ICLR 2025 Papers — Page 11
International Conference on Learning Representations · 3704 papers
EDiT: A Local-SGD-Based Efficient Distributed Training Method for Large Language Models
Jialiang Cheng (Ant Group), Jian Sha (Ant Group)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: For the distributed training of large-scale language models, two efficient distributed training methods, EDIt and A-EDIt, are proposed, integrating Local SGD with model sharding, and introducing techniques such as hierarchical synchronization, prefetching, and pseudo-gradient penalties to enhance efficiency and stability.
EditRoom: LLM-parameterized Graph Diffusion for Composable 3D Room Layout Editing
Kaizhi Zheng (University of California Santa Cruz), Xin Eric Wang (Microsoft)
GenerationData SynthesisGraph Neural NetworkTransformerLarge Language ModelDiffusion modelMultimodalityGraph
🎯 What it does: Proposes the EditRoom framework, which utilizes natural language instructions to achieve automated editing of 3D room layouts;
Effective and Efficient Time-Varying Counterfactual Prediction with State-Space Models
Haotian Wang (National University of Defense Technology), Wenjing Yang (National University of Defense Technology)
OptimizationTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: Mamba-CDSP is proposed, a time series causal prediction framework based on state space models, specifically designed for time-varying counterfactual prediction (TCP).
Effective Interplay between Sparsity and Quantization: From Theory to Practice
Simla Burcu Harma (EcoCloud), Amir Yazdanbakhsh (Google)
CompressionConvolutional Neural NetworkLarge Language ModelImageText
🎯 What it does: This study investigates the interaction between sparsity and quantization in large model compression, proving that the two are non-orthogonal and providing the optimal compression order;
Effective post-training embedding compression via temperature control in contrastive training
Georgiana Dinu (Amazon), Xing Niu (Amazon)
RetrievalCompressionContrastive LearningText
🎯 What it does: Investigate the impact of the temperature parameter in contrastive loss on text embeddings, and propose multi-temperature aggregation and temperature specialization training strategies to maintain high retrieval performance after compression.
Efficient Action-Constrained Reinforcement Learning via Acceptance-Rejection Method and Augmented MDPs
Wei Hung (National Yang Ming Chiao Tung University), Ping-Chun Hsieh (National Taiwan University)
Robotic IntelligenceReinforcement LearningAgentic AISequential
🎯 What it does: This paper proposes a framework called ARAM, which utilizes an acceptance-rejection method and enhanced MDP to achieve action-constrained learning based on unconstrained RL algorithms.
Efficient Active Imitation Learning with Random Network Distillation
Emilien Biré (Centrale Supelec), Rémy Portelas (Ubisoft La Forge)
Knowledge DistillationRobotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningAgentic AISequential
🎯 What it does: A method for active imitation learning based on Random Network Distillation (RND), called RND-DAgger, is proposed to reduce expert intervention and improve learning efficiency.
Efficient Alternating Minimization with Applications to Weighted Low Rank Approximation
Zhao Song (Simons Institute for the Theory of Computing UC Berkeley), Lichen Zhang (Massachusetts Institute of Technology)
OptimizationComputational Efficiency
🎯 What it does: This paper studies the problem of weighted low-rank approximation and proposes an alternating minimization framework that is approximately updatable, robust, and faster in runtime.
Efficient and Accurate Explanation Estimation with Distribution Compression
Hubert Baniecki (University of Warsaw), Przemyslaw Biecek (University of Warsaw)
Explainability and InterpretabilityComputational EfficiencyTabular
🎯 What it does: The study proposes the 'Compress Then Explain' (CTE) paradigm, which uses distribution compression to replace traditional i.i.d. sampling to improve the accuracy and efficiency of post-hoc interpretability methods.
Efficient and Context-Aware Label Propagation for Zero-/Few-Shot Training-Free Adaptation of Vision-Language Model
Yushu Li (South China University of Technology), Xun Xu (Institute for Infocomm Research A*STAR)
ClassificationDomain AdaptationComputational EfficiencyGraph Neural NetworkVision Language ModelImageMultimodality
🎯 What it does: A graph-based label propagation method is proposed for zero-shot/few-shot training adaptation of visual-language models, aimed at improving label efficiency and inference efficiency.
Efficient and Robust Neural Combinatorial Optimization via Wasserstein-Based Coresets
Xu Wang (University of Science and Technology of China), Yan Xiong (University of Science and Technology of China)
OptimizationGraph
🎯 What it does: This paper proposes a rigidity transformation invariance (RWD) metric based on Wasserstein distance and utilizes coreset technology to compress large-scale combinatorial optimization datasets, achieving efficiency and robustness in neural combinatorial optimization (NCO) training.
Efficient and Trustworthy Causal Discovery with Latent Variables and Complex Relations
Xiu-Chuan Li (Sydney AI Centre, University of Sydney), Tongliang Liu (Sydney AI Centre, University of Sydney)
Anomaly DetectionComputational EfficiencyGraph
🎯 What it does: This paper proposes an efficient and verifiable causal discovery algorithm based on the causal Markov assumption, which can identify latent variables and the complete causal structure in polynomial time, even in the presence of latent variables and complex causal relationships.
Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition
Aliyah R. Hsu (University of California Berkeley), Bin Yu (University of California Berkeley)
TransformerText
🎯 What it does: A context decomposition method suitable for Transformers (CD-T) is proposed, which enables automated circuit discovery and can quickly identify computational subgraphs responsible for specific tasks in the model without the need for training or manual examples.
Efficient Biological Data Acquisition through Inference Set Design
Ihor Neporozhnii (University of Toronto), Jason Hartford (Valence Labs)
Drug DiscoveryImageBiomedical Data
🎯 What it does: This study proposes a method for efficiently acquiring biological data through inference set design, aiming to achieve system target accuracy by selecting the smallest candidate set, thereby reducing experimental costs in drug discovery.
Efficient Causal Decision Making with One-sided Feedback
Jianing Chu (Amazon), PULAK GHOSH
Recommendation SystemOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackTabularFinance Related
🎯 What it does: A new value function is proposed for the decision-making problem with one-sided feedback, along with its identifiability, efficiency bounds, and efficient estimation methods.
Efficient Cross-Episode Meta-RL
Gresa Shala (University of Freiburg), Josif Grabocka (University of Technology Nuremberg)
Meta LearningTransformerReinforcement LearningSequential
🎯 What it does: This paper proposes an Efficient Cross-Episode Transformers (ECET) model for online meta reinforcement learning, which achieves rapid task adaptation by utilizing cross-episode and intra-episode experiences.
Efficient Dictionary Learning with Switch Sparse Autoencoders
Anish Mudide (Massachusetts Institute of Technology), Christian Schroeder de Witt (University of Oxford)
OptimizationComputational EfficiencyMixture of ExpertsAuto EncoderText
🎯 What it does: A Switch Sparse Autoencoder (Switch SAE) is proposed, which splits the training of sparse autoencoders into multiple small experts through a Mixture of Experts architecture, significantly reducing FLOPs and memory overhead while maintaining interpretable features.
Efficient Diffusion Transformer Policies with Mixture of Expert Denoisers for Multitask Learning
Moritz Reuss (Karlsruhe Institute of Technology), Rudolf Lioutikov (Massachusetts Institute of Technology)
Computational EfficiencyRobotic IntelligenceTransformerMixture of ExpertsDiffusion modelMultimodalityBenchmark
🎯 What it does: A diffusion strategy based on a mixture of experts, MoDE, is proposed, utilizing noise-conditioned routing to achieve multi-task learning.
Efficient Discovery of Pareto Front for Multi-Objective Reinforcement Learning
Ruohong Liu (University of Oxford), Jiang Bian (Microsoft Research Asia)
OptimizationReinforcement Learning
🎯 What it does: A two-stage C-MORL algorithm is proposed, which initializes training of multiple fixed preference policies using Pareto, and then extends the Pareto front and achieves policy allocation through constrained optimization.
Efficient Distribution Matching of Representations via Noise-Injected Deep InfoMax
Ivan Butakov (Moscow Institute of Physics and Technology), Alexey Frolov (Skolkovo Institute of Science and Technology)
ClassificationRepresentation LearningGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Under the Deep InfoMax framework, a lightweight method is proposed for automatic distribution matching (Gaussian or uniform distribution) by injecting independent noise at the normalized encoder output while maintaining the original InfoMax objective, without the need for additional networks.
Efficient Diversity-Preserving Diffusion Alignment via Gradient-Informed GFlowNets
Zhen Liu (Mila, Université de Montréal), Dinghuai Zhang (Mila, Université de Montréal)
GenerationOptimizationReinforcement LearningDiffusion modelImage
🎯 What it does: A gradient information-based GFlowNet method (∇-GFlowNet) is proposed for reward function alignment and fine-tuning of pre-trained diffusion models, capable of rapid convergence while maintaining diversity and prior distribution.
Efficient Evolutionary Search Over Chemical Space with Large Language Models
Haorui Wang (Georgia Institute of Technology), Chao Zhang (Georgia Institute of Technology)
OptimizationDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringTextBiomedical Data
🎯 What it does: MOLLEO is proposed, a framework that integrates large language models (LLMs) into evolutionary algorithms for generating compounds that meet multiple attribute objectives.
Efficient Exploration and Discriminative World Model Learning with an Object-Centric Abstraction
Anthony GX-Chen (New York University), Rob Fergus (New York University)
Robotic IntelligenceReinforcement LearningWorld ModelSequential
🎯 What it does: The study implements efficient exploration and world model learning in reinforcement learning using object-based abstraction (Ab-MDP), and proposes MEAD, a fully modeled, discriminative forward model + counting reward + MCTS exploration and Dijkstra planning method;
Efficient Imitation under Misspecification
Nicolas Espinosa-Dice (Cornell University), Gokul Swamy (Carnegie Mellon University)
OptimizationRobotic IntelligenceReinforcement LearningSequential
🎯 What it does: This paper studies the problem of imitation learning under incorrect specification and proposes an efficient inverse reinforcement learning algorithm aimed at reducing interaction with the environment while ensuring the learning of a strong policy.
Efficient Inference for Large Language Model-based Generative Recommendation
Xinyu Lin (National University of Singapore), Tat-Seng Chua (National University of Singapore)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A Speculative Decoding (SD) acceleration framework called AtSpeed is proposed for LLM generative recommendation, which designs two strategies: strict Top-K verification and relaxed sampling verification, and achieves efficient draft-then-verify inference based on this.
Efficient Interpolation between Extragradient and Proximal Methods for Weak MVIs
Thomas Pethick (Ecole Polytechnique Federale de Lausanne), Volkan Cevher (Ecole Polytechnique Federale de Lausanne)
Optimization
🎯 What it does: This study investigates non-monotonic games that satisfy weak Minty variational inequalities and proposes a fault-correcting approximate origin algorithm, achieving an O(1/ε) convergence rate for the first time across the entire range.
EFFICIENT JAILBREAK ATTACK SEQUENCES ON LARGE LANGUAGE MODELS VIA MULTI-ARMED BANDIT-BASED CONTEXT SWITCHING
Aditya Ramesh (Fujitsu Research of India Private Limited), Manohar Kaul (Fujitsu Research of India Private Limited)
OptimizationAdversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper studies an attack method implemented through a Sequence of Context (SoC) that gradually weakens the security barriers of large language models (LLMs) using a series of carefully designed context-switching queries (CSQs), ultimately achieving a jailbreak.
Efficient Learning with Sine-Activated Low-Rank Matrices
Yiping Ji (Australian Institute for Machine Learning), Simon Lucey (Australian Institute for Machine Learning)
CompressionOptimizationTransformerNeural Radiance FieldImage
🎯 What it does: Introducing sine activation to low-rank matrices to enhance their effective rank, thereby improving model accuracy while maintaining parameter compression.
Efficient Low-Bit Quantization with Adaptive Scales for Multi-Task Co-Training
Boyu Liu (Beihang University), Baochang Zhang (Beihang University)
ClassificationRestorationSuper ResolutionKnowledge DistillationImageMultimodality
🎯 What it does: Aiming at low-bit quantization for co-training models, the TSQ-MTC framework is proposed.
Efficient Masked AutoEncoder for Video Object Counting and A Large-Scale Benchmark
Bing Cao (Tianjin University), Qinghua Hu (Tianjin University)
Object DetectionObject TrackingTransformerAuto EncoderOptical FlowVideoBenchmark
🎯 What it does: An efficient mask autoencoder E-MAC based on density embedding is proposed for video object counting, along with the design of a Spatial Adaptive Mask (SAM) and Temporal Co-Fusion (TCF) module.
Efficient Model Editing with Task-Localized Sparse Fine-tuning
Leonardo Iurada (Politecnico di Torino), Tatiana Tommasi (Vector Institute)
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningImageText
🎯 What it does: A sparse fine-tuning method based on minimal sensitive parameters, called TaLoS, is proposed for model editing in task arithmetic.
Efficient Model-Based Reinforcement Learning Through Optimistic Thompson Sampling
Jasmine Bayrooti (University of Cambridge), Amanda Prorok (University of Cambridge)
Robotic IntelligenceReinforcement Learning
🎯 What it does: A HOT-GP algorithm based on joint Gaussian Process dynamics and reward models is proposed, achieving optimistic exploration in model-based reinforcement learning, thereby significantly improving sampling efficiency.
Efficient Multi-agent Offline Coordination via Diffusion-based Trajectory Stitching
Lei Yuan (Nanjing University), Yang Yu (Nanjing University)
Robotic IntelligenceReinforcement LearningDiffusion modelSequentialBenchmark
🎯 What it does: Utilizing diffusion models to generate high-quality multi-agent trajectory segments, enhancing offline data through head-tail concatenation, thereby improving the learning efficiency of offline multi-agent reinforcement learning.
Efficient Neuron Segmentation in Electron Microscopy by Affinity-Guided Queries
Hang Chen (Tsinghua University), Xiaolin Hu (Tsinghua University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: A query-based neuron segmentation method called AGQ is proposed, which directly predicts segmentation results on 3D electron microscopy images, eliminating the traditional watershed and clustering steps.
Efficient Off-Policy Learning for High-Dimensional Action Spaces
Fabian Otto (Microsoft Research), Gerhard Neumann (Karlsruhe Institute of Technology)
Reinforcement LearningSequential
🎯 What it does: A novel offline policy gradient algorithm called Vlearn is designed, achieving efficient offline reinforcement learning in high-dimensional action spaces using only the state value function.
Efficient Online Pruning and Abstraction for Imperfect Information Extensive-Form Games
Boning Li (Tsinghua University), Longbo Huang (Tsinghua University)
OptimizationComputational EfficiencyReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: The EVPA framework is proposed, aiming to achieve efficient pruning and information abstraction through online expected value estimation, in order to quickly solve approximate Nash equilibria in large incomplete information games.
Efficient Online Reinforcement Learning Fine-Tuning Need Not Retain Offline Data
Zhiyuan Zhou (University of California Berkeley), Aviral Kumar (Carnegie Mellon University)
Reinforcement LearningSequential
🎯 What it does: This study investigates an RL fine-tuning method that does not retain offline data, proposing WSRL for rapid fine-tuning through warmup and online RL.
Efficient Perplexity Bound and Ratio Matching in Discrete Diffusion Language Models
Etrit Haxholli (MetaDialog Research), Eli Waxman (MetaDialog Research)
GenerationOptimizationComputational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: This paper proposes a new KL divergence theorem and a more compact perplexity upper bound in the discrete diffusion model, and replaces ratio matching with cross-entropy training to achieve more efficient language modeling.
Efficient Policy Evaluation with Safety Constraint for Reinforcement Learning
Claire Chen (University of Virginia), Shangtong Zhang (University of Virginia)
OptimizationSafty and PrivacyReinforcement LearningSequential
🎯 What it does: An optimal variance minimization behavior policy is proposed under safety constraints for offline evaluation of the target policy's performance.
Efficient Reinforcement Learning with Large Language Model Priors
Xue Yan (Institute of Automation, Chinese Academy of Science), Jun Wang (University College London)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: The study treats large language models (LLMs) as action prior distributions, implemented through variational inference and posterior sampling within a reinforcement learning framework, significantly enhancing the sampling efficiency for online and offline text decision tasks.
Efficient Residual Learning with Mixture-of-Experts for Universal Dexterous Grasping
Ziye Huang (Peking University), Zongqing Lu (Peking University)
Robotic IntelligenceReinforcement LearningMixture of ExpertsImage
🎯 What it does: We propose ResDex, a framework that combines residual learning and mixture of experts (MoE) for efficient and scalable universal grasping on thousands of objects.
Efficient Source-Free Time-Series Adaptation via Parameter Subspace Disentanglement
Gaurav Patel (Purdue University), Juri Minxha (Apple)
Domain AdaptationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningTime Series
🎯 What it does: This paper proposes a framework for source-free domain adaptation (SFDA) in time series, which significantly improves parameter and sample efficiency by reparameterizing weights using low-rank Tucker decomposition during the source model preparation phase and only fine-tuning the core tensor at the target end.
Efficient Sparse PCA via Block-Diagonalization
Alberto Del Pia (University of Wisconsin Madison), Yinglun Zhu (University of California Riverside)
OptimizationComputational EfficiencyTextBiomedical Data
🎯 What it does: A framework based on matrix block diagonalization is proposed, utilizing existing Sparse PCA algorithms to accelerate the solution.
Efficient stagewise pretraining via progressive subnetworks
Abhishek Panigrahi (Princeton University), Sanjiv Kumar (Google)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a staged subnetwork training framework (RAPTR) that improves the computational efficiency of large model pre-training by randomly training sub-networks at each step and gradually increasing the size of the sub-networks.
Efficient Top-m Data Values Identification for Data Selection
Xiaoqiang Lin (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationComputational EfficiencyImage
🎯 What it does: This paper proposes the GPGapE algorithm for efficiently identifying the top m Shapley values in a dataset, further accelerating the process by calculating marginal contributions only on a small subset.
Efficient Training of Neural Stochastic Differential Equations by Matching Finite Dimensional Distributions
Jianxin Zhang (University of Michigan), Emily Pitler (Cisco Systems)
GenerationData SynthesisOptimizationComputational EfficiencyTime SeriesFinance RelatedStochastic Differential Equation
🎯 What it does: A training framework based on finite-dimensional distribution matching (FDM) is proposed for the efficient training of Neural Stochastic Differential Equations (Neural SDE).
Efficiently Democratizing Medical LLMs for 50 Languages via a Mixture of Language Family Experts
Guorui Zheng (Chinese University of Hong Kong), Benyou Wang (Chinese University of Hong Kong)
Large Language ModelMixture of ExpertsTextBiomedical Data
🎯 What it does: This paper constructs a high-quality medical dataset covering 12 high-resource languages and implements a large medical language model for 50 languages using a Mixture of Experts (MoE) architecture;
Efficiently Learning at Test-Time: Active Fine-Tuning of LLMs
Jonas Hübotter (ETH Zurich), Andreas Krause (ETH Zurich)
Computational EfficiencyData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented the SIFT algorithm for efficient data selection and fine-tuning of large language models during testing.
Efficiently Parameterized Neural Metriplectic Systems
Anthony Gruber (Sandia National Laboratories), Nathaniel Trask (University of Pennsylvania)
Physics RelatedOrdinary Differential Equation
🎯 What it does: A framework for learning finite-dimensional energy-entropy hyperbolic systems using neural networks (Neural Metriplectic Systems) is proposed.
EffoVPR: Effective Foundation Model Utilization for Visual Place Recognition
Issar Tzachor (OriginAI), Rami Ben-Ari (OriginAI)
RecognitionRetrievalTransformerContrastive LearningImage
🎯 What it does: The EffoVPR method is proposed, utilizing the self-attention layers within DINOv2 for feature pooling and reordering, supporting both single-stage and double-stage retrieval, and implementing both zero-shot and fine-tuning usage modes.
EG4D: Explicit Generation of 4D Object without Score Distillation
Qi Sun (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
GenerationData SynthesisDiffusion modelGaussian SplattingImageVideo
🎯 What it does: Generating 4D dynamic objects from a single image, first using a video diffusion model (SVD + SV3D) to generate temporally and spatially consistent videos without training, followed by rough reconstruction using 4D Gaussian Splatting, and refining it once using image-to-image diffusion.
EgoExo-Gen: Ego-centric Video Prediction by Watching Exo-centric Videos
Jilan Xu (Fudan University), Weidi Xie (Shanghai Jiao Tong University)
GenerationData SynthesisDiffusion modelVideoText
🎯 What it does: Under the conditions of synchronized third-person video, first-frame first-person video, and textual descriptions, a cross-perspective video prediction method (EgoExo-Gen) is proposed, which first predicts the hand-object interaction (HOI) mask in the future first-person perspective, and then uses the mask as a structural guide input to the video diffusion model to generate subsequent first-person video frames.
EgoSim: Egocentric Exploration in Virtual Worlds with Multi-modal Conditioning
Wei Yu (University of Toronto), Animesh Garg (Columbia University)
GenerationData SynthesisDiffusion modelVideoTextMultimodality
🎯 What it does: EgoSim is proposed, utilizing camera pose, contextual frames, and text as three conditions to achieve controllable generation of first-person perspective videos.
EIA: ENVIRONMENTAL INJECTION ATTACK ON GENERALIST WEB AGENTS FOR PRIVACY LEAKAGE
Zeyi Liao (Ohio State University), Huan Sun (Ohio State University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelAgentic AITextMultimodality
🎯 What it does: This study investigates the privacy leakage risks of general web proxies in a malicious website environment and proposes an Environment Injection Attack (EIA) scheme.
ElasticTok: Adaptive Tokenization for Image and Video
Wilson Yan (University of California Berkeley), Hao Liu (Google DeepMind)
GenerationCompressionTransformerVision Language ModelAuto EncoderImageVideo
🎯 What it does: ElasticTok, an adaptive visual tokenizer, has been developed, which dynamically allocates the number of tokens during encoding based on previous frame information using random masks, achieving variable-length encoding for images and long videos.
ELBOing Stein: Variational Bayes with Stein Mixture Inference
Ola Rønning (University of Copenhagen), Thomas Hamelryck (University of Copenhagen)
Mixture of ExpertsTabularSequential
🎯 What it does: A new particle variational Bayesian method called Stein Mixture Inference (SMI) is proposed, which approximates the posterior by treating each particle as a component of a mixture model, avoiding the variance collapse problem of traditional SVGD.
ELFS: Label-Free Coreset Selection with Proxy Training Dynamics
Haizhong Zheng (University of Michigan), Atul Prakash (University of Michigan)
ClassificationData-Centric LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A label-free unsupervised core sample selection method called ELFS is proposed, which can select a high-quality subset of training samples without relying on true labels.
ELICIT: LLM Augmentation Via External In-context Capability
Futing Wang (Westlake University), Tao Lin (Westlake University)
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A framework named ELICIT has been developed, which enhances the multi-task adaptability of large language models by externally storing task vectors and dynamically retrieving them during inference, without increasing the number of additional tokens.
Eliciting Human Preferences with Language Models
Belinda Z. Li (Massachusetts Institute of Technology), Jacob Andreas (Massachusetts Institute of Technology)
Recommendation SystemTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an interactive task requirement mining framework called GATE, which actively guides users to express their preferences by generating free-form questions and examples, and uses the obtained information to train personalized prediction models.
Eliminating Oversaturation and Artifacts of High Guidance Scales in Diffusion Models
Seyedmorteza Sadat (ETH Zurich), Romann M. Weber (Disney Research Studios)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: The paper proposes a new guidance method called Adaptive Projected Guidance (APG), which generates high-quality images in diffusion models with a higher guidance scale while significantly reducing saturation and artifacts.
Eliminating Position Bias of Language Models: A Mechanistic Approach
Ziqi Wang (University of Illinois Urbana-Champaign), Heng Ji (University of Illinois Urbana-Champaign)
TransformerLarge Language ModelText
🎯 What it does: A training-agnostic, zero-shot PINE method is proposed, utilizing cross-document bidirectional attention and position reallocation to eliminate the positional information bias of language models.
Elliptic Loss Regularization
Ali Hasan (Morgan Stanley), Vahid Tarokh (Duke University)
ClassificationOptimizationImageBiomedical DataStochastic Differential Equation
🎯 What it does: An elliptical loss regularization method is proposed, which controls the smoothness of the loss over the input space by ensuring that the loss function satisfies an elliptical partial differential equation.
Elucidating the Preconditioning in Consistency Distillation
Kaiwen Zheng (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationOptimizationKnowledge DistillationImageOrdinary Differential Equation
🎯 What it does: This paper studies a new preprocessing method called Analytic-Precond, designed for preprocessing in consistency distillation to accelerate the training process of multi-step generation.
EmbedLLM: Learning Compact Representations of Large Language Models
Richard Zhuang (University of California), Kannan Ramchandran (University of California)
OptimizationRepresentation LearningTransformerLarge Language ModelText
🎯 What it does: Proposes the EmbedLLM framework, which learns compact vector representations of LLMs for multi-model management and downstream tasks.
EmbodiedSAM: Online Segment Any 3D Thing in Real Time
Xiuwei Xu (Tsinghua University), Jiwen Lu (Tsinghua University)
Object DetectionSegmentationTransformerContrastive LearningImagePoint Cloud
🎯 What it does: This paper proposes an online real-time 3D instance segmentation framework called ESAM, which utilizes 2D masks generated by the 2D vision foundation model SAM. It converts these masks into 3D queries through a geometry-aware query enhancement module, refines the 3D masks using a dual-layer query decoder, and achieves efficient query matching and merging through auxiliary tasks.
Emergence of a High-Dimensional Abstraction Phase in Language Transformers
Emily Cheng (Universitat Pompeu Fabra), Marco Baroni (SISSA)
TransformerLarge Language ModelText
🎯 What it does: The study investigates the trajectory of the intrinsic dimensionality of representations at various layers of the Transformer language model as the number of layers changes, and finds a commonly occurring high-dimensional peak corresponding to a key stage in abstract language processing.
Emergence of meta-stable clustering in mean-field transformer models
Giuseppe Bruno (University of Bern), Andrea Agazzi (University of Bern)
Transformer
🎯 What it does: This paper rigorously analyzes the continuous-time mean-field PDE of token propagation across layers in a simplified Transformer model under large-scale token counts, revealing the dynamic evolution process from uniform distribution to meta-stable phase and finally to a single clustering.
Emergent Orientation Maps —— Mechanisms, Coding Efficiency and Robustness
Haixin Zhong (Fudan University), yuguo yu
Spiking Neural NetworkImage
🎯 What it does: A self-evolving synaptic plasticity spontaneous synaptic neural network (SESNN) was constructed, and through training with natural images, the effects of retinal-V1 visual overlap, neuron density, and connection range on the formation of cortical orientation preference maps were explored.
Emerging Safety Attack and Defense in Federated Instruction Tuning of Large Language Models
Rui Ye (Shanghai Jiao Tong University), Siheng Chen (Shanghai Jiao Tong University)
Federated LearningSafty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the security vulnerabilities in Federated Instruction Tuning (FedIT). It first proposes a covert security attack method based on misaligned data and, on this basis, designs a post-processing defense scheme that fine-tunes the global model using automatically generated secure aligned data to repair the security deviations caused by the attack.
EMMA: Empowering Multi-modal Mamba with Structural and Hierarchical Alignment
Yifei Xing (Institute of Computing Technology Chinese Academy of Sciences), Yaowei Wang (Pengcheng Laboratory)
Image TranslationRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: By pixel-level alignment and multi-scale feature fusion, the quality of Mamba-based multimodal LLM visual representations is improved.
EMOS: Embodiment-aware Heterogeneous Multi-robot Operating System with LLM Agents
Junting Chen (National University of Singapore), Lin Shao (National University of Singapore)
Robotic IntelligenceTransformerLarge Language ModelSimultaneous Localization and MappingMultimodalityBenchmark
🎯 What it does: This paper proposes EMOS, a framework for heterogeneous multi-robot system collaboration utilizing large language models, and achieves perception and task planning through the automatic generation of 'robot resumes'.
Empowering LLM Agents with Zero-Shot Optimal Decision-Making through Q-learning
Jiajun Chai (Institution of Automation, Chinese Academy of Sciences), Yuanheng Zhu (Institution of Automation, Chinese Academy of Sciences)
OptimizationTransformerLarge Language ModelReinforcement LearningWorld ModelTabular
🎯 What it does: A model-based LLM agent called MLAQ, which combines LLM and reinforcement learning, has been designed and implemented to achieve optimal decision-making with zero or minimal interactions.
Empowering Users in Digital Privacy Management through Interactive LLM-Based Agents
BOLUN SUN, Haiyun Jiang (Fudan University)
Safty and PrivacyTransformerLarge Language ModelAgentic AIText
🎯 What it does: This paper proposes an interactive LLM agent based on GPT-4o-mini to help users understand website privacy policies.
Enabling Realtime Reinforcement Learning at Scale with Staggered Asynchronous Inference
Matthew Riemer (Mila), Irina Rish (Mila)
Reinforcement LearningVideo
🎯 What it does: This study investigates the feasibility of large models in real-time reinforcement learning (RL), analyzes the reward decomposition of asynchronous interaction and learning, proposes algorithms for time-sharing inference and learning, and validates their effectiveness in real-time gaming environments.
Encryption-Friendly LLM Architecture
Donghwan Rho (Seoul National University), Jung Hee Cheon (Seoul National University)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a homomorphic encryption-friendly Transformer architecture and achieves efficient personalized fine-tuning and inference in an encrypted environment.
End-to-end Learning of Gaussian Mixture Priors for Diffusion Sampler
Denis Blessing (Karlsruhe Institute of Technology), Gerhard Neumann (FZI Research Center for Information Technology)
Diffusion modelTabular
🎯 What it does: This study investigates end-to-end learning of Gaussian mixture priors in diffusion models to enhance sampling performance.
Endless Jailbreaks with Bijection Learning
Brian R.Y. Huang (Haize Labs), Leonard Tang (Haize Labs)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A black-box general scalable jailbreak attack method based on bijective learning is proposed, which can bypass the security mechanisms of LLMs by randomly generating controllable complexity encodings.
Endowing Visual Reprogramming with Adversarial Robustness
Shengjie Zhou (Chongqing University), Lei Feng (Idealism Technology)
Domain AdaptationAdversarial AttackImage
🎯 What it does: This study investigates the robustness of Visual Reprogramming (VR) under adversarial attacks and proposes a strategy to enhance robustness by using pre-trained models from adversarial training and incorporating adversarial samples during the reprogramming process.
Energy-based Backdoor Defense Against Federated Graph Learning
Guancheng Wan (Wuhan University), Mang Ye (Wuhan University)
Federated LearningGraph Neural NetworkGraph
🎯 What it does: Designed and implemented the FedTGE framework, which utilizes an energy-based model to enable energy awareness in local graph neural networks for defense against backdoor attacks in federated graph learning.
Energy-Based Diffusion Language Models for Text Generation
Minkai Xu (Stanford University), Arash Vahdat (NVIDIA)
GenerationDiffusion modelText
🎯 What it does: Proposes EDLM, which combines energy models with discrete diffusion models to improve parallel text generation.
Energy-Weighted Flow Matching for Offline Reinforcement Learning
Shiyuan Zhang (Tsinghua University), Quanquan Gu (University of California)
OptimizationReinforcement LearningDiffusion modelFlow-based ModelTabularBenchmark
🎯 What it does: Proposes an energy-weighted flow matching and diffusion model that directly learns an energy-guided generative model and applies it to Q-weighted iterative policy optimization in offline reinforcement learning.
Enhance Multi-View Classification Through Multi-Scale Alignment and Expanded Boundary
Yuena Lin (Beijing University of Technology), Zhen Yang (Beijing University of Technology)
ClassificationAuto EncoderContrastive LearningImageVideo
🎯 What it does: For the multi-view classification task, this paper proposes a framework called MAMC that simultaneously addresses the issues of feature heterogeneity and information redundancy.
Enhanced Diffusion Sampling via Extrapolation with Multiple ODE Solutions
Jinyoung Choi (Seoul National University), Bohyung Han (Seoul National University)
GenerationData SynthesisDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A Richardson extrapolation-based ODE sampling method RX-DPM is proposed, which can significantly improve the sampling quality of diffusion models without increasing the number of function evaluations.
Enhancing Clustered Federated Learning: Integration of Strategies and Improved Methodologies
Yongxin Guo (Chinese University of Hong Kong), Tao Lin (Westlake University)
Federated LearningImage
🎯 What it does: Proposes the HCFL four-layer framework and the improved HCFL+, unifying and enhancing clustering federated learning methods to address data heterogeneity.
Enhancing Cognition and Explainability of Multimodal Foundation Models with Self-Synthesized Data
Yucheng Shi (University of Georgia), Ninghao Liu (University of Georgia)
ClassificationExplainability and InterpretabilitySupervised Fine-TuningContrastive LearningImageMultimodalityBiomedical Data
🎯 What it does: This paper proposes a visual fine-grained classification framework for self-synthesized interpretable answers, enhancing the recognition and explanation capabilities of large models through multi-round self-sampling and fine-tuning.
Enhancing Compositional Text-to-Image Generation with Reliable Random Seeds
Shuangqi Li (Stony Brook University), Mathieu Salzmann (École Polytechnique Fédérale de Lausanne)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: The study found that the initial random seed has a significant impact on the generation of text-to-image combinations, proposing to mine reliable seeds and use them to generate self-supervised data for fine-tuning, thereby significantly improving the accuracy of quantity and spatial relationship generation.
Enhancing Document Understanding with Group Position Embedding: A Novel Approach to Incorporate Layout Information
Yuke Zhu (MYbank), Sheng Guo (MYbank)
TransformerLarge Language ModelTextMultimodalityBenchmark
🎯 What it does: This paper proposes Group Position Embedding (GPE), which allocates different dimensions of positional information by group in multi-head attention, allowing large language models to learn document layout without changing the network structure or input format.
Enhancing End-to-End Autonomous Driving with Latent World Model
Yingyan Li (Institute of Automation, Chinese Academy of Sciences), Tieniu Tan (Institute of Automation, Chinese Academy of Sciences)
Autonomous DrivingOptimizationComputational EfficiencyTransformerWorld ModelImageVideo
🎯 What it does: A self-supervised latent world model (LAW) is proposed, which predicts future latent features based on the current scene's latent features and the ego vehicle's trajectory, and jointly optimizes scene representation and trajectory prediction for end-to-end driving.
Enhancing Federated Domain Adaptation with Multi-Domain Prototype-Based Federated Fine-Tuning
Jingyuan Zhang (Nanyang Technological University), Wei Yang Bryan Lim (Nanyang Technological University)
Domain AdaptationFederated LearningKnowledge DistillationAdversarial AttackImage
🎯 What it does: A multi-domain prototype-based federated fine-tuning framework (MPFT) is proposed, which addresses the issue of federated domain adaptation caused by the failure of average aggregation methods like FedAvg by using multi-domain prototypes generated by clients on the server.
Enhancing Graph Of Thought: Enhancing Prompts with LLM Rationales and Dynamic Temperature Control
SungUk Shin, Youngjoon Kim (Korea University)
Graph Neural NetworkLarge Language ModelPrompt EngineeringTextGraph
🎯 What it does: EGoT is a graph-structured prompt engineering framework that automates the generation of answers, evaluation of answers, and aggregation of reasoning, using LLMs to complete answering, scoring, and reasoning at each node, while dynamically adjusting the temperature across the entire graph.
Enhancing Language Model Agents using Diversity of Thoughts
Vijay Lingam (Amazon), Anoop Deoras (Amazon)
OptimizationAI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: A new framework is proposed to enhance the performance of language model agents through diversified reflection and cross-task memory—Diversity of Thoughts (DoT);
Enhancing Learning with Label Differential Privacy by Vector Approximation
Puning Zhao (Zhejiang Lab), Qingming Li (Zhejiang University)
OptimizationSafty and PrivacyConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Proposes a vector approximation-based local differential privacy learning method;
Enhancing Pre-trained Representation Classifiability can Boost its Interpretability
Shufan Shen (Key Lab of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences), Shuhui Wang (Huawei)
ClassificationExplainability and InterpretabilitySupervised Fine-TuningImage
🎯 What it does: This paper proposes a metric for measuring the inherent interpretability of pre-trained visual model representations—Inherent Interpretability Score (IIS)—and investigates the relationship between interpretability of representations and classification performance.
Enhancing Prediction Performance through Influence Measure
Shuguang Yu (Shanghai University of Finance and Economics), Fan Zhou (Shanghai University of Finance and Economics)
ClassificationComputational EfficiencyData-Centric LearningConvolutional Neural NetworkImageTabular
🎯 What it does: A new local influence metric (FI) is proposed to assess the impact of training samples on the performance of the validation set, and based on this, data pruning and active learning are implemented.
Enhancing Robust Fairness via Confusional Spectral Regularization
Gaojie Jin (Chinese Academy of Sciences), Ronghui Mu (University of Exeter)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: This study proposes a new regularization technique that enhances the robust fairness of deep neural networks through spectral regularization, particularly addressing the issue of robust accuracy discrepancies across different classes.
Enhancing the Scalability and Applicability of Kohn-Sham Hamiltonians for Molecular Systems
Yunyang Li (Yale University), Mark Gerstein
Drug DiscoveryMixture of ExpertsTabularPhysics Related
🎯 What it does: A large-scale molecular Hamiltonian prediction framework has been constructed, proposing the Wavefunction Alignment Loss (WALoss) and designing the WANet model based on eSCN, Mixture-of-Experts, and Many-Body Interaction;
Enhancing Uncertainty Estimation and Interpretability with Bayesian Non-negative Decision Layer
Xinyue Hu (Xidian University), Mingyuan Zhou (University of Texas at Austin)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkTransformerImage
🎯 What it does: A Bayesian Non-negative Decision Layer (BNDL) is proposed, introducing a sparse, non-negative probabilistic generative model in the final layer of deep networks to enhance uncertainty estimation and interpretability.
Enhancing Zeroth-order Fine-tuning for Language Models with Low-rank Structures
Yiming Chen (Peking University), Zaiwen Wen (Peking University)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a low-rank zero-order tuning method called LOZO, which performs zero-order gradient estimation based on the low-rank gradient structure of LLMs to reduce memory consumption and improve tuning effectiveness.
Ensembles of Low-Rank Expert Adapters
Yinghao Li (Amazon Web Service), MohamadAli Torkamani (Amazon Web Service)
Large Language ModelMixture of ExpertsText
🎯 What it does: The ELREA framework is proposed, which generates a low-rank expert adapter ensemble through gradient direction clustering to address gradient conflicts in large model fine-tuning.
Ensembling Diffusion Models via Adaptive Feature Aggregation
Cong Wang (Nanjing University), Jun Zhang (Nanjing University)
GenerationData SynthesisConvolutional Neural NetworkTransformerDiffusion modelImage
🎯 What it does: Proposes Adaptive Feature Aggregation (AFA), which dynamically fuses features in a multi-model UNet structure, adjusting the contributions of each model based on prompts, noise, steps, and spatial locations to enhance generation quality and contextual consistency.