NeurIPS 2025 Papers — Page 15
Conference on Neural Information Processing Systems · 5275 papers
Elucidated Rolling Diffusion Models for Probabilistic Forecasting of Complex Dynamics
Salva Rühling Cachay (University of California San Diego), Rose Yu (University of California San Diego)
Convolutional Neural NetworkDiffusion modelTime Series
🎯 What it does: A novel rolling diffusion model (ERDM) is proposed, capable of long-term probabilistic forecasting for complex dynamical systems.
Eluder dimension: localise it!
Alireza Bakhtiari (University of Alberta), Csaba Szepesvari
OptimizationReinforcement Learning
🎯 What it does: This paper studies low regret algorithms under first-order (small-cost) conditions using localized eluder dimension in generalized linear models (GLM) and finite-horizon reinforcement learning; it proposes two algorithms: ℓ-UCB (bandit) and ℓ-GOLF (RL).
Embedding Principle of Homogeneous Neural Network for Classification Problem
Jiahan Zhang (Shanghai Jiao Tong University), Tao Luo (Shanghai Jiao Tong University)
ClassificationOptimizationTabular
🎯 What it does: This study proves that in homogeneous (bias-free, positive 1-homogeneous activation) neural networks, linear isometric transformations using neuron or channel splitting can embed the KKT points of the minimum norm maximum margin problem of smaller networks into the corresponding KKT points of larger networks. Moreover, during the gradient flow training process, this mapping preserves the trajectory and direction limits.
Embeddings as Probabilistic Equivalence in Logic Programs
Jaron Maene (KU Leuven), Efthymia Tsamoura (Huawei)
Reinforcement LearningGraphSequential
🎯 What it does: A new probabilistic equivalent semantics is proposed, transforming the soft unification in neural symbolic models into a probability distribution over equivalence relations, thereby addressing the non-transitivity issue of soft unification in reasoning and learning.
Embodied Cognition Augmented End2End Autonomous Driving
Ling Niu (Tsinghua University), Jiangtao Gong (Tsinghua University)
Autonomous DrivingOptimizationTransformerContrastive LearningVideoMultimodality
🎯 What it does: This paper proposes the E AD 3 paradigm, which combines human driving cognition (EEG) with visual features through contrastive learning to train a driving thought model and integrate it into various end-to-end autonomous driving frameworks to enhance planning performance.
Embodied Crowd Counting
Runling Long (Harbin Institute of Technology), Liqiang Nie (City University of Hong Kong)
Object DetectionRobotic IntelligenceTransformerLarge Language ModelAgentic AIImageMultimodality
🎯 What it does: This paper proposes the task of 'Embodied Crowd Counting', utilizing drones to actively explore large outdoor scenes and perform crowd counting.
Emergence and Evolution of Interpretable Concepts in Diffusion Models
Berk Tinaz (University of Southern California), Mahdi Soltanolkotabi (University of Southern California)
GenerationExplainability and InterpretabilityDiffusion modelAuto EncoderImageText
🎯 What it does: This study investigates the internal mechanisms of diffusion models, analyzes the evolution of concepts during the generation process, and utilizes sparse autoencoders to achieve interpretable editing control.
Emergence and scaling laws in SGD learning of shallow neural networks
Yunwei Ren (Princeton University), Jason D. Lee (Princeton University)
Optimization
🎯 What it does: This study investigates the convergence complexity and dynamics of training a two-layer neural network using online stochastic gradient descent (SGD) under high-dimensional normal inputs to learn single/multi-exponential objective functions with information exponent greater than 2.
Emergence of Linear Truth Encodings in Language Models
Shauli Ravfogel (New York University), Alberto Bietti (Flatiron Institute)
TransformerText
🎯 What it does: A single-layer Transformer toy model was constructed, and a synthetic dataset was designed in accordance with the Truth Co-occurrence Hypothesis to study the generation mechanism of linear truth subspaces.
Emergent Risk Awareness in Rational Agents under Resource Constraints
Daniel Jarne Ornia (University of Oxford), Michael J. Wooldridge
OptimizationReinforcement LearningAgentic AITabularSequentialFinance Related
🎯 What it does: This paper proposes and analyzes a rational agent decision-making model under resource constraints, deriving the risk preference shift caused by resource depletion using the survival bandit framework, and provides theoretical proof and empirical validation.
Emergent Temporal Correspondences from Video Diffusion Transformers
Jisu Nam (Korea Advanced Institute of Science and Technology), Seungryong Kim (Korea Advanced Institute of Science and Technology)
Object TrackingGenerationTransformerDiffusion modelVideo
🎯 What it does: Proposed and implemented the DiffTrack framework for quantitative analysis of temporal correspondence in video Diffusion Transformer (DiT), which can directly extract motion trajectories from generated videos, further achieving unsupervised point tracking and motion-enhanced video generation.
EMLoC: Emulator-based Memory-efficient Fine-tuning with LoRA Correction
Hsi-Che Lin (National Taiwan University), Yu-Chiang Frank Wang (NVIDIA)
Large Language ModelSupervised Fine-TuningTextMultimodality
🎯 What it does: The EMLoC framework is proposed, achieving fine-tuning of large models under the same memory budget as inference, significantly reducing the memory overhead of fine-tuning.
Empirical Study on Robustness and Resilience in Cooperative Multi-Agent Reinforcement Learning
Simin Li (Beihang University), Xianglong Liu (Beihang University)
Robotic IntelligenceReinforcement Learning
🎯 What it does: Large-scale experiments were conducted in real environments to systematically evaluate the relationship between cooperation, robustness, and resilience, and to explore the impact of hyperparameters on all three;
Empower Words: DualGround for Structured Phrase and Sentence-Level Temporal Grounding
Minseok Kang (Yonsei University), Sangyoun Lee (Yonsei University)
RetrievalTransformerVision Language ModelVideoText
🎯 What it does: Proposes the DualGround dual-branch architecture, which separates sentence-level and phrase-level semantics to enhance video temporal alignment performance.
Empowering Decision Trees via Shape Function Branching
Nakul Upadhya (University of Toronto), Eldan Cohen (University of Toronto)
ClassificationExplainability and InterpretabilityTabularBiomedical DataFinance Related
🎯 What it does: A decision tree with shape function branches, called Shape Generalized Tree (SGT), is proposed, along with an efficient tree construction algorithm, ShapeCART, which is further extended to bivariate and multi-way branches.
Enabling Differentially Private Federated Learning for Speech Recognition: Benchmarks, Adaptive Optimizers, and Gradient Clipping
Martin Pelikan (Apple), Tatiana Likhomanenko (Apple)
RecognitionFederated LearningSafty and PrivacyTransformerBenchmarkAudio
🎯 What it does: This paper studies the combination of federated learning and differential privacy in speech recognition tasks, proposing a training strategy based on hierarchical gradient clipping and hierarchical gradient normalization, and providing theoretical convergence analysis and practical benchmarks.
Enabling Instructional Image Editing with In-Context Generation in Large Scale Diffusion Transformer
Zechuan Zhang (Zhejiang University), Yi Yang (Zhejiang University)
GenerationData SynthesisTransformerMixture of ExpertsVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: This paper presents ICEdit, an instruction-based image editing framework based on a large Diffusion Transformer, which achieves editing results comparable to or even better than SOTA using only 0.1% of the training data, leveraging contextual editing prompts, a minimal number of fine-tuning parameters, and early filtering inference scaling.
Encoder-Decoder Diffusion Language Models for Efficient Training and Inference
Marianne Arriola (Cornell University), Volodymyr Kuleshov (Cornell University)
GenerationComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Proposes the E2D2 encoding-decoding diffusion architecture, which splits the diffusion model into an encoder (handling clean context) and a lightweight decoder (multi-step denoising), significantly improving inference and training efficiency.
EnCompass: Enhancing Agent Programming with Search Over Program Execution Paths
Zhening Li (Asari AI), Stephan Zheng (Asari AI)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackAgentic AIText
🎯 What it does: The ENCOMPASS framework is proposed, which decouples agent programming and reasoning strategies through Probabilistic Angel Non-determinism (PAN), supports the insertion of branching points within the program to construct the search space, and allows for flexible use of various search strategies (such as beam, best-of-N, backtracking, etc.) for reasoning expansion.
End-to-End Low-Light Enhancement for Object Detection with Learned Metadata from RAWs
Xuelin Shen (Guangdong Laboratory of Artificial Intelligence and Digital Economy), Wenhan Yang (Peng Cheng Laboratory)
Object DetectionImage
🎯 What it does: Improving sRGB images using compressed raw image metadata in low-light environments to enhance object detection performance.
End-to-End Vision Tokenizer Tuning
Wenxuan Wang (Chinese Academy of Sciences), Xinlong Wang (Beijing Academy of Artificial Intelligence)
GenerationOptimizationTransformerLarge Language ModelVision Language ModelGenerative Adversarial NetworkImageTextMultimodality
🎯 What it does: An end-to-end visual tokenizer tuning method called ETT is proposed, enabling the visual tokenizer to be jointly optimized with downstream autoregressive tasks.
Energy Landscape-Aware Vision Transformers: Layerwise Dynamics and Adaptive Task-Specific Training via Hopfield States
Runze Xia (Lancaster University), Richard Jiang (Lancaster University)
TransformerSupervised Fine-TuningImage
🎯 What it does: This study investigates the energy landscape of ViT layers, proposing the Layer Instability Index (LII) and implementing adaptive layer freezing for fine-tuning based on this index.
Energy Loss Functions for Physical Systems
Sékou-Oumar Kaba (McGill University), Siamak Ravanbakhsh (McGill University)
GenerationData SynthesisOptimizationDrug DiscoveryGraph Neural NetworkDiffusion modelGraphPhysics Related
🎯 What it does: This paper proposes a framework that directly embeds physical energy into the loss function, implemented in both continuous and discrete physical systems, primarily targeting regression and generation tasks for molecular and spin systems.
Energy Matching: Unifying Flow Matching and Energy-Based Models for Generative Modeling
Michal Balcerak (University of Zurich), Bjoern Menze (University of Zurich)
GenerationData SynthesisDiffusion modelFlow-based ModelImage
🎯 What it does: Proposes the Energy Matching framework, which combines flow matching with energy-based models to achieve the OT path from noise to data and forms a Boltzmann distribution at the data approximation.
Energy-based generator matching: A neural sampler for general state space
Dongyeop Woo (Korea Advanced Institute of Science and Technology), Sungsoo Ahn (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisGenerative Adversarial NetworkTabularStochastic Differential Equation
🎯 What it does: This paper proposes an energy-based generator matching (EGM) framework that utilizes continuous-time Markov processes (such as diffusion, flow, jumps, etc.) to train the generator with only an energy function and without balanced samples, achieving high-quality sampling in continuous, discrete, and mixed state spaces.
EnerVerse: Envisioning Embodied Future Space for Robotics Manipulation
Siyuan Huang (Shanghai Jiao Tong University), Guanghui Ren (AgiBot)
GenerationRobotic IntelligenceDiffusion modelGaussian SplattingVideo
🎯 What it does: ENERVERSE is proposed, a robot generative foundation model based on video diffusion, capable of predicting and constructing a four-dimensional (3D + time) future space for robots according to instructions, and achieving practical control through an action policy head.
Enforcing convex constraints in Graph Neural Networks
Ahmed Rashwan (University of Bath), Lisa Maria Kreusser (University of Bath)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A graph neural network framework named ProjNet is proposed, aimed at satisfying input-dependent constraints, combining sparse vector clipping methods and the Component Average Dykstra (CAD) algorithm.
Enforcing Hard Linear Constraints in Deep Learning Models with Decision Rules
Gonzalo E. Constante Flores (Purdue University), Can Li (Purdue University)
OptimizationSafty and PrivacyTabularFinance Related
🎯 What it does: In safety-critical deep learning tasks, a model-agnostic framework is proposed, which combines the task network and safety network through a convex combination, and utilizes decision rules to ensure that the output satisfies linear equality and inequality constraints across the entire input space, without the need for iteration or online optimization during both training and inference phases.
Enhanced Cyclic Coordinate Descent Methods for Elastic Net Penalized Linear Models
Yixiao Wang (Duke University), Aditya Devarakonda (Wake Forest University)
OptimizationComputational EfficiencyTabularBenchmark
🎯 What it does: An Enhanced Coordinate Descent (ECCD) framework is proposed to solve generalized linear models with elastic network constraints, significantly reducing training time.
Enhanced Expert Merging for Mixture-of-Experts in Graph Foundation Models
Lei Liu (Wuhan University), Min Peng (Wuhan University)
Domain AdaptationComputational EfficiencyKnowledge DistillationGraph Neural NetworkMixture of ExpertsGraph
🎯 What it does: This paper conducts an in-depth study of the AnyGraph model based on Mixture of Experts (MoE) and proposes two enhanced expert fusion strategies that significantly improve the cross-domain generalization performance of multi-graph data.
Enhanced Self-Distillation Framework for Efficient Spiking Neural Network Training
Xiaochen Zhao (Zhejiang University), Aili Wang (Zhejiang University)
Computational EfficiencyKnowledge DistillationConvolutional Neural NetworkSpiking Neural NetworkSupervised Fine-TuningImage
🎯 What it does: An enhanced self-distillation framework is proposed, using a lightweight ANN branch to guide SNN training, balancing rate-coded gradients and significantly reducing BPTT costs.
Enhancing 3D Reconstruction for Dynamic Scenes
Jisang Han (KAIST), Seungryong Kim (KAIST)
Pose EstimationDepth EstimationAutonomous DrivingOptical FlowImagePoint Cloud
🎯 What it does: This paper proposes DUSt3R 2, which can directly regress the Static-Dynamic Alignment Point (SDAP) map in dynamic scenes, achieving 3D reconstruction with a single forward network.
Enhancing Bioactivity Prediction via Spatial Emptiness Representation of Protein-ligand Complex and Union of Multiple Pockets
Zhiyuan Zhou (Nanyang Technological University), Adams Wai-Kin Kong (Nanyang Technological University)
Drug DiscoveryGraph Neural NetworkBiomedical Data
🎯 What it does: The LigoSpace method is proposed, which enhances the accuracy of protein-ligand activity prediction by introducing spatial void representation GeoREC, joint multi-pocket information Union-Pocket, and pairwise loss in a 3D graph neural network.
Enhancing CLIP Robustness via Cross-Modality Alignment
Xingyu Zhu (University of Science and Technology of China), Hanwang Zhang (Nanyang Technological University)
Representation LearningAdversarial AttackTransformerContrastive LearningImage
🎯 What it does: A training-independent COLA method is proposed, which utilizes subspace projection to remove non-semantic noise from image features, and combines optimal transport (OT) to achieve global cross-modal alignment and local structural consistency, thereby enhancing the robustness of CLIP against adversarial attacks.
Enhancing Compositional Reasoning in CLIP via Reconstruction and Alignment of Text Descriptions
Jihoon Kwon (Seoul National University), Jy-yong Sohn (Yonsei University)
RetrievalRepresentation LearningTransformerContrastive LearningText
🎯 What it does: The READ method is proposed, which adds two auxiliary objectives of token-level reconstruction and sentence-level alignment based on the contrastive learning of CLIP to improve the model's performance on compositional reasoning tasks.
Enhancing Consistency of Flow-Based Image Editing through Kalman Control
Haozhe Chi (Peking University), Yadong MU
Image TranslationOptimizationFlow-based ModelRectified FlowImage
🎯 What it does: This paper proposes a Kalman control-based streaming image editing framework called Kalman-Edit, which addresses the shortcomings of traditional control methods in maintaining structural consistency in non-target areas. Furthermore, an accelerated version, Kalman-Edit*, is introduced to achieve more efficient editing.
Enhancing Contrastive Learning with Variable Similarity
Haowen Cui (Nanjing University of Science and Technology), Jian Yang (Nanjing University of Science and Technology)
Object DetectionRetrievalRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A Variable Similarity mechanism is proposed to dynamically adjust the similarity of positive sample pairs based on different levels of data augmentation in contrastive learning, thereby overcoming the semantic inconsistency caused by augmentation.
Enhancing Deep Batch Active Learning for Regression with Imperfect Data Guided Selection
Yinjie Min (Nankai University), Yongdao Zhou (Nankai University)
OptimizationData-Centric LearningTabularTime SeriesFinance Related
🎯 What it does: This paper proposes to improve active learning for regression tasks by using auxiliary data through density ratio weighted loss estimation, overcoming the challenge of estimating prediction uncertainty when there are no labels.
Enhancing Diffusion-based Unrestricted Adversarial Attacks via Adversary Preferences Alignment
Kaixun Jiang (Fudan University), Wenqiang Zhang (Peking University)
Adversarial AttackConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: This paper views unconstrained adversarial attacks as a preference alignment problem and proposes a two-stage framework called APA: first aligning visual consistency using LoRA and a differentiable similarity reward, and then enhancing attack effectiveness through dual-path attack guidance and diffusion enhancement, significantly improving black-box transfer rates.
Enhancing Graph Classification Robustness with Singular Pooling
Sofiane ENNADIR, Johannes F. Lutzeyer (École Polytechnique)
ClassificationAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: This paper proposes a robust pooling method based on the principal singular vector, RS-Pool, to enhance the adversarial robustness of GNNs in graph classification tasks.
Enhancing GUI Agent with Uncertainty-Aware Self-Trained Evaluator
Gongwei Chen (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
OptimizationTransformerLarge Language ModelReinforcement LearningContrastive LearningMultimodalitySequential
🎯 What it does: A framework for Uncertainty-Aware Reinforcement Self-Training (URST) is proposed, which trains a GUI trajectory evaluator using a lightweight multimodal large language model (MLLM) that can self-generate reasoning and judgments and continuously improve.
Enhancing Infrared Vision: Progressive Prompt Fusion Network and Benchmark
Jinyuan Liu (Dalian University of Technology), Risheng Liu (Dalian University of Technology)
RestorationTransformerPrompt EngineeringImageBenchmark
🎯 What it does: This paper addresses the problem of thermal infrared image enhancement by proposing the Progressive Prompt Fusion Network (PPFN) and Selective Progressive Training (SPT) methods, which can simultaneously handle noise, blur, and low contrast degradation in single or composite degradation scenarios.
Enhancing Interpretability in Deep Reinforcement Learning through Semantic Clustering
Liang Zhang (University of Arizona), Adarsh Pyarelal (University of Arizona)
Explainability and InterpretabilityReinforcement LearningImage
🎯 What it does: This study investigates the semantic clustering characteristics in deep reinforcement learning (DRL) and proposes a semantic clustering module that combines feature dimensionality reduction with online clustering to enhance the interpretability of DRL models.
Enhancing LLM Planning for Robotics Manipulation through Hierarchical Procedural Knowledge Graphs
Jiacong Zhou (Hangzhou Dianzi University), Jun Yu (Harbin Institute of Technology)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A hierarchical program knowledge graph (HP-KG) has been constructed, and this graph is used to enhance the capabilities of LLMs in robotic operation planning.
Enhancing LLM Watermark Resilience Against Both Scrubbing and Spoofing Attacks
Huanming Shen (University of Electronic Science and Technology of China), Xiaojun Wan (Peking University)
GenerationAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: By introducing an equivalent texture key mechanism and the SEEK scheme, the robustness of LLM watermarking against text rewriting (scrubbing) and forgery (spoofing) attacks is enhanced.
Enhancing Optimizer Stability: Momentum Adaptation of The NGN Step-size
Rustem Islamov (University of Basel), Aurelien Lucchi (Max Planck Institute for Intelligent Systems)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: An optimizer combining NGN step size and momentum, NGN-M, and its coordinate-level variant NGN-MD are proposed, and their stability and performance are validated in various deep learning tasks.
Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward
Yanming Wan (University of Washington), Natasha Jaques (University of California, Berkeley)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a curiosity-driven user model reward (CURIO) to enhance personalized interactions of LLMs in multi-turn dialogues.
Enhancing Privacy in Multimodal Federated Learning with Information Theory
Tianzhe Xiao (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
Federated LearningSafty and PrivacyImageTextMultimodality
🎯 What it does: Designed and implemented Sec-MMFL, a privacy protection framework for multimodal federated learning based on information theory.
Enhancing Safety in Reinforcement Learning with Human Feedback via Rectified Policy Optimization
Xiyue Peng (ShanghaiTech University), Xin Liu (ShanghaiTech University)
OptimizationSafty and PrivacyReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a reinforcement learning method called Rectified Policy Optimization (RePO) to address the 'safety compensation' problem that arises during the safety alignment of large language models (LLMs), where merely satisfying expected safety constraints still leaves some prompt-response pairs unsafe.
Enhancing Sample Selection Against Label Noise by Cutting Mislabeled Easy Examples
Suqin Yuan (University of Sydney), Tongliang Liu (University of Sydney)
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A sample selection strategy called Early Cutting is proposed, which recalibrates and removes easily mislearned noisy samples (Mislabeled Easy Examples, MEEs) that are mislearned in the early stages of learning, improving the model's generalization performance on noisy labeled data.
Enhancing Tactile-based Reinforcement Learning for Robotic Control
Elle Miller (University of Edinburgh), Sethu Vijayakumar (University of Edinburgh)
Robotic IntelligenceReinforcement LearningMultimodalityBenchmark
🎯 What it does: In the absence of vision, using sparse binary tactile sensors and self-supervised learning to enhance the robustness and dexterity of robotic manipulation.
Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders
Ali Rasekh (Leibniz University Hannover), Mohsen Fayyaz (Microsoft)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoMultimodality
🎯 What it does: In the Video-LLM framework, a time attention module is proposed and implemented in the visual encoder, enabling the model to better capture temporal information in video question answering and action recognition tasks.
Enhancing Text-to-Image Diffusion Transformer via Split-Text Conditioning
Yu Zhang (Tongji University), Longbing Cao (Macquarie University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: The DiT-ST framework is proposed, which alleviates the understanding deficiencies of DiT regarding complete text by breaking down the full text prompt into hierarchical simplified sentences (split-text) and gradually injecting them in the order of object-relation-attribute during the diffusion process.
Enhancing the Maximum Effective Window for Long-Term Time Series Forecasting
Jiahui Zhang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
TransformerTime Series
🎯 What it does: This paper proposes the Maximum Effective Window (MEW) metric and constructs two modules: the Information Bottleneck Filter (IBF) and the Hybrid Transformer-Mamba (HTM), to enhance the Transformer model's utilization of long-window time series. Based on this, the PIH model is designed to achieve a longer backtracking window and state-of-the-art (SOTA) prediction performance.
Enhancing the Outcome Reward-based RL Training of MLLMs with Self-Consistency Sampling
Jiahao Wang (Xi'an Jiaotong University), Jinguo Zhu (Shanghai Artificial Intelligence Laboratory)
TransformerLarge Language ModelReinforcement LearningVision Language ModelTextMultimodality
🎯 What it does: Design and implement the Self-Consistency Sampling (SCS) method, which uses truncated resampling and visual perturbations to generate consistency rewards, thereby enhancing the reliability and accuracy of reward-based reinforcement learning in multi-choice tasks of multimodal large language models.
Enhancing Training Data Attribution with Representational Optimization
Weiwei Sun (Carnegie Mellon University), Yiming Yang (Carnegie Mellon University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A scalable training data attribution method called AirRep is proposed, which measures the impact of training data on model predictions by learning representations aligned with tasks and models.
Enhancing Vision-Language Model Reliability with Uncertainty-Guided Dropout Decoding
Yixiong Fang (Carnegie Mellon University), Jiawei Zhou (Stony Brook University)
RecognitionGenerationData-Centric LearningTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A method for uncertainty-guided dropout decoding of visual tokens during the inference phase is proposed, which enhances the reliability of large visual-language models (LVLMs) by removing high-uncertainty tokens.
Enhancing Visual Prompting through Expanded Transformation Space and Overfitting Mitigation
Shohei Enomoto (NTT)
ClassificationRecognitionConvolutional Neural NetworkSupervised Fine-TuningPrompt EngineeringImage
🎯 What it does: A new visual prompting method ACAVP is proposed, which combines affine, color, and additive transformations to achieve task-specific fine-tuning of pre-trained visual models.
Enhancing Zero-Shot Black-Box Optimization via Pretrained Models with Efficient Population Modeling, Interaction, and Stable Gradient Approximation
Muqi Han (Xidian University), Handing Wang (Xidian University)
OptimizationTransformerTabularTime SeriesBenchmark
🎯 What it does: Proposed the EPOM (Enhanced Pretrained Optimization Model) to achieve zero-shot black-box optimization;
Enigmata: Scaling Logical Reasoning in Large Language Models with Synthetic Verifiable Puzzles
Jiangjie Chen (ByteDance), Mingxuan Wang (ByteDance)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The ENIGMATA suite is proposed, which includes 36 types of logic puzzles that can be generated and verified, and trains LLMs based on RLVR to enhance logical reasoning abilities.
ENMA: Tokenwise Autoregression for Continuous Neural PDE Operators
Armand Kassaï Koupaï (Sorbonne Université), Patrick Gallinari (Criteo AI Lab)
GenerationData SynthesisCompressionTransformerFlow-based ModelAuto EncoderTime SeriesSequential
🎯 What it does: The ENMA model is designed, which is a continuous hidden layer autoregressive generative neural operator for the generation and prediction of time-parameterized PDEs.
Entropic Time Schedulers for Generative Diffusion Models
Dejan Stancevic (Radboud University), Luca Ambrogioni (Radboud University)
GenerationDiffusion modelImageStochastic Differential Equation
🎯 What it does: This paper proposes a time scheduler based on conditional entropy, which selects sampling points in an entropy-balanced manner to enhance the inference performance of diffusion models.
Entropy Rectifying Guidance for Diffusion and Flow Models
Tariq Berrada (Meta), Karteek Alahari (Montreal Institute for Learning Algorithms)
GenerationData SynthesisTransformerDiffusion modelFlow-based ModelImageTextMultimodality
🎯 What it does: A guidance method for adjusting the energy of attention layers during inference, called Entropy Rectifying Guidance (ERG), is proposed for sampling in diffusion and flow models, aiming to simultaneously enhance image quality, diversity, and consistency.
Entropy-Calibrated Label Distribution Learning
Yunan Lu (Hong Kong Polytechnic University), Lei Yang (Nanjing University of Science and Technology)
Supervised Fine-TuningTabular
🎯 What it does: This paper studies the prediction error bias of low-entropy samples in label distribution learning (LDL), proposing to reduce anchor vector cohesion through angular regularization and introducing the Entropy Calibration Aggregation (ECA) method to eliminate numerical imbalance in evaluation bias.
Environment Inference for Learning Generalizable Dynamical System
Shixuan Liu (National University of Defense Technology), Zhong Liu (National University of Defense Technology)
Domain AdaptationOptimizationTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes a prediction error-based unlabeled environment inference method called DynaInfer, which automatically assigns environment labels to trajectories in multi-environment dynamic system learning and trains a generalizable dynamic model.
EnzyControl: Adding Functional and Substrate-Specific Control for Enzyme Backbone Generation
Chao Song (Northwestern Polytechnical University), Yang Zhang (National University of Singapore)
GenerationDrug DiscoveryProtein Structure PredictionFlow-based ModelTabular
🎯 What it does: A method called EnzyControl is proposed, which can achieve both functional site retention and substrate specificity control during the enzyme scaffold generation process.
EPA: Boosting Event-based Video Frame Interpolation with Perceptually Aligned Learning
Yuhan Liu (Xiamen University), Yongjian Deng (Beijing University of Technology)
Image TranslationRestorationConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: Proposes the EPA framework, which achieves lossless and degradation-resistant video frame interpolation by utilizing the semantic perception features of visual foundation models and event-guided alignment.
Epistemic Uncertainty Estimation in Regression Ensemble Models with Pairwise Epistemic Estimators
Lucas Berry (McGill University), David Meger (McGill University)
OptimizationComputational EfficiencyFlow-based ModelTabularTime Series
🎯 What it does: A non-sampling method based on Pairwise Distance Estimation (PaiDEs) called PairEpEsts is proposed for efficiently estimating concept uncertainty in regression ensemble models, and it is applied to regularized Normalizing Flow (Nflows Base) and Probabilistic Network Ensembles (PNEs), further utilized in active learning scenarios.
Epistemic Uncertainty for Generated Image Detection
Jun Nie (University of Science and Technology of China), Xinmei Tian (University of Science and Technology of China)
Anomaly DetectionTransformerContrastive LearningImage
🎯 What it does: A method is proposed to detect AI-generated images by estimating empirical uncertainty through weight perturbation, utilizing the prior knowledge of large-scale visual models.
Equi-mRNA: Protein Translation Equivariant Encoding for mRNA Language Models
Mehdi Yazdani-Jahromi (University of Central Florida), Ozlem Garibay
Protein Structure PredictionTransformerLarge Language ModelBiomedical Data
🎯 What it does: A covariant encoding framework called Equi-mRNA based on SO(2) group theory is proposed, which explicitly encodes the symmetry of synonymous codons in mRNA, enhancing the performance of language models for protein-coding sequences.
Equilibrium Policy Generalization: A Reinforcement Learning Framework for Cross-Graph Zero-Shot Generalization in Pursuit-Evasion Games
Runyu Lu (University of Chinese Academy of Sciences), Cesare Alippi (Università della Svizzera italiana)
Graph Neural NetworkReinforcement LearningGraph
🎯 What it does: Designed the Equilibrium Policy Generalization (EPG) framework for training pursuit-avoidance game strategies with zero-shot generalization performance under different graph structures.
EquiTabPFN: A Target-Permutation Equivariant Prior Fitted Network
Michael Arbel (INRIA), Frank Hutter (University of Freiburg)
ClassificationComputational EfficiencyTransformerTabular
🎯 What it does: Proposes the EquiTabPFN architecture, achieving unsupervised learning of tabular data for any number of categories.
Equivariance by Contrast: Identifiable Equivariant Embeddings from Unlabeled Finite Group Actions
Tobias Schmidt (Institute of Computational Biology Helmholtz Munich), Matthias Bethge (Tübingen AI Center)
Representation LearningContrastive LearningImage
🎯 What it does: In the context of observing actions of unlabeled finite groups, a recognizable equivariant embedding space is constructed using contrastive learning, and a linear mapping of group actions is achieved through implicit linear group representations;
Equivariance Everywhere All At Once: A Recipe for Graph Foundation Models
Ben Finkelshtein (University of Oxford), Ron Levie (Technion Israel Institute of Technology)
ClassificationGraph Neural NetworkGraph
🎯 What it does: A universal graph foundational model framework based on triple symmetry (node, label permutation equivalence, feature permutation invariance) is proposed, and its universal approximation capability on multi-set functions is proven.
Equivariant Eikonal Neural Networks: Grid-Free, Scalable Travel-Time Prediction on Homogeneous Spaces
Alejandro García-Castellanos (Amsterdam Machine Learning Lab University of Amsterdam), Erik J Bekkers
OptimizationMeta LearningPoint CloudPhysics Related
🎯 What it does: This paper proposes a mesh-free, scalable Equivariant Neural Eikonal Solver (E-NES) that combines equivariant neural fields with physics-informed networks to predict travel times of the Eikonal equation on Riemannian homogeneous spaces.
EraseFlow: Learning Concept Erasure Policies via GFlowNet-Driven Alignment
Naga Sai Abhiram kusumba, Yezhou Yang (Arizona State University)
GenerationAdversarial AttackDiffusion modelImageText
🎯 What it does: The EraseFlow framework is proposed, which achieves concept elimination for pre-trained text-to-image diffusion models through the trajectory balance of GFlowNets.
Erasing Conceptual Knowledge from Language Models
Rohit Gandikota (Northeastern University), David Bau (Northeastern University)
TransformerLarge Language ModelText
🎯 What it does: The paper proposes a method for achieving concept-level 'forgetting' through the self-classification ability of the language model itself, namely Erasing Language Memory (ELM).
Error Broadcast and Decorrelation as a Potential Artificial and Natural Learning Mechanism
Mete Erdogan (Koc University), Alper Tunga Erdogan
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A new neural network learning framework called Error Broadcast and Decorrelation (EBD) is proposed, which addresses the credit assignment problem by minimizing the correlation between output error and layer activations across layers.
Error Feedback under $(L_0,L_1)$-Smoothness: Normalization and Momentum
Sarit Khirirat (King Abdullah University of Science and Technology), Peter Richtárik (King Abdullah University of Science and Technology)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes and analyzes a normalized error feedback algorithm (||EF21|| and its momentum version ||EF21-SGDM||) for distributed non-convex optimization, providing convergence rates under generalized smoothness ((L,L0,1)-smooth) conditions;
Error Forcing in Recurrent Neural Networks
A Erdem Sağtekin (New York University), Cristina Savin (New York University)
Recurrent Neural NetworkTime SeriesSequential
🎯 What it does: An Error Forcing (EF) mechanism is proposed to dynamically guide network activity towards the zero-error manifold in recurrent neural network learning, improving temporal credit assignment;
ErrorTrace: A Black-Box Traceability Mechanism Based on Model Family Error Space
Chuanchao Zang (Shandong University), Shanqing Guo (Shandong University)
Large Language ModelText
🎯 What it does: This paper proposes ErrorTrace, a black-box traceability mechanism based on the error space of model families;
ESCA: Contextualizing Embodied Agents via Scene-Graph Generation
Jiani Huang (University of Pennsylvania), Mayur Naik (University of Pennsylvania)
RecognitionObject DetectionRobotic IntelligenceTransformerLarge Language ModelVision Language ModelContrastive LearningVideoTextMultimodality
🎯 What it does: The ESCA framework is proposed, utilizing the open-source CLIP model SGClip to generate scene graphs to enhance the perception and planning of multimodal large language models (MLLMs) in embodied agents.
ESCA: Enabling Seamless Codec Avatar Execution through Algorithm and Hardware Co-Optimization for Virtual Reality
Mingzhi Zhu (New York University), Sai Qian Zhang (New York University)
CompressionOptimizationComputational EfficiencySupervised Fine-TuningVideo
🎯 What it does: The ESCA framework is proposed to achieve real-time inference of Codec Avatar (PCA) on VR devices, with algorithm and hardware co-optimization.
Escaping Collapse: The Strength of Weak Data for Large Language Model Training
Kareem Amin (Google Research), Sergei Vassilvitskii (Google Research)
GenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A theoretical framework based on the perspective of boosting is proposed, studying how to achieve continuous improvement of large models using only weak labels (β-weak labeler) and synthetic data by gradually focusing on hard examples. A convergence proof is provided along with a corresponding iterative algorithm (Algorithm 2).
Escaping saddle points without Lipschitz smoothness: the power of nonlinear preconditioning
Alexander Bodard (KU Leuven), Panagiotis Patrinos (KU Leuven)
Optimization
🎯 What it does: This paper studies the convergence and saddle point escape properties of nonlinear preconditioned gradient descent in non-convex optimization, and proposes a new smoothness condition, proving its validity in key problems such as phase recovery and matrix decomposition.
ESCORT: Efficient Stein-variational and Sliced Consistency-Optimized Temporal Belief Representation for POMDPs
Yunuo Zhang (Vanderbilt University), Abhishek Dubey (Vanderbilt University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: Proposes the ESCORT framework, which combines particle evolution with SVGD, correlation projection, and temporal consistency to improve the belief representation of POMDP.
Establishing Linear Surrogate Regret Bounds for Convex Smooth Losses via Convolutional Fenchel–Young Losses
Yuzhou Cao (Nanyang Technological University), Bo An (Nanyang Technological University)
Optimization
🎯 What it does: This paper proposes a general construction method that utilizes convolutional Fenchel-Young loss and infimal convolution to construct a surrogate loss that is both convex and smooth, and satisfies the linear surrogate regret bound.
Estimating cognitive biases with attention-aware inverse planning
Sounak Banerjee (New York University), Mark K Ho
Autonomous DrivingReinforcement LearningSequential
🎯 What it does: This paper proposes and implements an attention-aware inverse planning framework to infer cognitive biases from human behavior trajectories.
Estimating Hitting Times Locally at Scale
Themistoklis Haris (Boston University), Charalampos Tsourakakis (Boston University)
Graph
🎯 What it does: This paper proposes two local algorithms that estimate the hitting time between any two points by performing a small number of random walks only near the target node.
Estimating Interventional Distributions with Uncertain Causal Graphs through Meta-Learning
Anish Dhir (Imperial College London), Mark van der Wilk (University of Oxford)
Meta LearningTransformerBiomedical Data
🎯 What it does: This paper proposes an end-to-end Meta-learning framework MACE-TNP for directly predicting the post-intervention distribution from observational data in the presence of uncertainty in causal graph structures.
Estimating Model Performance Under Covariate Shift Without Labels
Jakub Białek (NannyML NV), Nikolaos Perrakis (NannyML NV)
ClassificationDomain AdaptationTabularBenchmark
🎯 What it does: This paper proposes a label-free performance estimation method called PAPE, specifically designed to evaluate the performance of binary classifiers under covariate shift.
Estimation of Stochastic Optimal Transport Maps
Sloan Nietert (École Polytechnique Fédérale de Lausanne), Ziv Goldfeld (Cornell University)
OptimizationTabular
🎯 What it does: This paper proposes a new error metric E_p for evaluating the quality of random transport maps (OT maps) under conditions that do not satisfy the traditional Brenier theorem, and based on this, designs an efficient 'rounding' estimator.
Estimation of Treatment Effects in Extreme and Unobserved Data
Jiyuan Tan (Stanford University), Jose Blanchet (Stanford University)
TabularTime Series
🎯 What it does: This study investigates how to estimate treatment effects in extreme rare events and provides corresponding theories and estimation methods.
EUGens: Efficient, Unified and General Dense Layers
Sang Min Kim (Seoul National University), Krzysztof Marcin Choromanski (Google)
Computational EfficiencyKnowledge DistillationImageTextPoint Cloud
🎯 What it does: An efficient unified general dense layer (EUGen) based on random features is proposed, which can replace traditional fully connected feedforward layers to reduce computation and parameter count.
Eulerian Neural Network Informed by Chemical Transport for Air Quality Forecasting
Xukai Zhang (Beijing Institute of Technology), Sijie Ruan (Beijing Institute of Technology)
Convolutional Neural NetworkRecurrent Neural NetworkTime Series
🎯 What it does: A physics-informed neural network CTENet based on Eulerian representation is proposed for air quality prediction;
EvaLearn: Quantifying the Learning Capability and Efficiency of LLMs via Sequential Problem Solving
Shihan Dou (Fudan University), Xuanjing Huang (Fudan University)
TransformerLarge Language ModelSupervised Fine-TuningTextSequentialBenchmark
🎯 What it does: The EvaLearn benchmark is proposed, which evaluates the learning ability and efficiency of LLMs in a sequential manner across six types of tasks with 182 sequences (each containing 7 questions, totaling 648 challenging questions) and provides five comprehensive learning metrics.
Evaluating and Learning Optimal Dynamic Treatment Regimes under Truncation by Death
Sihyung Park (North Carolina State University), Shu Yang (North Carolina State University)
Biomedical DataElectronic Health Records
🎯 What it does: A new principle stratification method for evaluating and learning multi-stage dynamic treatment regimes (DTR) under the problem of truncated death is proposed, focusing on the value function of the always-survivor subgroup.
Evaluating LLM-contaminated Crowdsourcing Data Without Ground Truth
Yichi Zhang (Rutgers University), Yang Liu (University of California)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a conditional evaluation mechanism that utilizes 'peer prediction' and low-cost signals generated by LLMs in the absence of true labels, aimed at identifying low-effort, LLM-assisted cheating crowd workers.
Evaluating LLMs in Open-Source Games
Swadesh Sistla (University of Washington), Max Kleiman-Weiner (University of Washington)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper studies the performance of large language models (LLMs) in open-source games, exploring their understanding of code, strategy generation, and evolutionary balance.
Evaluating multiple models using labeled and unlabeled data
Divya M Shanmugam, Emma Pierson (University of California Berkeley)
ClassificationAnomaly DetectionOptimizationLarge Language ModelTextBiomedical DataElectronic Health Records
🎯 What it does: Proposes the Semi-Supervised Model Evaluation (SSME) method, which evaluates the performance of multiple models using a small number of labeled samples and a large number of unlabeled samples simultaneously.