NeurIPS 2025 Papers — Page 14
Conference on Neural Information Processing Systems · 5275 papers
Dynamic Regret Reduces to Kernelized Static Regret
Andrew Jacobsen (University of Milan), Nicolò Cesa-Bianchi (University of Milan)
Optimization
🎯 What it does: This paper proposes transforming the dynamic regret problem into a static optimization problem in the Reproducing Kernel Hilbert Space (RKHS), and achieves optimal path length dependence by designing appropriate kernel functions.
Dynamic Semantic-Aware Correlation Modeling for UAV Tracking
Xinyu Zhou (Fudan University), Wenqiang Zhang (Fudan University)
Object TrackingTransformerVideo
🎯 What it does: A dynamic semantic-aware association modeling framework for drone tracking, DSATrack, is proposed to enhance the semantic matching and localization accuracy between the template and the search area.
Dynamic Shadow Unveils Invisible Semantics for Video Outpainting
Ruilin Li (Shanghai University), Jiayan Qiu (University of Leicester)
RestorationGenerationData SynthesisTransformerDiffusion modelOptical FlowVideo
🎯 What it does: A video inpainting framework that utilizes dynamic shadow information to achieve instance-aware generation, capable of producing coherent instances and lighting in occluded or disappeared areas.
Dynamic Siamese Expansion Framework for Improving Robustness in Online Continual Learning
Fei Ye (University of Electronic Science and Technology of China), shijie zhou
ClassificationAdversarial AttackTransformerContrastive LearningImage
🎯 What it does: This paper proposes the Dynamic Siamese Expansion Framework (DSEF), which dynamically generates lightweight experts through a static + dynamic Siamese backbone network in online continual learning, while maintaining the robustness of historical knowledge when learning new tasks.
Dynamic Test-Time Compute Scaling in Control Policy: Difficulty-Aware Stochastic Interpolant Policy
Inkook Chun (New York University), Eric Vanden-Eijnden (Harvard University)
Computational EfficiencyRobotic IntelligenceConvolutional Neural NetworkReinforcement LearningVision Language ModelMultimodalityStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes the Difficulty-Aware Stochastic Interpolant Policy (DA-SIP), which dynamically adjusts the integration step size, solver, and ODE/SDE mode based on real-time difficulty in robot control, achieving adaptive scaling during testing.
Dynamic View Synthesis as an Inverse Problem
Hidir Yesiltepe (Virginia Tech), Pinar Yanardag (Virginia Tech)
GenerationData SynthesisDiffusion modelVideo
🎯 What it does: A training-free dynamic view synthesis method is proposed, which generates new perspectives from monocular videos by simply changing the noise initialization of a pre-trained video diffusion model.
Dynamical Decoupling of Generalization and Overfitting in Large Two-Layer Networks
Andrea Montanari (Stanford University), Pierfrancesco Urbani (Université Paris-Saclay)
TabularPhysics Related
🎯 What it does: This study investigates the learning dynamics of large-scale two-layer neural networks, utilizing dynamic mean field theory to analyze the dynamic characteristics of training algorithms.
Dynamical Low-Rank Compression of Neural Networks with Robustness under Adversarial Attacks
Steffen Schotthöfer (Oak Ridge National Laboratory), Stefan Schnake (Oak Ridge National Laboratory)
CompressionOptimizationAdversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A robust low-rank compression method based on dynamic low-rank training and spectral regularization is proposed, which can achieve up to 94% parameter compression while maintaining or improving adversarial robustness.
Dynamical modeling of nonlinear latent factors in multiscale neural activity with real-time inference
Eray Erturk (University of Southern California), Maryam M. Shanechi (University of Southern California)
GenerationData SynthesisRepresentation LearningRecurrent Neural NetworkMultimodalityTime SeriesStochastic Differential Equation
🎯 What it does: A framework (MRINE) is proposed and implemented for nonlinear fusion and dynamic modeling of multimodal neural signals with different time scales, distributions, and missing samples in real-time environments.
Dynamical Properties of Tokens in Self-Attention and Effects of Positional Encoding
Duy-Tung Pham (FPT Software AI Center), Thieu Vo
TransformerImageTextOrdinary Differential Equation
🎯 What it does: This study investigates the continuous time dynamics of self-attention in pre-trained Transformers, analyzing the convergence and divergence behaviors of tokens, exploring the effects of absolute and rotational position encodings, and proposing lightweight adjustments to mitigate performance degradation caused by convergence.
DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation
Jiashuo Sun (University of Illinois Urbana-Champaign), Jiawei Han (University of Illinois Urbana-Champaign)
RetrievalOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Proposes the DynamicRAG framework, which uses RL to dynamically adjust the order and quantity of retrieved documents, allowing the generator and reranker to be trained simultaneously.
Dynamics of Spontaneous Topic Changes in Next Token Prediction with Self-Attention
Mumin Jia (York University), Jairo Diaz-Rodriguez (York University)
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This study investigates the spontaneous topic change mechanism of self-attention models in next-word prediction and validates it through single-layer self-attention theory and modern LLM experiments.
Dynamics-Aligned Latent Imagination in Contextual World Models for Zero-Shot Generalization
Frank Röder (Institute for Data Science Foundations), Pradeep Kr. Banerjee (Institute for Data Science Foundations)
Robotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningWorld ModelSequential
🎯 What it does: In the DreamerV3 framework, a self-supervised context encoder is added, which can infer implicit environmental context from the agent-environment interaction history, thus achieving zero-shot generalization.
DynamicVerse: A Physically-Aware Multimodal Framework for 4D World Modeling
Kairun Wen (Xiamen University), Zhiwen Fan (University of Texas at Austin)
SegmentationPose EstimationDepth EstimationOptimizationLarge Language ModelVision Language ModelSimultaneous Localization and MappingVideoTextMultimodality
🎯 What it does: A large-scale 4D dynamic scene dataset named DynamicVerse has been constructed, and an automated data collection and processing pipeline called DynamicGen has been proposed to generate physically scaled geometry, motion, instance segmentation, and multimodal textual descriptions from internet videos.
DynaNav: Dynamic Feature and Layer Selection for Efficient Visual Navigation
Jiahui Wang (National University of Singapore), Changhao Chen (Hong Kong University of Science and Technology)
Autonomous DrivingComputational EfficiencyTransformerReinforcement LearningImage
🎯 What it does: This paper proposes DynaNav, a visual navigation framework based on dynamic feature selection and early exit, which can adaptively activate Transformer layers and features according to scene complexity, significantly reducing computational load while maintaining performance.
DynaPhArM: Adaptive and Physics-Constrained Modeling for Target-Drug Complexes with Drug-Specific Adaptations
Diya Zhang (Central South University), Fei Guo (Central South University)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkTransformerDiffusion modelBiomedical Data
🎯 What it does: We propose DynaPhArM, a framework based on SE(3)-equivariant Transformers and diffusion models for adaptively simulating atomic-level interactions and conformational changes of target-drug complexes.
DynaPipe: Dynamic Layer Redistribution for Efficient Serving of LLMs with Pipeline Parallelism
HongXin Xu, Xianwei Zhang (Sun Yat-Sen University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A technique for dynamically reallocating pipeline layers during LLM inference is proposed to address the issue of pipeline idleness caused by tail sampling;
DynaRend: Learning 3D Dynamics via Masked Future Rendering for Robotic Manipulation
Jingyi Tian (Xi'an Jiaotong University), Gang Hua (Amazon)
Robotic IntelligenceTransformerReinforcement LearningImagePoint Cloud
🎯 What it does: This paper proposes the DynaRend framework, which utilizes multi-view RGB-D data for self-supervised pre-training of 3D semantics, geometry, and dynamics through a differential renderable triplane representation, and transfers the learned representations to keyframe-based control tasks such as robotic grasping and stacking.
E-BATS: Efficient Backpropagation-Free Test-Time Adaptation for Speech Foundation Models
Jiaheng Dong (University of Melbourne), Ting Dang (University of Melbourne)
Domain AdaptationOptimizationComputational EfficiencyConvolutional Neural NetworkPrompt EngineeringAudio
🎯 What it does: This paper proposes E-BATS, a gradient-free adaptive method for speech foundation models that can adapt in real-time to changes in different acoustic domains without using source data or labels.
E-MoFlow: Learning Egomotion and Optical Flow from Event Data via Implicit Regularization
Wenpu Li (Westlake University), Peidong Liu (Westlake University)
Pose EstimationAutonomous DrivingOptical FlowVideoOrdinary Differential Equation
🎯 What it does: An unsupervised framework E-MoFlow is proposed, which jointly learns the six degrees of freedom inertial navigation of event cameras and optical flow.
E2E-VGuard: Adversarial Prevention for Production LLM-based End-To-End Speech Synthesis
Zhisheng Zhang (Tsinghua University), Jie Hao
GenerationAdversarial AttackTransformerLarge Language ModelAudio
🎯 What it does: Proposes the E2E-VGuard defense framework, which utilizes adversarial perturbations to simultaneously disrupt voice characteristics and pronunciation, preventing voice cloning attacks based on LLM and end-to-end (E2E) scenarios.
E2Former: An Efficient and Equivariant Transformer with Linear-Scaling Tensor Products
Yunyang Li (Yale University), Jia Zhang (Ubiquant)
Computational EfficiencyDrug DiscoveryProtein Structure PredictionTransformerGraphBiomedical Data
🎯 What it does: The E2Former architecture is proposed, utilizing Wigner 6j convolutions to transfer edge-level tensor multiplication to node-level, achieving an efficient SO(3) equivariant Transformer for molecular modeling.
EA3D: Online Open-World 3D Object Extraction from Streaming Videos
Xiaoyu Zhou (Peking University), Ming-Hsuan Yang (Google DeepMind)
Object DetectionSegmentationVision Language ModelGaussian SplattingSimultaneous Localization and MappingVideo
🎯 What it does: An online, open-world 3D object extraction framework EA3D is proposed, capable of real-time 3D reconstruction and semantic understanding without prior geometric or pose information.
Each Complexity Deserves a Pruning Policy
Hanshi Wang (Chinese Academy of Sciences), Zhipeng Zhang
Autonomous DrivingComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: This paper proposes an adaptive pruning framework for untrained visual-language models, AutoPrune, which generates personalized token retention strategies for each input while adhering to a fixed computational budget.
EAG3R: Event-Augmented 3D Geometry Estimation for Dynamic and Extreme-Lighting Scenes
Xiaoshan Wu (University of Hong Kong), XIAOJUAN QI
Object TrackingPose EstimationDepth EstimationTransformerPoint Cloud
🎯 What it does: A 3D geometric estimation framework EAG3R based on event camera enhancement is proposed, capable of robust geometric reconstruction and camera pose tracking in dynamically dark scenes.
Eagle 2.5: Boosting Long-Context Post-Training for Frontier Vision-Language Models
Guo Chen (Nanjing University), Guilin Liu (NVIDIA)
RecognitionGenerationRetrievalTransformerLarge Language ModelVision Language ModelImageVideoTextMultimodality
🎯 What it does: Developed and trained the long-context visual language model Eagle 2.5, combining information-prioritized sampling and progressive post-training, and constructed the dual-layer annotated Eagle-Video-110K dataset to enhance the understanding of long videos and high-resolution images.
EAGLE-3: Scaling up Inference Acceleration of Large Language Models via Training-Time Test
Yuhui Li (Peking University), Hongyang Zhang (University of Waterloo)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes EAGLE-3, a lossless acceleration method that utilizes training-time tests during the inference phase and integrates multi-layer features;
EAP-GP: Mitigating Saturation Effect in Gradient-based Automated Circuit Identification
Lin Zhang (King Abdullah University of Science and Technology), Di Wang (King Abdullah University of Science and Technology)
TransformerLarge Language ModelText
🎯 What it does: A new gradient-based automatic circuit recognition method EAP-GP is proposed to address the saturation effect problem of existing methods.
EAReranker: Efficient Embedding Adequacy Assessment for Retrieval Augmented Generation
Dongyang Zeng (Guangzhou University), Binxing Fang (Guangzhou University)
GenerationRetrievalExplainability and InterpretabilityComputational EfficiencyTransformerTextRetrieval-Augmented Generation
🎯 What it does: Proposes EAReranker, a RAG pre-ranking model that evaluates the sufficiency of retrieval results using only embedding vectors.
EasySpec: Layer-Parallel Speculative Decoding for Efficient Multi-GPU Utilization
Yize Wu (Intelligent Software Research Center, Institute of Software, Chinese Academy of Sciences), Yanjun Wu (Intelligent Software Research Center, Institute of Software, Chinese Academy of Sciences)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper presents EasySpec, a layer-parallel speculation strategy for the Drafting phase of Speculative Decoding in multi-GPU environments, significantly improving inference throughput.
EchoShot: Multi-Shot Portrait Video Generation
Jiahao Wang (Xi'an Jiaotong University), Jieping Ye (Alibaba Cloud Computing)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: This paper proposes the EchoShot framework, which enables multi-camera portrait video generation, supporting identity consistency and fine-grained control.
ECO: Evolving Core Knowledge for Efficient Transfer
Fu Feng (Southeast University), Xin Geng (Southeast University)
Computational EfficiencyKnowledge DistillationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes the ECO framework, which extracts core knowledge into reusable learngenes through evolutionary learning, enabling efficient knowledge transfer across models and tasks.
EddyFormer: Accelerated Neural Simulations of Three-Dimensional Turbulence at Scale
Yiheng Du (University of California Berkeley), Aditi S. Krishnapriyan (University of California Berkeley)
TransformerPhysics Related
🎯 What it does: EddyFormer is proposed, a three-dimensional turbulent numerical simulation model based on Transformer and Spectral Element Method (SEM) technology, achieving DNS-level accuracy and significant acceleration;
EDELINE: Enhancing Memory in Diffusion-based World Models via Linear-Time Sequence Modeling
Jia-Hua Lee (National Tsing Hua University), Chun-Yi Lee (National Taiwan University)
GenerationReinforcement LearningDiffusion modelWorld ModelSequential
🎯 What it does: A unified world model architecture EDELINE is proposed, which combines state space models and diffusion models to enhance the memory and generation quality of diffusion-based world models.
Edit Flows: Variable Length Discrete Flow Matching with Sequence-Level Edit Operations
Marton Havasi (Meta Platforms), Ricky T. Q. Chen (Meta Platforms)
GenerationFlow-based ModelImageText
🎯 What it does: This paper proposes Edit Flows, a non-autoregressive sequence generation framework based on continuous-time Markov chains, which utilizes three editing operations: insertion, deletion, and replacement to achieve variable-length generation.
Edit Less, Achieve More: Dynamic Sparse Neuron Masking for Lifelong Knowledge Editing in LLMs
Jinzhe Liu (University of Chinese Academy of Sciences), Shuhui Wang (University of Chinese Academy of Sciences)
TransformerLarge Language ModelText
🎯 What it does: Proposed Neuron-Specific Masked Knowledge Editing (NMKE), which achieves lifelong knowledge editing of LLMs through neuron-level attribution and dynamic sparse masking;
EditInfinity: Image Editing with Binary-Quantized Generative Models
Jiahuan Wang (Harbin Institute of Technology), Wenjie Pei (Harbin Institute of Technology)
Image TranslationGenerationPrompt EngineeringDiffusion modelImageBenchmark
🎯 What it does: This paper proposes EditInfinity, which utilizes the binary characteristics of the VQ quantization model Infinity to achieve text-driven high-fidelity image editing.
EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting
Bohao Liao, Zheng-Jun Zha
RestorationOptimizationNeural Radiance FieldGaussian SplattingVideoBenchmark
🎯 What it does: Utilizing event cameras to assist in free trajectory 3D Gaussian rendering, jointly optimizing camera trajectories and 3D scenes to enhance reconstruction and view synthesis quality under high-speed video.
Effective Neural Approximations for Geometric Optimization Problems
Samantha Chen (University of California), Yusu Wang (University of California)
OptimizationTransformerAuto EncoderPoint CloudMesh
🎯 What it does: A neural network approximation framework is proposed to solve geometric optimization problems, particularly the geometric 'range metric' problems, such as minimum enclosing spheres and shape fitting descriptors for rings.
Effective Policy Learning for Multi-Agent Online Coordination Beyond Submodular Objectives
Qixin Zhang (Nanyang Technological University), Dacheng Tao (Nanyang Technological University)
OptimizationReinforcement LearningGraph
🎯 What it does: A policy learning framework suitable for multi-agent online coordination (MA-OC) problems is proposed, achieving optimal approximation ratios under weakly submodular objectives.
Effects of Dropout on Performance in Long-range Graph Learning Tasks
Jasraj Singh (Nanyang Technological University), Laura Toni (University College London)
ClassificationGraph Neural NetworkGraph
🎯 What it does: This study investigates the impact of dropout-style algorithms such as DropEdge, DropNode, DropAgg, DropGNN, and DropMessage in long-distance graph learning tasks, and proposes a new method called DropSens.
Efficient $k$-Sparse Band–Limited Interpolation with Improved Approximation Ratio
Yang Cao (Wyoming Seminary), Chiwun Yang (Sun Yat-sen University)
OptimizationComputational Efficiency
🎯 What it does: This paper proposes a new multi-layer frequency decomposition framework for high-precision recovery of k-sparse band-limited signals from limited noisy time-domain samples, and presents a polynomial-time algorithm that proves its approximation ratio does not exceed 3 + √2 + ε, significantly lower than the previous constant upper limit of 100.
Efficient Adaptive Experimentation with Noncompliance
Miruna Oprescu (Cornell University), Nathan Kallus (Cornell University)
Tabular
🎯 What it does: In an adaptive experimental environment where only binary instrumental variables can be randomly assigned and treatment cannot be directly enforced, estimate the Average Treatment Effect (ATE).
Efficient Adaptive Federated Optimization
Su Hyeong Lee (University of Chicago), Tian Li (University of Chicago)
OptimizationFederated LearningComputational EfficiencyTransformerImageText
🎯 What it does: Two efficient joint adaptive federated learning algorithms, FedAda 2 and FedAda++ 2, are proposed to avoid the transmission of preconditioners and reduce device memory usage.
Efficient Algorithms for Robust and Partial Semi-Discrete Optimal Transport
Pankaj K Agarwal, Keegan Yao (Duke University)
OptimizationTabular
🎯 What it does: This paper studies the theoretical properties and algorithms of semi-discrete robust and partially optimal transport (α‑OPT and λ‑ROT), showing that their optimal solutions can be represented by constrained Laguerre diagrams.
Efficient Allocation of Working Memory Resource for Utility Maximization in Humans and Recurrent Neural Networks
Qingqing Yang (Ohio State University), Hsin-Hung Li (Ohio State University)
OptimizationReinforcement Learning from Human FeedbackRecurrent Neural NetworkReinforcement LearningSequential
🎯 What it does: This study investigates how humans allocate memory resources in working memory tasks based on learned rewards and natural statistical distributions, and validates the resource allocation mechanism under utility maximization through experiments, theoretical extensions, and recurrent neural network models.
Efficient and Generalizable Mixed-Precision Quantization via Topological Entropy
Nan Li (Shanxi University), Lianbo Ma (Northeastern University)
Object DetectionOptimizationComputational EfficiencyImage
🎯 What it does: This paper proposes an efficient transferable mixed-precision quantization method GMPQ-TE based on topological entropy.
Efficient and Near-Optimal Algorithm for Contextual Dueling Bandits with Offline Regression Oracles
Aadirupa Saha (University of Illinois), Robert E. Schapire (Microsoft Research)
OptimizationReinforcement LearningTabular
🎯 What it does: An efficient and approximately optimal algorithm is proposed for addressing the contextual bandit problem with an offline regression oracle, particularly in continuous action spaces.
Efficient Bayesian Experiment Design with Equivariant Networks
Conor Igoe (Carnegie Mellon University), Jeff Schneider (Carnegie Mellon University)
OptimizationGraph Neural NetworkTransformerReinforcement LearningSequential
🎯 What it does: This paper studies the issue of sample inefficiency caused by 'belief explosion' in Bayesian Experimental Design (BED) and proposes the use of Graph Neural Networks (GNN) to introduce domain permutation equivariance as a network prior to alleviate this problem. The efficiency of GNN in both offline and online learning is validated through two training methods: behavior cloning and reinforcement learning.
Efficient Data Selection at Scale via Influence Distillation
Mahdi Nikdan (Institute of Science and Technology Austria and Google Research), Vahab Mirrokni (Google Research)
Computational EfficiencyKnowledge DistillationData-Centric LearningLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies a second-order approximation-based influence distillation method for scalable and efficient sample selection of training data for large language models.
Efficient Fairness-Performance Pareto Front Computation
Mark Kozdoba (Technion Israel Institute of Technology), Shie Mannor (NVIDIA)
OptimizationComputational EfficiencyTabularBiomedical Data
🎯 What it does: A new method is proposed to compute the optimal Pareto front between fairness and performance, which does not require training complex fairness representation models.
Efficient Federated Learning against Byzantine Attacks and Data Heterogeneity via Aggregating Normalized Gradients
Shiyuan Zuo (Beijing Institute of Technology), Han Hu (Beijing Institute of Technology)
OptimizationFederated LearningConvolutional Neural NetworkImage
🎯 What it does: A Fed-NGA algorithm is proposed, utilizing normalized gradient aggregation to achieve robustness against Byzantine attacks and data heterogeneity in federated learning.
Efficient Hybrid Language Model Compression through Group-Aware SSM Pruning
Ali Taghibakhshi (NVIDIA), Pavlo Molchanov (NVIDIA)
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a group-aware pruning method for the Mamba layer and combines it with FFN, embedding dimension, and layer depth pruning to construct a unified compression scheme.
Efficient Kernelized Learning in Polyhedral Games beyond Full Information: From Colonel Blotto to Congestion Games
Andreas Kontogiannis (National Technical University of Athens and Archimedes Athena Research Center), Ioannis Panageas (University of California Irvine and Archimedes Athena Research Center)
Reinforcement Learning
🎯 What it does: This paper studies polygonal games with multiple players and multi-dimensional action sets (such as Colonel Blotto, graphical congestion games, and network congestion games), proposing a kernelization method to achieve efficient no-regret learning under partial information (bandit/semibandit), which converges to coarse correlated equilibria (CCE).
Efficient Knowledge Transfer in Federated Recommendation for Joint Venture Ecosystem
Yichen Li (Huazhong University of Science and Technology), Ruixuan Li (Chongqing Ant Consumer Finance Co., Ltd)
Recommendation SystemFederated LearningSafty and PrivacyComputational EfficiencyTabular
🎯 What it does: A federated recommendation framework FR-JVE aimed at joint enterprise ecosystems is proposed to address the issue of sharing partial user and product data among subsidiaries without compromising user privacy.
Efficient Large Language Model Inference with Neural Block Linearization
Mete Erdogan (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A method called Neural Block Linearization (NBL) is proposed, which compresses large-scale language models by replacing the self-attention layers of Transformers with linear layers derived from linear minimum mean square error (LMMSE), achieving inference acceleration.
Efficient Last-Iterate Convergence in Solving Extensive-Form Games
Linjian Meng (Nanjing University), Yang Gao (Nanjing University)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: The paper proposes an algorithm called Reward Transformation CFR+ (RTCFR+), which utilizes the parameter-independent RM+ (CFR+) to solve the perturbed regularization game at each iteration and achieves the last iteration convergence of the original game through the reward transformation framework.
Efficient Low Rank Attention for Long-Context Inference in Large Language Models
Li Tenghui, Qibin Zhao (RIKEN AIP)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A low-rank query and key attention (LRQK) framework is proposed, which supports efficient long-context reasoning on resource-constrained devices.
Efficient Multi-bit Quantization Network Training via Weight Bias Correction and Bit-wise Coreset Sampling
Jinhee Kim (Duke University), Jong Hwan Ko (Sungkyunkwan University)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: This study investigates the training efficiency of multi-bit quantized networks and proposes two methods: weight bias correction and bit-wise core sampling, which significantly reduce training costs without lowering or potentially improving the accuracy of various bit-width models.
Efficient Multi-modal Large Language Models via Progressive Consistency Distillation
Zichen Wen (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A training framework for multimodal large model visual token compression is proposed through Enhanced Progressive Consistency Distillation (EPIC), which is compatible with various compression methods without altering the model structure.
Efficient Multimodal Dataset Distillation via Generative Models
Zhenghao Zhao (University of Illinois Chicago), Yan Yan (University of Illinois Chicago)
RetrievalKnowledge DistillationDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: The EDGE method is proposed and implemented, utilizing a generative diffusion model for efficient distillation of image-text datasets, generating a small number of high-quality image-text pairs, and further enhancing retrieval performance through post-caption synthesis.
Efficient PAC Learning for Realizable-Statistic Models via Convex Surrogates
Shivani Agarwal (University of Pennsylvania)
🎯 What it does: This paper studies a class of intermediate PAC learning models—Realizable Statistical Models (RSM)—and presents a computationally feasible learning algorithm with finite sample complexity under this model, achieved through convex 'strongly composable' surrogate loss.
Efficient Parametric SVD of Koopman Operator for Stochastic Dynamical Systems
Minchan Jeong (KAIST), Gregory W. Wornell (MIT)
OptimizationRepresentation LearningContrastive LearningTime SeriesPhysics Related
🎯 What it does: A method for learning the Koopman operator's singular subspace based on low-rank approximation (LoRA) is proposed, which avoids numerically unstable operations and demonstrates its feasibility in downstream tasks such as multi-step prediction and feature extraction.
Efficient Part-level 3D Object Generation via Dual Volume Packing
Jiaxiang Tang (Peking University), Tsung-Yi Lin (Nvidia Research)
GenerationData SynthesisDiffusion modelRectified FlowAuto EncoderImagePoint CloudMesh
🎯 What it does: An end-to-end framework is proposed to directly generate editable part-level 3D models from single-view images.
Efficient Policy Optimization in Robust Constrained MDPs with Iteration Complexity Guarantees
Sourav Ganguly (New Jersey Institute of Technology), Adam Wierman (California Institute of Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: A natural policy gradient algorithm for robust constrained Markov decision processes (RCMDP) is proposed, which can satisfy constraints and obtain an approximately optimal policy simultaneously without the need for binary search.
Efficient Pre-Training of LLMs via Topology-Aware Communication Alignment on More Than 9600 GPUs
Guoliang HE, Eiko Yoneki (University of Cambridge)
Large Language ModelMixture of Experts
🎯 What it does: Proposed and implemented the Arnold scheduling system, aligning the communication patterns for large-scale LLM pre-training tasks on multi-layer network topologies to enhance training performance.
Efficient Preference-Based Reinforcement Learning: Randomized Exploration meets Experimental Design
Andreas Schlaginhaufen (Ecole Polytechnique Federale de Lausanne), Maryam Kamgarpour (Ecole Polytechnique Federale de Lausanne)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper proposes a series of meta-algorithms based on RLoracle, which learn the reward function from trajectory-level preference feedback in Markov decision processes using random exploration and maximum likelihood estimation. Two versions are provided: one for minimizing online cumulative returns (Algorithm 1) and another for preference-independent exploration and single batch estimation (Algorithm 2), along with an improved version that combines lazy random exploration with optimal design (Algorithm 3).
Efficient Prompt Compression with Evaluator Heads for Long-Context Transformer Inference
Weizhi Fei (Tsinghua University), Wei Han (Huawei Technologies Co., Ltd.)
CompressionComputational EfficiencyTransformerPrompt EngineeringTextBenchmark
🎯 What it does: A training-independent prompt compression method called EHPC is proposed, which utilizes evaluator heads in the Transformer to quickly identify important tokens during the pre-filling stage, thereby generating shorter prompts for LLM inference.
Efficient Quadratic Corrections for Frank-Wolfe Algorithms
Jannis Halbey (Zuse Institute Berlin), Sebastian Pokutta (Zuse Institute Berlin)
Optimization
🎯 What it does: The Corrective Frank-Wolfe (CFW) framework is proposed, and two efficient correction steps for quadratic objectives (QC-LP and QC-MNP) are designed on it, further accelerating algorithms such as Split Conditional Gradient (SCG) and Second-Order Conditional Gradient Sliding (SOCGS).
Efficient Randomized Experiments Using Foundation Models
Piersilvio De Bartolomeis (ETH Zurich and Harvard University), Issa Dahabreh
Large Language ModelTextTabular
🎯 What it does: A new random experiment estimation method H-AIPW is proposed, which utilizes predictions from multiple base models to improve efficiency while maintaining valid statistical inference.
Efficient RAW Image Deblurring with Adaptive Frequency Modulation
Wenlong Jiao (Tianjin University), Dongwei Ren (Tianjin University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: A frequency domain enhancement network specifically for RAW images, FrENet, is proposed, achieving RAW-to-RAW deblurring.
Efficient Rectified Flow for Image Fusion
Zirui Wang (City University of Hong Kong), Jinyuan Liu (Zhejiang University)
Image TranslationGenerationComputational EfficiencyDiffusion modelRectified FlowAuto EncoderImage
🎯 What it does: This paper proposes a single-step diffusion model RFfusion based on Rectified Flow for efficient image fusion, balancing fusion quality and speed.
Efficient Representativeness-Aware Coreset Selection
Zihao Cheng (Fudan University), WEIZHONG ZHANG
ClassificationOptimizationComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a lightweight dynamic core subset selection method called ERACS, which can dynamically track and maintain the representativeness of the core subset during the training process of deep networks.
Efficient Safe Meta-Reinforcement Learning: Provable Near-Optimality and Anytime Safety
Siyuan Xu (Pennsylvania State University), Minghui Zhu (Pennsylvania State University)
Computational EfficiencyMeta LearningReinforcement Learning
🎯 What it does: This paper proposes an efficient safe meta-reinforcement learning framework that ensures safety and near-optimality at all times while adapting to unknown tasks.
Efficient semantic uncertainty quantification in language models via diversity-steered sampling
Ji Won Park (Genentech), Kyunghyun Cho (Genentech)
GenerationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A diversity-driven sampling method is proposed, which generates semantically diverse samples by penalizing semantically similar outputs during language model generation, thereby more efficiently estimating semantic uncertainty in free-form question answering.
Efficient Spectral Control of Partially Observed Linear Dynamical Systems
Anand Paresh Brahmbhatt (Princeton University), Elad Hazan (Google DeepMind)
OptimizationComputational EfficiencyReinforcement Learning
🎯 What it does: A new Double Spectral Control (DSC) algorithm is proposed for controlling linear dynamical systems under partial observation, adversarial disturbances, and convex loss.
Efficient Speech Language Modeling via Energy Distance in Continuous Latent Space
Zhengrui Ma (Chinese Academy of Sciences), Min zhang
GenerationData SynthesisComputational EfficiencyTransformerLarge Language ModelAudio
🎯 What it does: A new framework for speech language modeling in continuous latent space, called SLED, is proposed, which uses Energy Distance as the training objective and directly performs autoregressive modeling on continuous acoustic vectors.
Efficient Training of Minimal and Maximal Low-Rank Recurrent Neural Networks
Anushri Arora (Princeton University), Jonathan W. Pillow (Princeton University)
Recurrent Neural NetworkTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: This paper proposes an efficient method for training low-rank recurrent neural networks (RNNs), including online recursive least squares (RLS) learning based on a neural engineering framework, using continuous orthogonal matching pursuit (OMP) to find minimal networks, mapping infinite-width low-rank RNNs to Gaussian processes (GP) and optimizing basis function distributions through maximum marginal likelihood, as well as an active learning sampling strategy based on maximum predictive variance.
Efficient Training-Free Online Routing for High-Volume Multi-LLM Serving
Fangzhou Wu (University of Wisconsin-Madison), Sandeep Silwal (University of Wisconsin-Madison)
Large Language ModelBenchmark
🎯 What it does: A training-free online multi-LLM routing algorithm is proposed, which can efficiently allocate requests to the best model under high query volume and limited token budget.
Efficient Utility-Preserving Machine Unlearning with Implicit Gradient Surgery
Shiji Zhou (Beihang University), Han Zhao (University of Illinois at Urbana-Champaign)
OptimizationComputational EfficiencyDiffusion modelImage
🎯 What it does: This paper proposes an efficient Utility-Preserving Machine Unlearning (EUPMU) method, which views model forgetting as a constrained optimization problem that maximizes the forgetting objective while maintaining original performance constraints, and solves it using a one-sided gradient surgery.
Efficient Verified Unlearning For Distillation
Yijun Quan (Warwick Manufacturing Group University of Warwick), Giovanni Montana (Warwick Manufacturing Group University of Warwick)
Computational EfficiencyKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: The PURGE framework is proposed, achieving efficient and verifiable machine unlearning of teacher data in a knowledge distillation environment.
Efficiently Escaping Saddle Points under Generalized Smoothness via Self-Bounding Regularity
Daniel Yiming Cao (Cornell University), Benjamin Tang (Cornell University)
Optimization
🎯 What it does: The study investigates the convergence of first-order methods for optimizing non-convex functions under generalized smoothness, particularly the convergence of first and second-order stationary points.
Efficiently Maintaining the Multilingual Capacity of MCLIP in Downstream Cross-Modal Retrieval Tasks
Fengmao Lv (Southwest Jiaotong University), Tianrui Li (Southwest Jiaotong University)
RetrievalComputational EfficiencyTransformerPrompt EngineeringContrastive LearningMultimodality
🎯 What it does: Two efficient multilingual CLIP downstream fine-tuning strategies (TaPCL and CiPCL) are proposed, achieving significant reductions in computational costs while maintaining multilingual capabilities.
Efficiently Scaling LLM Reasoning Programs with Certaindex
Yichao Fu (University of California San Diego), Hao Zhang (Carnegie Mellon University)
OptimizationComputational EfficiencyTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This paper proposes a model-independent metric called Certaindex by studying the answer stability during the reasoning process of LLMs, and utilizes this metric to achieve early stopping in the reasoning process.
Efficiently Verifiable Proofs of Data Attribution
Ari Karchmer (Morgan Stanley Machine Learning Research), Martin Pawelczyk (Harvard Business School)
🎯 What it does: A two-step interactive verification protocol is proposed, allowing a third party to verify the approximate optimality of data attribution results without executing complete model training.
EfficientNav: Towards On-Device Object-Goal Navigation with Navigation Map Caching and Retrieval
Zebin Yang (Peking University), Meng Li (Peking University)
OptimizationComputational EfficiencyRobotic IntelligenceTransformerLarge Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: A zero-range target navigation system, EfficientNav, has been implemented on local devices, using a small LLM as a planner. It compresses and reuses navigation map information through discrete KV caching, attention clustering, and semantic-aware retrieval to complete path planning from the starting point to the designated object.
EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models
Yantai Yang (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)
Computational EfficiencyRobotic IntelligenceTransformerVision-Language-Action ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: EfficientVLA is proposed, a fully training-independent and structured inference acceleration framework designed for Diffusion-based Vision-Language-Action (VLA) models. It achieves significant improvements in inference speed and computational cost by pruning language module layers, selecting task-aware visual tokens, and caching intermediate features of the diffusion action head.
Effortless, Simulation-Efficient Bayesian Inference using Tabular Foundation Models
Julius Vetter (University of Tübingen), Jakob H. Macke (Max Planck Institute for Intelligent Systems)
TransformerTabularBenchmark
🎯 What it does: A training-free simulation-based Bayesian inference method NPE-PFN is proposed, using the pre-trained TabPFN as a conditional density estimator to perform autoregressive inference of multi-dimensional posteriors.
EGGS: Exchangeable 2D/3D Gaussian Splatting for Geometry-Appearance Balanced Novel View Synthesis
Yancheng Zhang (Institute of Artificial Intelligence University of Central Florida), Chen Chen (Institute of Artificial Intelligence University of Central Florida)
GenerationData SynthesisOptimizationGaussian SplattingImagePoint Cloud
🎯 What it does: An exchangeable 2D/3D Gaussian scattering framework, EGGS, is proposed for achieving high-quality appearance rendering and precise geometric reconstruction simultaneously.
EgoBridge: Domain Adaptation for Generalizable Imitation from Egocentric Human Data
Ryan Punamiya (Georgia Institute of Technology), Danfei Xu (Georgia Institute of Technology)
Domain AdaptationRobotic IntelligenceTransformerReinforcement LearningVideo
🎯 What it does: Proposes the EgoBridge framework, which utilizes differentiable Optimal Transport for joint domain adaptation of human egocentric perspective data and robot demonstrations, thereby achieving cross-body imitation learning.
EgoDTM: Towards 3D-Aware Egocentric Video-Language Pretraining
Boshen Xu (Renmin University of China), Qin Jin (Renmin University of China)
Depth EstimationRetrievalRobotic IntelligenceTransformerLarge Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: This paper proposes EgoDTM, a first-person video-language pre-training model that integrates a lightweight 3D visual decoder and spatially aware subtitles.
EgoThinker: Unveiling Egocentric Reasoning with Spatio-Temporal CoT
Baoqi Pei (Shanghai Artificial Intelligence Laboratory), Yu Qiao
Object DetectionTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoMultimodalityChain-of-Thought
🎯 What it does: Proposes the EgoThinker framework, constructs a large-scale egocentric QA dataset EgoRe-5M, and achieves significant improvements in first-person video reasoning and fine-grained spatiotemporal localization through two-stage supervised fine-tuning and reinforcement learning training.
Elastic Robust Unlearning of Specific Knowledge in Large Language Models
Yize Sui (National University of Defense Technology), Ji Wang (National University of Defense Technology)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the Elastic Robust Unlearning (ERU) framework to address the rigidity of reward settings and insufficient robustness in LLM unlearning.
Elastic ViTs from Pretrained Models without Retraining
Walter Simoncini (University of Technology Nuremberg), Yuki M Asano
ClassificationSegmentationTransformerContrastive LearningImage
🎯 What it does: This paper proposes a SnapViT method that enables one-time structured pruning of a pre-trained Vision Transformer (ViT) without the need for retraining, generating an elastic model that can be used at any sparsity level.
ElasticMM: Efficient Multimodal LLMs Serving with Elastic Multimodal Parallelism
Zedong Liu (Institute of Computing Technology, Chinese Academy of Sciences), Dingwen Tao (Institute of Computing Technology, Chinese Academy of Sciences)
OptimizationComputational EfficiencyTransformerLarge Language ModelVision Language ModelMultimodality
🎯 What it does: Proposes ElasticMM, a MLLM service architecture based on Elastic Multi-Modal Parallel (EMP), achieving decoupling of request types and inference stages, as well as dynamic resource allocation;
ELDET: Early-Learning Distillation with Noisy Labels for Object Detection
Dongmin Choi (SAIGE), Frank C. Park (Seoul National University)
Object DetectionKnowledge DistillationImageBiomedical Data
🎯 What it does: This paper proposes a knowledge distillation framework called ELDET, which utilizes models from the early learning stage as a teacher network to simultaneously suppress classification noise and localization noise in object detection, thereby enhancing detection robustness.
ELECTRA: A Cartesian Network for 3D Charge Density Prediction with Floating Orbitals
Jonas Elsborg (Technical University of Denmark), Arghya Bhowmik (Technical University of Denmark)
GraphPhysics Related
🎯 What it does: This paper proposes the ELECTRA model, which utilizes floating orbitals to predict three-dimensional electron density, addressing the limitations of traditional atomic-centered orbital approaches.
Elevating Visual Perception in Multimodal LLMs with Visual Embedding Distillation
Jitesh Jain (Georgia Tech), Jianwei Yang (Meta Superintelligence Labs)
Knowledge DistillationRepresentation LearningTransformerLarge Language ModelVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a multimodal large language model (MLLM) training framework named VisPer-LM, which enhances visual perception capabilities by optimizing predictive embedding of intermediate layer representations in LLM during the pre-training phase, distilling knowledge from expert visual perception encoders into the hidden representations of LLM.
Eliciting Reasoning in Language Models with Cognitive Tools
Brown Ebouky (IBM Research), Mattia Rigotti (IBM Research)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A modular tool invocation framework based on cognitive architecture is proposed to stimulate reasoning abilities within LLMs through four cognitive tools (understanding the problem, recalling relevant information, checking answers, and backtracking).
ElliCE: Efficient and Provably Robust Algorithmic Recourse via the Rashomon Sets
Bohdan Turbal (Princeton University), Lesia Semenova (Rutgers University)
OptimizationExplainability and InterpretabilityTabularFinance Related
🎯 What it does: This paper proposes a robust algorithmic explanation framework called ElliCE, based on the Rashomon set elliptical approximation, which can generate effective counterfactual explanations under various approximately optimal models.