ICML 2025 Papers — Page 10
International Conference on Machine Learning · 3257 papers
EGPlace: An Efficient Macro Placement Method via Evolutionary Search with Greedy Repositioning Guided Mutation
ji deng, Jun Gao (Peking University)
OptimizationReinforcement LearningTabularBenchmark
🎯 What it does: A macro cell placement method called EGPlace is proposed, which is based on evolutionary search and greedy relocation-guided mutation for efficient macro cell layout.
Ehrenfeucht-Haussler Rank and Chain of Thought
Pablo Barcelo, Tomasz Steifer (Institute of Fundamental Technological Research)
TransformerChain-of-Thought
🎯 What it does: This paper establishes an equivalence between the Ehrenfeucht–Haussler rank of Boolean functions and the number of Chain-of-Thought steps required by a single-layer Transformer (hard attention) in generating answers, and provides precise upper bounds on the order of iterative function compositions and the k-th One function.
Eigen Analysis of Conjugate Kernel and Neural Tangent Kernel
Xiangchao Li (University of Science and Technology of China), Qing Yang (University of Science and Technology of China)
ClassificationImage
🎯 What it does: This study investigates the spectral properties of the conjugate kernel (CK) and neural tangent kernel (NTK) under random weights in deep fully connected networks, revealing how isolated eigenvalues and their eigenvectors capture the intra-group structure of input Gaussian mixture models in the high-dimensional limit.
Eigenspectrum Analysis of Neural Networks without Aspect Ratio Bias
Yuanzhe Hu (Dartmouth College), Yaoqing Yang (Dartmouth College)
OptimizationHyperparameter SearchConvolutional Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: The FARMS method is proposed and validated, utilizing fixed aspect ratio submatrix sampling and average ESD to eliminate the bias of the weight matrix aspect ratio on the analysis of significant feature spectra, and improving hierarchical hyperparameter tuning effects in various scenarios such as LLM pruning, image classification, and scientific machine learning fine-tuning.
ELEMENTAL: Interactive Learning from Demonstrations and Vision-Language Models for Reward Design in Robotics
Letian Chen (Georgia Institute of Technology), Matthew Craig Gombolay
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelMultimodality
🎯 What it does: The ELEMENTAL framework is proposed, which combines visual demonstrations with large language models to achieve interactive learning of user intentions by robots.
Eliciting Language Model Behaviors with Investigator Agents
Xiang Lisa Li (Stanford University), Jacob Steinhardt (UC Berkeley)
OptimizationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIPrompt EngineeringText
🎯 What it does: Train and use investigator agents to automatically generate prompts that can induce the target language model to exhibit specified behaviors (such as false reporting, malicious output, or biased behavior);
ELITE: Enhanced Language-Image Toxicity Evaluation for Safety
Wonjun Lee (Yonsei University), Suhyun Kim (Kyung Hee University)
GenerationSafty and PrivacyTransformerLarge Language ModelVision Language ModelImageTextBenchmark
🎯 What it does: This paper proposes the ELITE evaluator and the ELITE benchmark for accurately assessing the safety of visual language models, filtering, and generating diverse image-text pairs.
ELMO : Efficiency via Low-precision and Peak Memory Optimization in Large Output Spaces
Jinbin Zhang (Aalto University), Rohit Babbar (University of Bath)
OptimizationComputational EfficiencyLarge Language ModelTextBenchmark
🎯 What it does: Proposed the ELMO framework, which uses pure low precision (BF16/FP8) to train extremely large output space models, achieving significant reductions in memory usage and computation.
ELoRA: Low-Rank Adaptation for Equivariant GNNs
Chen Wang (Institute of Computing Technology, Chinese Academy of Sciences), Weile Jia (Institute of Computing Technology, Chinese Academy of Sciences)
Graph Neural NetworkSupervised Fine-TuningGraph
🎯 What it does: A low-rank adaptation PEFT method ELoRA for SO(3) equivariant graph neural networks is proposed and implemented, and fine-tuned on pre-trained atomic potentials.
Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers
Zhengchao Wan (University of Missouri), Yusu Wang (University of California San Diego)
Flow-based ModelStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Study the ODE sampling dynamics of flow matching (FM) models and analyze their trajectories through data geometry and denoisers.
Elucidating the design space of language models for image generation
Xuantong LIU, Yuan Yao (Hong Kong University of Science and Technology)
GenerationData SynthesisTransformerLarge Language ModelAuto EncoderImage
🎯 What it does: This study systematically explores the design space of using large language models (LLMs) for image generation, evaluating tokenizers, modeling approaches, scanning patterns, vocabulary design, and sampling strategies one by one, ultimately constructing an ELM model based on the BAE tokenizer and autoregressive Transformer.
Elucidating the Design Space of Multimodal Protein Language Models
Cheng-Yen Hsieh (ByteDance), Quanquan Gu (ByteDance)
Protein Structure PredictionMultimodalityBiomedical Data
🎯 What it does: This paper systematically studies and improves the design space of multimodal protein language models, proposing a technical path that removes structural quantization loss, incorporates word-level supervision, residual diffusion, geometric perception modules, representation alignment, and multi-chain data.
Embedding Safety into RL: A New Take on Trust Region Methods
Nikola Milosevic (Max Planck Institute for Human Cognitive and Brain Sciences), Nico Scherf (Max Planck Institute for Human Cognitive and Brain Sciences)
OptimizationSafty and PrivacyReinforcement LearningTabularBenchmark
🎯 What it does: In the framework of Constrained Markov Decision Processes (CMDP), the authors propose Constrained Trust Region Policy Optimization (C-TRPO) and Constrained Natural Policy Gradient (C-NPG), which construct a safe trust region in the policy space to ensure that the training process always meets safety constraints.
EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents
Rui Yang (University of Illinois Urbana-Champaign), Tong Zhang (Northwestern University)
Robotic IntelligenceTransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmark
🎯 What it does: Proposes the EMBODIEDBENCH benchmark for evaluating the performance of multimodal large language models in vision-driven embodied agent tasks, covering 4 environments, 1128 test instances, and 6 fine-grained subsets.
Emergence and Effectiveness of Task Vectors in In-Context Learning: An Encoder Decoder Perspective
Seungwook Han (Massachusetts Institute of Technology), Pulkit Agrawal (Massachusetts Institute of Technology)
TransformerPrompt EngineeringText
🎯 What it does: The study investigates the emergence and effects of task vectors in ICL using autoregressive Transformers, elucidating the formation and performance correlation of task vectors from an encoding-decoding perspective.
Emergence in non-neural models: grokking modular arithmetic via average gradient outer product
Neil Rohit Mallinar (University of California San Diego), Mikhail Belkin (University of California San Diego)
🎯 What it does: This paper demonstrates the 'grokking' phenomenon of modular arithmetic tasks in non-neural network models through Recursive Feature Machines (RFM) and Average Gradient Outer Product (AGOP), revealing that feature learning is the fundamental cause of this phenomenon.
Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs
Jan Betley (Truthful AI), Owain Evans (University of California Berkeley)
Safty and PrivacyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The study investigates the fine-tuning of large language models like GPT-4o on generating unsafe code, revealing widespread misalignment behavior in non-coding tasks.
Emergent Response Planning in LLMs
Zhichen Dong (Shanghai Artificial Intelligence Laboratory), Chaochao Lu (Shanghai Artificial Intelligence Laboratory)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper demonstrates that the model can implicitly encode the global attributes of a complete future response during the prompt phase by probing hidden representations, thereby achieving response planning.
Emergent Symbolic Mechanisms Support Abstract Reasoning in Large Language Models
Yukang Yang (Princeton University), Taylor Whittington Webb
TransformerLarge Language ModelText
🎯 What it does: This paper explores the mechanisms supporting abstract reasoning within large language models through a series of causal mediation, attention, and similarity analyses, revealing a three-stage symbolic architecture composed of symbolic abstraction, symbolic induction, and retrieval heads.
EmoGrowth: Incremental Multi-label Emotion Decoding with Augmented Emotional Relation Graph
Kaicheng Fu (Chinese Academy of Sciences), Huiguang He (Chinese Academy of Sciences)
ClassificationKnowledge DistillationGraph Neural NetworkAuto EncoderVideoAudio
🎯 What it does: The AESL framework is proposed for multi-label incremental emotion decoding, addressing the issues of missing labels in the past and future.
Emoji Attack: Enhancing Jailbreak Attacks Against Judge LLM Detection
Zhipeng Wei (International Computer Science Institute), N. Benjamin Erichson (Lawrence Berkeley National Laboratory)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The study investigates and verifies the vulnerability of Judge LLM under token segmentation bias and proposes an attack method utilizing emoji injection—Emoji Attack—to enhance the success rate of jailbreak evasion detection.
Emotional Face-to-Speech
Jiaxin Ye (Fudan University), Hongming Shan (Fudan University)
GenerationData SynthesisDiffusion modelMultimodalityAudio
🎯 What it does: This paper proposes the Emotion-Focused Speech (eF2S) task, which generates speech with emotion directly from facial expressions while maintaining the speaker's identity.
Empirical Privacy Variance
Yuzheng Hu (University of Illinois Urbana-Champaign), David Forsyth (University of Illinois Urbana-Champaign)
Safty and PrivacyTransformerSupervised Fine-TuningText
🎯 What it does: Under the same (ε,δ)-DP guarantee, this study investigates the empirical privacy differences caused by different DP-SGD hyperparameter configurations and quantifies these differences.
Empower Structure-Based Molecule Optimization with Gradient Guided Bayesian Flow Networks
Keyue Qiu (Tsinghua University), Wei-Ying Ma (Tsinghua University)
OptimizationDrug DiscoveryFlow-based ModelBiomedical Data
🎯 What it does: Proposed the MolJO structural basic molecular optimization framework, utilizing Bayesian Flow Networks for joint gradient guidance of continuous coordinates and discrete atom types, and introducing backward correction sampling to achieve efficient 3D molecular optimization.
Empowering World Models with Reflection for Embodied Video Prediction
Xiaowei Chi (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
GenerationRobotic IntelligenceVision Language ModelDiffusion modelWorld ModelVideoTextBenchmark
🎯 What it does: This paper proposes a Reflection-of-Generation (RoG) mechanism based on intermediate reasoning, and constructs the EVA (Embodied Video Anticipator) model and the EVA-Bench evaluation framework to enhance the capabilities of embodied video prediction and world modeling.
Enabling Optimal Decisions in Rehearsal Learning under CARE Condition
Wen-Bo Du (Nanjing University), Zhi-Hua Zhou (Nanjing University)
OptimizationReinforcement LearningTabular
🎯 What it does: Under the premise of given observational data and the Structured Reproduction Model (SRM), this paper proposes the CARE condition, which transforms the Avoiding Unfavorable Future (AUF) problem into a convex optimization problem, allowing for the direct solution of the optimal decision that maximizes the AUF probability; a projection Newton algorithm (super-linear convergence) is designed for this condition, along with a closed-form solution when |Y|=1; an embedded CARE scheme is provided for cases outside the CARE condition.
ENAHPool: The Edge-Node Attention-based Hierarchical Pooling for Graph Neural Networks
Zhehan Zhao (Beijing Normal University), Edwin Hancock (University of York)
ClassificationGraph Neural NetworkGraph
🎯 What it does: This paper proposes a hierarchical pooling method based on edge-node attention, ENAHPool, and the corresponding MD-MPNN network, aimed at enhancing the expressive power of graph classification tasks.
EncryptedLLM: Privacy-Preserving Large Language Model Inference via GPU-Accelerated Fully Homomorphic Encryption
Leo de Castro (J.P. Morgan), Manuela Veloso (J.P. Morgan)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposed a GPU-accelerated implementation of Fully Homomorphic Encryption (FHE) that supports encrypted LLM inference.
End-to-End Learning Framework for Solving Non-Markovian Optimal Control
Xiaole Zhang (University of Southern California), Paul Bogdan (University of Southern California)
OptimizationReinforcement Learning from Human FeedbackRecurrent Neural NetworkTransformerReinforcement LearningTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: Proposes an optimal control theory based on fractional-order linear time-invariant systems and implements an end-to-end learning framework.
Energy-Based Flow Matching for Generating 3D Molecular Structure
Wenyin Zhou (KTH Royal Institute of Technology), Hossein Azizpour (Science for Life Laboratory)
GenerationDrug DiscoveryProtein Structure PredictionFlow-based ModelContrastive LearningGraph
🎯 What it does: This paper proposes an energy-based flow matching model (IDFlow) that generates 3D molecular structures by learning an iterative and approximately unitary flow graph, which is used in docking and protein backbone generation tasks.
Energy-Based Preference Model Offers Better Offline Alignment than the Bradley-Terry Preference Model
Yuzhong Hong (Zuoyebang Education Technology), yang song
Recommendation SystemReinforcement Learning from Human FeedbackReinforcement LearningContrastive LearningTabular
🎯 What it does: Proposes an energy-based preference model (IPM) and its EPA loss, replacing the traditional Bradley-Terry model for offline RLHF alignment.
Enforcing Idempotency in Neural Networks
Nikolaj Banke Jensen (University of Oxford), Jamie Vicary (University of Cambridge)
GenerationOptimizationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A universal power series corrector g(K) = 3K² - 2K³ is proposed to enforce idempotency in neural networks (f(x) = f(f(x))), and based on this, a 'Modified Backpropagation' training scheme is designed.
Enforcing Latent Euclidean Geometry in Single-Cell VAEs for Manifold Interpolation
Alessandro Palma (Institute of Computational Biology), Fabian J Theis
Auto EncoderBiomedical Data
🎯 What it does: The FlatVI method is proposed, which enforces the latent space to exhibit Euclidean geometry in single-cell VAE by stretching the flattening loss, allowing linear interpolation to be closer to the geodesics on the decoded data manifold.
Enhancing Adversarial Robustness with Conformal Prediction: A Framework for Guaranteed Model Reliability
Jie Bao (Chengdu Research Institute City University of Hong Kong), Zhixin Zhou (City University of Hong Kong)
ClassificationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A framework that combines adversarial training with conformal prediction is proposed, designing the OPSA attack that maximizes the size of the prediction set without requiring coverage information, along with its corresponding OPSA-AT defense method.
Enhancing Certified Robustness via Block Reflector Orthogonal Layers and Logit Annealing Loss
Bo-Han Lai (National Taiwan University), Shang-Tse Chen (National Taiwan University)
ClassificationComputational EfficiencyConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposed the Block Reflector Orthogonal (BRO) layer and Logit Annealing Loss to construct efficient 1-Lipschitz networks, achieving provable robustness.
Enhancing Cooperative Multi-Agent Reinforcement Learning with State Modelling and Adversarial Exploration
Andreas Kontogiannis (National Technical University of Athens), George Vouros (University of Piraeus)
Reinforcement LearningAuto Encoder
🎯 What it does: A multi-agent reinforcement learning framework called SMPE 2 is proposed, which is based on state modeling and adversarial exploration to enhance cooperation in partially observable environments without communication.
Enhancing Decision-Making of Large Language Models via Actor-Critic
Heng Dong (Tsinghua University), Chongjie Zhang (Washington University in St. Louis)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextChain-of-Thought
🎯 What it does: A LLM-based Actor-Critic framework called LAC is designed, integrating LLM prior strategies with action evaluation for long-term planning to achieve efficient multi-step decision-making.
Enhancing Diversity In Parallel Agents: A Maximum State Entropy Exploration Story
Vincenzo de paola, Marcello Restelli (Politecnico di Milano)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: By introducing a state entropy maximization framework in a multi-agent parallel simulation environment, a policy gradient-based PGPSE algorithm is proposed to achieve diversified exploration under limited samples;
Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization
Luca Masserano (Carnegie Mellon University), Bernie Wang
TransformerTime Series
🎯 What it does: WaveToken is proposed, which quantizes time series data into discrete vocabulary after decomposition through wavelet transform, and trains an autoregressive Transformer for prediction.
Enhancing Foundation Models with Federated Domain Knowledge Infusion
Jiaqi Wang (Pennsylvania State University), Fenglong Ma (Pennsylvania State University)
Domain AdaptationFederated LearningKnowledge DistillationDiffusion modelImage
🎯 What it does: A cross-tunnel federated learning framework called FedAG is proposed, which uses multi-domain fine-grained adapters to inject domain knowledge into the visual foundation model CLIP and achieve cross-domain generalization.
Enhancing Graph Contrastive Learning for Protein Graphs from Perspective of Invariance
Yusong Wang (Guangdong Institute of Intelligence Science and Technology), Mingkun Xu (Guangdong Institute of Intelligence Science and Technology)
Protein Structure PredictionGraph Neural NetworkContrastive LearningGraphBiomedical Data
🎯 What it does: Using graph contrastive learning to enhance protein graph representations, a graph augmentation strategy based on Functional Community Invariance (FCI) and 3D Protein Structure Invariance (3-PSI) is proposed, integrating both augmentations into a unified GCL framework.
Enhancing Graph Invariant Learning from a Negative Inference Perspective
Kuo Yang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
Domain AdaptationAnomaly DetectionGraph Neural NetworkTransformerGraph
🎯 What it does: A negative inference graph OOD framework called NeGo is proposed, which utilizes negative prompt reasoning environmental factors and enhances causal subgraph learning to improve the generalization ability of graph learning under complex environmental shifts.
Enhancing Ligand Validity and Affinity in Structure-Based Drug Design with Multi-Reward Optimization
Seungbeom Lee (POSTECH), Dongwoo Kim (POSTECH)
OptimizationDrug DiscoveryReinforcement LearningFlow-based ModelBiomedical Data
🎯 What it does: A multi-reward optimization framework is proposed, using Direct Preference Optimization (DPO) on Bayesian flow networks to generate ligands in structure-based drug design, balancing objectives such as ligand efficacy, affinity, and drug similarity.
Enhancing Logits Distillation with Plug&Play Kendall's $\tau$ Ranking Loss
Yuchen Guan (Tsinghua University), Chun Yuan (Tsinghua University)
Object DetectionKnowledge DistillationConvolutional Neural NetworkTransformerImage
🎯 What it does: A pluggable ranking loss based on Kendall τ is proposed to supplement the KL divergence in knowledge distillation, addressing its insufficient focus on low-probability channels and the resulting suboptimal solutions.
Enhancing Parallelism in Decentralized Stochastic Convex Optimization
Ofri Eisen (Technion), Kfir Yehuda Levy
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A distributed stochastic convex optimization algorithm named Decentralized Anytime SGD (DAT-SGD) is proposed, significantly enhancing the scalability parallelism threshold of decentralized networks.
Enhancing Performance of Explainable AI Models with Constrained Concept Refinement
Geyu Liang (University of Michigan), Salar Fattahi (University of Michigan)
ClassificationOptimizationExplainability and InterpretabilityComputational EfficiencyContrastive LearningImageMultimodality
🎯 What it does: This paper proposes a Constrained Concept Refinement (CCR) method for concept embedding in interpretable AI models, which enhances predictive performance while maintaining interpretability, achieving significant computational efficiency improvements in large-scale image classification tasks.
Enhancing Rating-Based Reinforcement Learning to Effectively Leverage Feedback from Large Vision-Language Models
Tung Minh Luu, Chang D. Yoo (Korea Advanced Institute of Science and Technology)
Robotic IntelligenceTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: This paper proposes a reinforcement learning method called ERL-VLM, based on large-scale visual-language models (VLM), to generate absolute scores. It uses VLM to evaluate trajectory segments on a Likert scale, thereby learning a reward function and training an RL policy.
Enhancing Spectral GNNs: From Topology and Perturbation Perspectives
Taoyang Qin (Nanjing University of Posts and Telecommunications), Zheng Liu (Nanjing University of Posts and Telecommunications)
ClassificationGraph Neural NetworkGraph
🎯 What it does: A high-dimensional cellular Laplacian matrix (Perturbed Sheaf Laplacian, PSL) is proposed, obtained by applying small orthogonal perturbations to the block structure of the normalized graph Laplacian matrix. This matrix is embedded into various spectral GNN models (such as GCN, GPRGNN, APPNP, Graph-Heat, BernNet, Jacobi) to enhance node classification performance.
Enhancing Statistical Validity and Power in Hybrid Controlled Trials: A Randomization Inference Approach with Conformal Selective Borrowing
Ke Zhu (North Carolina State University), Xiaofei Wang (Duke University)
Biomedical Data
🎯 What it does: In a mixed control trial, a precise randomization inference framework based on Fisher's randomization test is proposed, and selective borrowing from external controls (EC) is achieved through an adaptive conformal selection method, thereby maintaining precise control of the first type error and enhancing test power in small sample environments.
Enhancing Target-unspecific Tasks through a Features Matrix
Fangming Cui (Shanghai Jiao Tong University), Jun Yu (Harbin Institute of Technology)
ClassificationDomain AdaptationTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: By constructing feature matrices generated from various handcrafted templates, the generalization ability of CLIP in target-agnostic tasks is enhanced.
Enhancing the Influence of Labels on Unlabeled Nodes in Graph Convolutional Networks
Jincheng Huang (University of Electronic Science and Technology of China), Xiaofeng Zhu (University of Electronic Science and Technology of China)
Graph Neural NetworkContrastive LearningGraph
🎯 What it does: A new two-step framework called ELU-GCN is proposed to enhance the influence of labels on unlabelled nodes in Graph Convolutional Networks (GCN). This framework optimizes the utilization of labels through graph learning and graph contrastive learning.
Enhancing Treatment Effect Estimation via Active Learning: A Counterfactual Covering Perspective
Hechuan Wen (University of Queensland), Hongzhi Yin (University of Queensland)
Tabular
🎯 What it does: Under the constraint of a limited label budget, an active learning strategy is utilized to select samples from observational data to enhance the accuracy of treatment effect estimation.
Enhancing Visual Localization with Cross-Domain Image Generation
Yuanze Wang (Shanghai Jiao Tong University), Dianxi Shi (Advanced Institute of Big Data)
GenerationData SynthesisPose EstimationDomain AdaptationGaussian SplattingImage
🎯 What it does: By constructing a cross-domain 3D Gaussian splatting model and combining image editing, anchor point generation, and positional attention, high-quality training images are automatically generated across cameras and environments, enhancing the performance of absolute pose regression visual localization.
EnIGMA: Interactive Tools Substantially Assist LM Agents in Finding Security Vulnerabilities
Talor Abramovich (Tel Aviv University), Ofir Press (Princeton Language and Intelligence)
OptimizationTransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: EnIGMA is a language model agent designed for Capture The Flag (CTF) challenges, which automates the process of finding and exploiting security vulnerabilities by integrating interactive tools (such as the gdb debugger and remote connection tools) within a Docker environment.
Ensemble Distribution Distillation via Flow Matching
Jonggeon Park (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
Domain AdaptationAnomaly DetectionKnowledge DistillationFlow-based ModelRectified FlowImageTextOrdinary Differential Equation
🎯 What it does: This paper proposes an ensemble distribution distillation method based on flow matching (EDFM), which trains a lightweight student network to learn the mapping of single model predictions to the complete prediction distribution of the ensemble teacher, achieving performance comparable to or better than the ensemble teacher while maintaining a small model size.
Ensemble Learned Bloom Filters: Two Oracles are Better than One
Ming Lin (Sun Yat-sen University), Lin CHEN
Anomaly DetectionOptimizationTabular
🎯 What it does: This paper proposes a structure of an Ensemble Learning-based Learning Bloom Filter (ELBF), which replaces a single large pre-filter with multiple small learning pre-filters equipped with corresponding backup Bloom filters, significantly reducing the overall false positive rate under a given memory budget.
EnsLoss: Stochastic Calibrated Loss Ensembles for Preventing Overfitting in Classification
Ben Dai (Chinese University of Hong Kong)
ClassificationOptimizationConvolutional Neural NetworkTabular
🎯 What it does: This paper proposes ENSLOSS, an ensemble method that combines calibration loss functions randomly to achieve the effects of multiple losses in a single training session.
ENSUR: Equitable and Statistically Unbiased Recommendation
Nitin Bisht (University of Technology Sydney), Guandong Xu (Education University of Hong Kong)
Recommendation SystemTabular
🎯 What it does: The ENSUR framework is proposed to generate minimal prediction sets with statistical confidence and fairness guarantees.
EPIC: Efficient Position-Independent Caching for Serving Large Language Models
Junhao Hu (Peking University), Tao Xie (Peking University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The EPIC system and LegoLink algorithm are proposed to achieve position-independent (PIC) caching in large language model services, significantly improving inference efficiency by reusing cached KV vectors without the need for precise prefix matching.
EpiCoder: Encompassing Diversity and Complexity in Code Generation
Yaoxiang Wang (Xiamen University), Scarlett Li (Microsoft)
GenerationData SynthesisAI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: A code synthesis framework based on feature trees is proposed, utilizing hierarchical code features to achieve controllable complexity and diversity in instruction data generation.
Epsilon-VAE: Denoising as Visual Decoding
Long Zhao (Google DeepMind), Ting Liu (Google DeepMind)
RestorationGenerationDiffusion modelRectified FlowAuto EncoderImage
🎯 What it does: Proposes Epsilon-VAE, which replaces the single-step decoder in traditional autoencoders with a diffusion-based iterative denoising decoder, achieving high-quality reconstruction and generation.
EQ-VAE: Equivariance Regularized Latent Space for Improved Generative Image Modeling
Theodoros Kouzelis (Athena Research Center), Nikos Komodakis
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: Proposes EQ-VAE, which simplifies the latent space during the training phase of the autoencoder using equivariance regularization, improving the quality and efficiency of latent generative models.
Equivalence is All: A Unified View for Self-supervised Graph Learning
Yejiang Wang (Northeastern University), Xingwei Wang
Representation LearningGraph Neural NetworkGraph
🎯 What it does: Proposes the GALE framework, which implements unsupervised graph representation learning using a unified equivalence class of automatic isomorphism and attribute isomorphism.
EquivaMap: Leveraging LLMs for Automatic Equivalence Checking of Optimization Formulations
Haotian Zhai (University of Texas at Austin), Liu Leqi (Stanford University)
OptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The EquivaMap framework is proposed, which utilizes large language models to automatically identify and verify equivalent formulations of two identical optimization problems, based on the Quasi-Karp equivalence definition.
Equivariant Neural Tangent Kernels
Philipp Misof (Chalmers University of Technology and the University of Gothenburg), Jan E Gerken
Data SynthesisDrug DiscoveryConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper derives the recursive relationship of the neural tangent kernel (NTK) for group convolutional equivariant neural networks (GCNN) and proves that in the infinite width limit, the training dynamics of data-augmented non-equivariant networks are the same as those of equivariant networks.
Equivariant Polynomial Functional Networks
Thieu Vo, Tan Minh Nguyen
ClassificationGenerationConvolutional Neural NetworkImage
🎯 What it does: A new neural functional network MAGEP-NFN is proposed, utilizing polynomial equivariant layers to achieve equivariance to both permutation symmetry and scaling/sign-flip symmetry in the weight space;
EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers
Daiheng Gao (University of Science and Technology of China), Weiming Zhang (University of Science and Technology of China)
GenerationOptimizationTransformerRectified FlowContrastive LearningImageText
🎯 What it does: This paper proposes the EraseAnything method, specifically designed for Transformer visual-text models like Flux, to delete undesirable concepts while maintaining generation quality.
Ergodic Generative Flows
Leo Maxime Brunswic (Huawei Technologies Canada), Yinchuan Li (Huawei)
GenerationReinforcement LearningFlow-based ModelTabularTime Series
🎯 What it does: A class of ergodicity-based generative flow models—Ergodic Generative Flows (EGF)—is designed to address flow matching issues in continuous space and imitation learning, and a KL-weakFM loss is proposed without an additional reward model.
ERICT: Enhancing Robustness by Identifying Concept Tokens in Zero-Shot Vision Language Models
Xinpeng Dong (Zhejiang University), Kun Kuang (Zhejiang University)
RecognitionObject DetectionTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: The ERICT method is proposed, which enhances the robustness of VLM by identifying concept tokens during the inference phase and masking the CLS attention weights.
Erwin: A Tree-based Hierarchical Transformer for Large-scale Physical Systems
Maksim Zhdanov (University of Amsterdam), Jan-Willem van de Meent (University of Amsterdam)
TransformerGraphPhysics Related
🎯 What it does: A Transformer model called Erwin, based on a spherical tree hierarchy, is proposed for efficiently handling large-scale physical systems.
ESPFormer: Doubly-Stochastic Attention with Expected Sliced Transport Plans
Ashkan Shahbazi (Vanderbilt University), Soheil Kolouri (Vanderbilt University)
ClassificationComputational EfficiencyTransformerImageTextPoint Cloud
🎯 What it does: This paper proposes a dual stochastic attention mechanism called ESPFormer based on Expected Sliced Transport Plans (ESP), which replaces the traditional Sinkhorn iteration to achieve dual stochasticity, significantly enhancing the efficiency and expressiveness of Transformers.
ETTA: Elucidating the Design Space of Text-to-Audio Models
Sang-gil Lee (NVIDIA), Bryan Catanzaro (NVIDIA)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderTextAudio
🎯 What it does: Conducted systematic large-scale experiments on the design space of text-to-audio generation models, proposed a large high-quality synthetic subtitle dataset AF-Synthetic, and built the ETTA model based on it, covering VAE, improved DiT implementation, different training objectives, and sampling strategies.
Evaluating Judges as Evaluators: The JETTS Benchmark of LLM-as-Judges as Test-Time Scaling Evaluators
Yilun Zhou (Salesforce AI Research), Shafiq Joty (Salesforce AI Research)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper constructs the JETTS benchmark to systematically evaluate the performance of LLM judges in testing scenarios involving expansion (reordering, beam search, critical improvement).
Evaluating LLMs Across Multi-Cognitive Levels: From Medical Knowledge Mastery to Scenario-Based Problem Solving
Yuxuan Zhou (Tsinghua University), Ji Wu (Tsinghua University)
Large Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: A multi-cognitive level medical assessment framework based on Bloom's taxonomy was designed and constructed, and it was used to systematically evaluate the performance of existing LLMs at different cognitive levels in the medical field.
Evaluating Neuron Explanations: A Unified Framework with Sanity Checks
Tuomas Oikarinen (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerVision Language ModelImageText
🎯 What it does: A unified neuron explanation evaluation framework, NeuronEval, is proposed, and within this framework, theoretical and experimental validation of 18 existing evaluation metrics is conducted, introducing two testing methods: missing label testing and additional label testing.
Event-Customized Image Generation
Zhen Wang (Zhejiang University), Long Chen (Hong Kong University of Science and Technology)
GenerationRetrievalDiffusion modelImage
🎯 What it does: Proposed the task of event-customized image generation and introduced the training-free method FreeEvent;
Everything Everywhere All at Once: LLMs can In-Context Learn Multiple Tasks in Superposition
Zheyang Xiong (University of Wisconsin-Madison), Dimitris Papailiopoulos (University of Wisconsin-Madison)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the phenomenon of large language models (LLMs) being able to simultaneously perform multiple different tasks through contextual examples during a single inference—referred to as 'task superposition'—and conducts systematic experiments and theoretical analyses of its performance across various models and scales.
EvFocus: Learning to Reconstruct Sharp Images from Out-of-Focus Event Streams
Lin Zhu (Beijing Institute of Technology), Hua Huang (Beijing Normal University)
RestorationRecurrent Neural NetworkTransformerImageVideo
🎯 What it does: The EvFocus framework is constructed to reconstruct sharp images from defocused event streams.
EvoControl: Multi-Frequency Bi-Level Control for High-Frequency Continuous Control
Samuel Holt (University of Cambridge), Nicolas Heess (Google DeepMind)
Robotic IntelligenceReinforcement LearningTime Series
🎯 What it does: Designed and implemented the EvoControl framework, using PPO to train a low-frequency high-level policy and Evolution Strategies to train a high-frequency low-level controller, achieving two-layer policy learning for high-frequency continuous control.
EVOLvE: Evaluating and Optimizing LLMs For In-Context Exploration
Allen Nie (Stanford University), Minmin Chen (Google DeepMind)
OptimizationKnowledge DistillationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes the BanditBench benchmark, conducting unsupervised 'self-exploration' experiments with large language models (LLMs) in multi-armed and contextual bandit environments, and investigates two methods to enhance LLM exploration capabilities: algorithm guidance and algorithm distillation.
Evolving Minds: Logic-Informed Inference from Temporal Action Patterns
Chao Yang (Chinese University of Hong Kong), Shuang Li (Chinese University of Hong Kong)
Anomaly DetectionExplainability and InterpretabilityReinforcement LearningTime SeriesSequential
🎯 What it does: A logical information-based point process framework is proposed, which, combined with a variational EM that can be diluted, is capable of jointly inferring hidden psychological event sequences and predicting future actions.
Evolving Prompts In-Context: An Open-ended, Self-replicating Perspective
Jianyu Wang (Alibaba Group), Lidong Bing (MiroMind)
GenerationOptimizationLarge Language ModelPrompt EngineeringText
🎯 What it does: A method for pruning examples in in-context learning (ICL) is proposed, discovering that randomly pruning examples to generate 'garbled' prompts can actually enhance the task performance of large language models (LLMs). An automatic search for effective pruning strategies is conducted through the self-evolving PromptQuine framework.
EvoMesh: Adaptive Physical Simulation with Hierarchical Graph Evolutions
Huayu Deng (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
Graph Neural NetworkMeshGraphPhysics Related
🎯 What it does: This study investigates physical simulations on adaptive multi-scale graph structures, proposing the EvoMesh framework that can adaptively generate time-evolving graph layers during the training process and perform physical dynamics predictions.
EvoPress: Accurate Dynamic Model Compression via Evolutionary Search
Oliver Sieberling (ETH Zurich), Dan Alistarh (IST Austria)
CompressionOptimizationLarge Language ModelText
🎯 What it does: EvoPress is proposed—a dynamic compression framework based on evolutionary search, designed to automatically find the optimal compression configuration for each layer (including layer pruning, sparsification, and quantization) while satisfying overall compression constraints.
Ex-VAD: Explainable Fine-grained Video Anomaly Detection Based on Visual-Language Models
Chao Huang (Sun Yat-Sen University), Xiaochun Cao (Sun Yat-Sen University)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelVision Language ModelVideoMultimodality
🎯 What it does: Proposes the Ex-VAD method, which utilizes visual language models (VLM) and large language models (LLM) to generate anomaly explanations and integrate multimodal features, achieving interpretable fine-grained video anomaly detection.
Exact Recovery of Sparse Binary Vectors from Generalized Linear Measurements
Arya Mazumdar (University of California San Diego), Neha Sangwan (University of California San Diego)
🎯 What it does: The research presents methods for accurately recovering sparse binary vectors from generalized linear measurements (such as logistic regression and noisy one-bit compressed sensing), providing the optimal sample complexity and proving the information-theoretic lower bound.
Exact risk curves of signSGD in High-Dimensions: quantifying preconditioning and noise-compression effects
Ke Liang Xiao (McGill University), Elliot Paquette
OptimizationTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper provides a theoretical analysis of signSGD in the high-dimensional limit, deriving its stochastic differential equation (SDE) and deterministic ordinary differential equation (ODE), and using them to describe the evolution of the risk curve.
Exact Upper and Lower Bounds for the Output Distribution of Neural Networks with Random Inputs
Andrey Kofnov (TU Wien), Efstathia Bura (TU Wien)
Tabular
🎯 What it does: A method is proposed for accurately estimating the upper and lower bounds of the cumulative distribution function (CDF) of neural network output under random inputs, which guarantees conservative error bounds across the entire support set and proves that these bounds can converge to the true CDF as resolution increases.
Exactly Tight Information-theoretic Generalization Bounds via Binary Jensen-Shannon Divergence
Yuxin Dong (Xi'an Jiaotong University), Chen Li (Xi'an Jiaotong University)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A generalized error upper bound based on binary Jensen–Shannon divergence is proposed, constructing a theoretically achievable precise and tight upper bound in random learning algorithms;
ExLM: Rethinking the Impact of $\texttt{[MASK]}$ Tokens in Masked Language Models
Kangjie Zheng (Peking University), Ming Zhang (Peking University)
TransformerLarge Language ModelTextBiomedical Data
🎯 What it does: This paper systematically studies the 'semantic corruption' problem caused by the [MASK] token in masked language models and proposes the EXLM model, which enhances the model's contextual understanding and downstream performance by performing multi-state expansion for each [MASK] token and explicitly modeling the semantic dependencies between states.
Exogenous Isomorphism for Counterfactual Identifiability
Yikang Chen (Shanghai Key Laboratory of Trustworthy Computing), Dehui du
Flow-based ModelTabularOrdinary Differential Equation
🎯 What it does: The concept of 'exogenous isomorphism' is proposed, and ~EI-identifiability is introduced to simplify the complete causal identifiability analysis of the L3 layer; identifiability conditions are provided for two types of models: Bijective SCM and Triangular Monotonic SCM, and based on this, a neural TM-SCM is constructed for counterfactual consistency inference in practice.
Expected Variational Inequalities
Brian Hu Zhang (Carnegie Mellon University), Tuomas Sandholm (Carnegie Mellon University)
Optimization
🎯 What it does: This paper proposes and analyzes the Expected Variational Inequality (EVI) as a natural relaxation of the traditional Variational Inequality (VI), proving its existence, complexity, and algorithmic feasibility under various settings.
Expert Race: A Flexible Routing Strategy for Scaling Diffusion Transformer with Mixture of Experts
Yike Yuan (ByteDance), Qiyang Min (ByteDance)
GenerationTransformerMixture of ExpertsDiffusion modelImage
🎯 What it does: We propose Race-DiT, a flexible routing strategy using Mixture of Experts (MoE) in diffusion transformers.
Explainable Concept Generation through Vision-Language Preference Learning for Understanding Neural Networks' Internal Representations
Aditya Taparia (Arizona State University), Ransalu Senanayake (Arizona State University)
GenerationExplainability and InterpretabilityReinforcement LearningDiffusion modelImageMultimodality
🎯 What it does: This paper proposes a method for automatically generating conceptual images that explain the internal representations of neural networks, viewing concept generation as an image generation problem.
Explaining the role of Intrinsic Dimensionality in Adversarial Training
Enes Altinisik (Qatar Computing Research Institute), Sanjay Chawla (Qatar Computing Research Institute)
RetrievalAdversarial AttackConvolutional Neural NetworkTransformerLarge Language ModelGenerative Adversarial NetworkImageText
🎯 What it does: This study investigates the robustness and generalization differences of adversarial training in visual models, decoder language models, and encoder language models, providing explanations through intrinsic dimensions (ID). Based on this, a scalable adversarial training method called SMAAT is proposed, which generates adversarial samples using the minimum layer of hierarchical ID, significantly reducing training costs.
Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations
Shahaf Bassan (Hebrew University of Jerusalem), Guy Katz (Hebrew University of Jerusalem)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: An algorithm based on abstraction-refinement is proposed, which can efficiently generate provably minimal sufficient explanations for neural network predictions.
Explanatory Instructions: Towards Unified Vision Tasks Understanding and Zero-shot Generalization
Yang Shen (Nanjing University of Science and Technology), Errui Ding (Baidu)
Image TranslationRestorationSegmentationGenerationDepth EstimationTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper introduces the concept of 'Explanatory Instructions', which describes the transformation between images and target images in natural language, thereby reconstructing the objectives of visual tasks. Based on this, a dataset of 12M image-instruction-output triples, DECVT, was constructed, and a self-regressive visual language model was fine-tuned on this dataset, achieving zero-shot generalization at both the instruction and task levels.
Explicit Discovery of Nonlinear Symmetries from Dynamic Data
Lexiang Hu (Peking University), Zhouchen Lin (Peking University)
Neural Architecture SearchTime SeriesSequentialPhysics RelatedOrdinary Differential Equation
🎯 What it does: The LieNLSD method is proposed, which can explicitly discover nonlinear symmetries from dynamic data and automatically determine the number and expressions of generators.
Explicit Exploration for High-Welfare Equilibria in Game-Theoretic Multiagent Reinforcement Learning
Austin A. Nguyen (University of Michigan), Michael P. Wellman (University of Michigan)
Reinforcement LearningAgentic AI
🎯 What it does: Proposes Ex2PSRO, which incorporates an exploration strategy based on high welfare trajectories and KL regularization into the best response training of PSRO, encouraging the search for equilibria with higher welfare.
Explicit Preference Optimization: No Need for an Implicit Reward Model
Xiangkun Hu (Amazon Web Services), David Wipf (Amazon Web Services)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: This paper proposes an explicit preference optimization framework named EXPO, which replaces traditional implicit reward methods such as DPO and IPO to address the complexities and limitations of RLHF.