mohammad Mohammadi Amiri (Rensselaer Polytechnic Institute)
OptimizationSafty and PrivacyConvolutional Neural NetworkTransformerImageText
🎯 What it does: This paper proposes the OFMU (Optimization-Driven Framework for Machine Unlearning) framework, which enables a trained model to proactively 'forget' specified data without retraining, while maintaining its performance on the remaining data.
Omni-Captioner: Data Pipeline, Models, and Benchmark for Omni Detailed Perception
Ziyang Ma (Shanghai Jiao Tong University), Xie Chen (Shanghai Jiao Tong University)
GenerationData SynthesisLarge Language ModelSupervised Fine-TuningAgentic AIVision Language ModelVideoMultimodalityBenchmarkAudio
🎯 What it does: Designed and implemented the Omni-Detective data generation pipeline, two fine-grained description models (Audio-Captioner and Omni-Captioner), and proposed the Omni-Cloze benchmark for unified evaluation of multimodal fine-grained perception capabilities.
Omni-iEEG: A Large-Scale, Comprehensive iEEG Dataset and Benchmark for Epilepsy Research
Chenda Duan (UCLA Samueli School of Engineering), vwani Roychowdhury
Convolutional Neural NetworkTransformerSupervised Fine-TuningBiomedical DataBenchmark
🎯 What it does: Constructed and released the Omni-iEEG large-scale, standardized pre-surgical iEEG dataset, containing 302 patients, 178 hours of high-resolution recordings, and providing 36K expert-verified HFO annotations along with clinical metadata.
Omni-IML: Towards Unified Interpretable Image Manipulation Localization
Chenfan Qu (South China University of Technology), Lianwen Jin (South China University of Technology)
Anomaly DetectionExplainability and InterpretabilityLarge Language ModelImageTextMultimodalityChain-of-Thought
🎯 What it does: Developed a general-purpose image manipulation localization model called Omni-IML and constructed a large-scale, structured interpretable dataset named Omni-273k.
Omni-Reward: Towards Generalist Omni-Modal Reward Modeling with Free-Form Preferences
Zhuoran Jin (University of Chinese Academy of Sciences), Jun Zhao (University of Chinese Academy of Sciences)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningImageVideoTextMultimodalityBenchmarkAudio
🎯 What it does: Propose the Omni-Reward framework, constructing Omni-RewardBench, Omni-RewardData, and Omni-RewardModel to achieve multi-modal free-form preference reward modeling.
Omni-View: Unlocking How Generation Facilitates Understanding in Unified 3D Model based on Multiview images
JiaKui Hu (Institute of Medical Technology Peking University), Yanye Lu (Institute of Medical Technology Peking University)
GenerationDepth EstimationTransformerDiffusion modelFlow-based ModelImageTextMultimodality
🎯 What it does: Propose Omni-View, a unified model integrating 3D scene understanding and generation, enhancing multi-view image understanding through texture and geometry generation modules.
Omni-Weather: A Unified Multimodal Model for Weather Radar Understanding and Generation
Zhiwang Zhou (Tongji University), LEI BAI
Image TranslationGenerationExplainability and InterpretabilityTransformerSupervised Fine-TuningVision Language ModelAuto EncoderImageVideoTextMultimodalityChain-of-Thought
🎯 What it does: Propose Omni-Weather, a unified multimodal foundation model that simultaneously performs weather forecasting (e.g., radar nowcasting, satellite radar inversion) and weather understanding (e.g., radar report generation, question answering)
OmniActor: A Generalist GUI and Embodied Agent for 2D&3D Worlds
Longrong Yang, Xi Li (Zhejiang University)
Robotic IntelligenceLarge Language ModelReinforcement LearningMixture of ExpertsVision-Language-Action ModelMultimodality
🎯 What it does: Developed OmniActor, a general-purpose multimodal agent capable of performing tasks simultaneously in 2D GUI and 3D physical worlds