ICML 2026 Papers — Page 18
International Conference on Machine Learning · 6554 papers
Effective Distillation to Hybrid xLSTM Architectures
Lukas Hauzenberger (Johannes Kepler University), Sepp Hochreiter (Johannes Kepler University)
Computational EfficiencyKnowledge DistillationRepresentation LearningRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsContrastive LearningTextBenchmark
🎯 What it does: This paper proposes an effective distillation pipeline that replaces the quadratic attention mechanism of large-scale language models with an xLSTM structure combining mLSTM and sliding window attention (SWA), achieving an efficient linearized model;
Effective Model Pruning : Measuring the Redundancy of Model Components
Yixuan Wang (University of Florida), Warren Dixon
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: Proposed a general sparsification threshold — effective sample number (N_eff) — to directly determine the number of components to retain from any importance score distribution.
Effective MoE-based LLM Compression by Exploiting Heterogeneous Inter-Group Experts Routing Frequency and Information Density
Zhendong Mi (Stevens Institute of Technology), Shaoyi Huang (Stevens Institute of Technology)
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelMixture of ExpertsText
🎯 What it does: Dynamically allocate the SVD rank using routing frequency and information density, and combine sparse projection to recover residual information, achieving efficient compression of Mixture-of-Experts large models.
Effective Reasoning Chains Reduce Intrinsic Dimensionality
Archiki Prasad (UNC Chapel Hill), Peter Shaw (Google DeepMind)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper investigates the impact of chain-of-thought (CoT) strategies on the reasoning performance of large language models, using the intrinsic dimensionality of tasks as a quantitative metric to evaluate the effectiveness of different reasoning strategies.
Effects of Structural Reward Shaping on Biophysical Properties in RL-Trained Plasmid Generators
McClain Thiel (University College London), Chris P Barnes
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical Data
🎯 What it does: Studied the application of structure-based reward shaping in reinforcement learning (GRPO) for generating complete plasmid sequences that meet quality control standards on PlasmidGPT, and compared it with supervised fine-tuning and baseline models.
EffGen: Enabling Small Language Models as Capable Autonomous Agents
Gaurav Srivastava (Virginia Tech), Xuan Wang (Virginia Tech)
Autonomous DrivingOptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the EFFGEN framework, enabling small language models (SLM) to efficiently and securely perform multi-tool calling, task decomposition, and memory management, constructing a locally deployable agent system.
Efficient Adaptive Testing via Gradient Path Matching Subset Selection for AI Education
Yan Zhuang (Nanjing University of Aeronautics and Astronautics), Daoqiang Zhang (Nanjing University of Aeronautics and Astronautics)
Recommendation SystemOptimizationComputational EfficiencySpiking Neural NetworkTransformerReinforcement LearningContrastive LearningTextTabularBenchmark
🎯 What it does: Proposes the Gradient Path Matching (GPM) framework, which utilizes gradient path matching for subset selection of questions, thereby achieving more accurate ability estimation with fewer questions in AI education adaptive assessment.
Efficient and Minimax Optimal In-context Nonparametric Regression with Transformers
Michelle Ching (University of Cambridge), Richard J. Samworth (University of Cambridge)
OptimizationComputational EfficiencyRepresentation LearningTransformerAuto EncoderContrastive LearningTabular
🎯 What it does: This paper studies the use of pre-trained Transformers for non-parametric regression within a context learning framework, and proves that it can achieve the optimal asymptotic error rate on Holder-smooth regression functions.
Efficient and Safe Molecular Assembly via Reinforcement Learning and Constraint Solving
Stefan Pranger (Graz University of Technology), Bettina Könighofer
OptimizationDrug DiscoveryReinforcement LearningGraphTabular
🎯 What it does: Propose an AI-based molecular assembly planning framework and a high-fidelity simulation environment called NANOASSEMBLYGYM, which can compute collision-free assignments and scheduling, and train reinforcement learning (RL) agents to perform STM operations, completing the structural assembly of up to 420 molecules.
Efficient and Uncertainty-Aware Diffusion Framework for Offline-to-Online Reinforcement Learning
Ha Manh Bui (Johns Hopkins University), Anqi Liu (Johns Hopkins University)
Computational EfficiencyKnowledge DistillationTransformerReinforcement LearningDiffusion modelTabularTime SeriesSequential
🎯 What it does: The DUAL framework distills an offline-trained diffusion planner into an efficient diffusion actor and introduces uncertainty assessment during the online phase to balance exploration and exploitation.
Efficient Bayesian Inference from Noisy Pairwise Comparisons
Till Aczel (ETH Zurich), Roger Wattenhofer (ETH Zurich)
Recommendation SystemExplainability and InterpretabilityComputational EfficiencyMixture of ExpertsTextTabularBenchmark
🎯 What it does: This paper proposes a Bayesian Bradley-Terry model (BBQ), which can simultaneously estimate the quality of the generative model and the reliability of the evaluators, and achieve efficient inference through the EM algorithm.
Efficient Bilevel Optimization for CKA-Guided MoE Upcycling
Zhiyuan Yu (Sun Yat-sen University), Li Shen (Sun Yat-sen University)
OptimizationComputational EfficiencyKnowledge DistillationNeural Architecture SearchLarge Language ModelReinforcement LearningMixture of ExpertsContrastive LearningTextBenchmark
🎯 What it does: Dynamically and selectively expand Mixture-of-Experts (MoE) networks through double optimization and CKA-guided differentiable NAS, achieving efficient continual learning and parameter reuse.
Efficient Code Analysis via Graph Representation Learning-Guided Large Language Models
Hang Gao (Key Laboratory of System Software, Institute of Software, Chinese Academy of Sciences), Jian Zhang (Key Laboratory of System Software, Institute of Software, Chinese Academy of Sciences)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringContrastive LearningTextGraph
🎯 What it does: Propose a malicious code detection framework called GMLLM that combines graph representation learning with large language models (LLMs). It uses graph neural networks (GNNs) to identify key code snippets and then guides the LLM to perform in-depth analysis on them.
Efficient Continuous-Depth Modeling with GRU Equivalents
Ayan Banerjee (Arizona State University), Sandeep Gupta (Arizona State University)
Computational EfficiencyRepresentation LearningRecurrent Neural NetworkAuto EncoderVideoTextMultimodalityStochastic Differential EquationOrdinary Differential EquationAudio
🎯 What it does: Propose the CoDA framework, replacing the nonlinear ODE layers in continuous deep networks (such as NODE, LTC-NN) with GRU + linear ODE + dense layers, thereby achieving acceleration without loss of accuracy.
Efficient Diffusion LLMs via Temporal-Spatial Parallel Decoding and Confidence Extrapolation
Zekai Li (Advanced Micro Devices, Inc.), Emad Barsoum (Advanced Micro Devices, Inc.)
GenerationComputational EfficiencyAI Code AssistantRecurrent Neural NetworkTransformerLarge Language ModelMixture of ExpertsDiffusion modelAuto EncoderContrastive LearningText
🎯 What it does: Propose a method to accelerate the inference of diffusion-based large language models by using temporal-spatial parallel decoding and confidence extrapolation, significantly reducing unnecessary denoising steps while maintaining text quality.
Efficient Diffusion Models under Nonconvex Equality and Inequality constraints via Landing
Kijung Jeon (Georgia Institute of Technology), Molei Tao (Georgia Institute of Technology)
GenerationOptimizationComputational EfficiencyDiffusion modelScore-based ModelPoint CloudMeshTabularTime SeriesSequentialBiomedical DataPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose a constraint diffusion model based on Landing, which can generate high-quality samples that satisfy constraints under arbitrary non-convex equality and inequality constraints, unifying overdamped and underdamped dynamics;
Efficient Diffusion Models via Time Step Optimization with Consistent Training and Inference Constraints
Binrui Wu (Alibaba International Digital Commerce Group), WEIZHONG ZHANG
GenerationOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerReinforcement LearningDiffusion modelScore-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes a time-step optimization framework based on policy gradients, which reparameterizes the sampling schedule using a probabilistic mask and simultaneously trains the model and time steps through bi-level optimization, significantly improving the sampling efficiency and generation quality of diffusion models.
Efficient Distributed MLLM Training with Cornstarch
Insu Jang (University of Michigan), Mosharaf Chowdhury (University of Michigan)
OptimizationComputational EfficiencyTransformerLarge Language ModelMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodalityAudio
🎯 What it does: This paper proposes the Cornstarch framework, specifically designed to achieve efficient distributed training for multi-modal large language models (MLLM), addressing performance bottlenecks caused by frozen parameters, non-causal attention patterns, and model/data heterogeneity.
Efficient Distributionally Robust Assortment Optimization in MNL Bandits
Yunfan Zhang (New York University), Zhengyuan Zhou (New York University)
Recommendation SystemOptimizationTabularBenchmark
🎯 What it does: Studies the distributionally robust product assortment optimization (DRAO) problem under the context of multinomial logit (MNL) choice models, aiming to maximize revenue in the worst-case scenario.
Efficient DP-SGD for LLMs with Randomized Clipping
Enayat Ullah (Meta Platforms Inc), Meisam Razaviyayn (University of Southern California)
Safty and PrivacyComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes a DP-SGD variant based on random cropping called DPSGD-RC, aimed at achieving differential privacy in the training of large language models.
Efficient Equivariant High-Order Crystal Tensor Prediction via Cartesian Local-Environment Many-Body Coupling
Dian Jin (Hong Kong Polytechnic University), Xiaoming Tao (Hong Kong Polytechnic University)
Computational EfficiencyDrug DiscoveryGraph Neural NetworkTransformerAuto EncoderContrastive LearningGraphTabularPhysics Related
🎯 What it does: Proposed a new equivariant higher-order crystal tensor prediction framework, CEITNet, which can directly and end-to-end predict second- to fourth-order tensors (dielectric, piezoelectric, elastic tensors) from crystal structures.
Efficient Generative Modeling beyond Memoryless Diffusion via Adjoint Schrödinger Bridge Matching
Jeongwoo Shin (Seoul National University), Jaemoo Choi (Seoul National University)
GenerationComputational EfficiencyDiffusion modelScore-based ModelImageStochastic Differential Equation
🎯 What it does: Propose a two-stage generative framework called Adjoint Schrödinger Bridge Matching (ASBM), which first treats the forward dynamics of the Schrödinger Bridge as data-to-energy sampling, learning the optimal coupling, and then uses bridge matching to supervise the reverse dynamics;
Efficient Hallucination Detection for LLMs Using Uncertainty-Aware Attention Heads
Artem Vazhentsev (Mohamed bin Zayed University of Artificial Intelligence), Artem Shelmanov (Mohamed bin Zayed University of Artificial Intelligence)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose an unsupervised, single-forward-pass hallucination detection framework named RAUQ, which identifies factual errors in LLM-generated text by leveraging the change in attention head focus on the previous token from the Transformer.
Efficient Inference for Noisy LLM-as-a-Judge Evaluation
Yiqun T. Chen (Johns Hopkins University), Shengyi Li (Johns Hopkins University)
Explainability and InterpretabilityComputational EfficiencyData-Centric LearningTransformerLarge Language ModelTextTabular
🎯 What it does: Studied and compared methods for correcting average metrics under the LLM-as-a-judge setting, using measurement error correction (Rogan-Gladen, MLE) and prediction-driven inference (PPI, PPI++, EIF), and provided theoretical efficiency analysis and empirical validation.
Efficient Learned Image Compression without Entropy Coding
Hao Cao (Tsinghua University), Jungong Han (Tsinghua University)
CompressionAuto EncoderImage
🎯 What it does: This paper proposes an entropy-free learning-based image compression framework called EF-LIC, which achieves efficient and low-latency image compression on GPUs using unconstrained vector quantization and context autoregressive transformations.
Efficient Learning of Deep State Space Models via Importance Smoothing
John-Joseph Brady (King's College London), Yunpeng Li (King's College London)
GenerationData SynthesisOptimizationComputational EfficiencyReinforcement Learning from Human FeedbackDiffusion modelScore-based ModelAuto EncoderContrastive LearningTabularTime SeriesSequentialFinance RelatedPhysics RelatedStochastic Differential Equation
🎯 What it does: Proposed and implemented a parallel differentiable particle smoothing method called PVMC for efficiently training deep state space models, avoiding the serial limitations of traditional particle filters.
Efficient LLM Moderation with Multi-Layer Latent Prototypes
Maciej Chrabaszcz (Nask National Research Institute), Sebastian Cygert
Safty and PrivacyComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: Proposes a lightweight input moderation method called Multi-Layer Prototype Moderator (MLPM), which uses multi-layer internal representations and prototype distances to determine whether prompts may lead to unsafe outputs.
Efficient Mismatch-Tolerant Coding for Model-Driven Compression
Aviv Adler (Analog Devices Inc.), Jennifer Tang (College of the Holy Cross)
CompressionLarge Language ModelText
🎯 What it does: Propose an improved PMATIC encoding algorithm to achieve robust decoding in model-driven compression, addressing probability mismatch caused by model non-determinism.
Efficient Multi-Agent Reasoning via Confidence-Guided Adaptive Debate
Seungdong Yoa (LG AI Research), Junhyun Lee (Hankuk University of Foreign Studies)
Explainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelAgentic AIMixture of ExpertsTextBenchmarkChain-of-Thought
🎯 What it does: Proposes the LASE framework, based on a leader-follower structure and confidence-guided adaptive debate, triggering multi-agent interaction only when the model is uncertain.
Efficient Multi-modal Dataset Distillation via Analytic Parameter Matching
Deyu Bo (National University Of Singapore), Xinchao Wang (National University Of Singapore)
Data SynthesisComputational EfficiencyKnowledge DistillationRepresentation LearningAuto EncoderContrastive LearningImageTextMultimodalityAudio
🎯 What it does: This paper proposes a multi-modal dataset distillation method based on analytical parameter matching.
Efficient Multi-round LLM Inference over Disaggregated Serving
Wenhao He (Shanghai Jiao Tong University), Fangcheng Fu (Shanghai Jiao Tong University)
Computational EfficiencyTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This study focuses on the efficient service of multi-round large language model inference in a prefill-decode decoupled environment, proposing the AMPD framework.
Efficient Neural Controlled Differential Equations via Attentive Kernel Smoothing
Egor Serov (Applied Ai Institute), Alexey Zaytsev (Applied Ai Institute)
ClassificationComputational EfficiencyTransformerTime SeriesStochastic Differential Equation
🎯 What it does: Propose a control trajectory construction method based on kernel smoothing and Gaussian processes, combined with a multi-view attention mechanism, significantly reducing the numerical solving cost of Neural CDE
Efficient numeracy in language models through single-token number embeddings
Linus Kreitner (Technical University of Munich), Martin J. Menten (Technical University of Munich)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Propose a numerical encoding for individual tokens called BitTokens, enabling language models to perform arithmetic operations in a single forward pass, thereby improving the model's numerical reasoning efficiency and accuracy.
Efficient Online Influence Maximization under the Independent Cascade Model with Node-Level Feedback
Arpit Agarwal (Indian Institute of Technology Bombay), Rohan Ghuge (University of Texas at Austin)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes an online influence maximization algorithm that achieves sublinear regret using only node-level feedback under the independent cascade model.
Efficient Online Variational Estimation via Monte Carlo Sampling
Mathis Chagneux (Institut Polytechnique de Paris), Jimmy Olsson (KTH Royal Institute of Technology)
OptimizationComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackRecurrent Neural NetworkDiffusion modelScore-based ModelAuto EncoderContrastive LearningTabularTime SeriesSequentialStochastic Differential Equation
🎯 What it does: Proposed an online variational inference framework based on Monte Carlo importance sampling (RMCVI), which can simultaneously learn state space model parameters and posterior distributions without requiring a regression step.
Efficient Parallel Samplers for Recurrent-Depth Models and Their Connection to Diffusion Language Models
Jonas Geiping (ELLIS Institute Tübingen & Max-Planck Institute for Intelligent Systems), Guinan Su (ELLIS Institute Tübingen & Max-Planck Institute for Intelligent Systems)
GenerationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelDiffusion modelScore-based ModelText
🎯 What it does: Propose a parallel sampler based on diffusion forcing, which can achieve parallel generation on recurrent-depth Transformers and significantly improve inference speed.
Efficient Prediction of SO(3)-Equivariant Hamiltonian Matrices via SO(2) Local Frames
Haiyang Yu (Texas A&M University), Shuiwang Ji (Texas A&M University)
Computational EfficiencyDrug DiscoveryGraph Neural NetworkContrastive LearningGraphTabularBenchmarkPhysics Related
🎯 What it does: Propose the QHNetV2 network, which achieves global equivariance under SO(3) symmetry through the SO(2) local frame, and predicts the quantum Hamiltonian matrix, thereby significantly accelerating the computation.
Efficient Preference Poisoning Attack on Offline RLHF
Chenye Yang (University of California Davis), Lifeng Lai (University of California Davis)
OptimizationAdversarial AttackReinforcement Learning from Human FeedbackReinforcement LearningPrompt EngineeringContrastive LearningTextChain-of-Thought
🎯 What it does: This paper studies label flipping attacks in offline RLHF (Direct Preference Optimization), provides theoretical analysis and proposes two efficient attack algorithms;
Efficient privacy loss accounting for subsampling and random allocation
Vitaly Feldman (Apple), Moshe Shenfeld (Hebrew university of Jerusalem)
Safty and PrivacyComputational EfficiencyContrastive LearningGaussian SplattingImageTabular
🎯 What it does: Proposed an efficient privacy loss counting (PLD) method for random allocation (k-out-of-t sampling) and Poisson sampling, along with implementation details and complexity analysis.
Efficient Public Verification of Private ML via Regularization
Zoë Ruha Bell (University of California Berkeley), Shafi Goldwasser (University of California Berkeley)
OptimizationSafty and PrivacyContrastive LearningTabular
🎯 What it does: A new differential privacy (DP) algorithm was designed that can verify the DP guarantees of a model at a cost lower than training, achieving near-optimal privacy-utility trade-offs in DP stochastic convex optimization (DP-SCO).
Efficient Reasoning with Hidden Thinking
Xuan Shen (Zhejiang University), Jiuxiang Gu (Adobe)
Explainability and InterpretabilityComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextMultimodalityBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose the Heima framework, which compresses Chain-of-Thought into sparse thinking tokens, achieving efficient inference in multimodal large language models.
Efficient RL Training for LLMs with Experience Replay
Charles Arnal (Meta), Rémi Munos (Meta)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper investigates the feasibility and benefits of introducing experience replay (replay buffer) in the fine-tuning of large language models (LLMs) with reinforcement learning, and proposes how to effectively use replay within an asynchronous training framework to reduce inference computational load;
Efficient Skill Grounding via Code Refactoring with Small Language Models
Sera Choi (Sungkyunkwan University), Honguk Woo (Sungkyunkwan University)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelPrompt EngineeringDiffusion modelTextRetrieval-Augmented Generation
🎯 What it does: Proposes the RECENT framework, which utilizes a small language model to achieve efficient embodiment and immediate adaptation of skills across different robots and environments through code refactoring, supporting long-term control tasks.
Efficient Stochastic Optimisation via Sequential Monte Carlo
James Cuin (Imperial College London), O. Deniz Akyildiz (Sorbonne Université)
OptimizationScore-based ModelImageTabularTime SeriesStochastic Differential Equation
🎯 What it does: This paper proposes a general Sequential Monte Carlo (SMC) framework to efficiently estimate gradients in optimization problems where gradients are not analytically available, and uses these gradients to drive parameter updates, significantly reducing the computational cost of traditional MCMC loops.
Efficient Synthetic Network Generation via Latent Embedding Reconstruction
Feifan Jiang (University of Michigan), Ji Zhu (University of Michigan)
GenerationData SynthesisComputational EfficiencyRepresentation LearningGraph Neural NetworkTransformerDiffusion modelScore-based ModelAuto EncoderContrastive LearningGraph
🎯 What it does: This paper proposes a two-stage generative framework named SyNGLER, which first learns low-dimensional node embeddings using a latent space model, and then trains a distribution-agnostic generator in this embedding space, ultimately generating synthetic networks that are structurally similar to the original network.
Efficient Tail-Aware Generative Optimization via Flow Model Fine-Tuning
Zifan Wang (Kth Royal Institute Of Technology), Karl Henrik Johansson
GenerationOptimizationDrug DiscoveryReinforcement LearningDiffusion modelScore-based ModelFlow-based ModelImageTextGraphTabular
🎯 What it does: Propose the TFFT (Tail-aware Flow Fine-Tuning) algorithm, which utilizes the variational duality of CVaR to perform tail-aware fine-tuning of pre-trained flow/diffusion models, aiming to pursue high-reward samples (right CVaR) while suppressing low-reward failures (left CVaR).
Efficient Test-time Inference for Generative Planning Models with OCL Search
Robert Gieselmann (Amazon), Jeremy L Wyatt
OptimizationReinforcement Learning from Human FeedbackTransformerTabularSequential
🎯 What it does: This paper proposes OCLGEN, a search algorithm that utilizes generative models and heuristic models for efficient reasoning under the OCL framework, significantly improving plan quality during testing.
Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models
Jinbin Bai (National University of Singapore), Ming-Hsuan Yang (UC Merced)
Computational EfficiencyAI Code AssistantTransformerPrompt EngineeringDiffusion modelTextRetrieval-Augmented Generation
🎯 What it does: Propose the PRISM framework to achieve efficient test-time scaling (Test-Time Scaling) for discrete diffusion language models (dLLMs)
Efficient Training of Boltzmann Generators Using Off-Policy Log-Dispersion Regularization
Henrik Schopmans (Karlsruhe Institute of Technology), Pascal Friederich (Karlsruhe Institute of Technology)
GenerationData SynthesisOptimizationComputational EfficiencyScore-based ModelFlow-based ModelTabularBiomedical DataPhysics Related
🎯 What it does: Proposed and verified an offline (off-policy) log-dispersion regularization (LDR) framework to improve the data efficiency and final performance of Boltzmann Generators in unbiased, biased, and purely variational training scenarios.
Efficient Training-Free Multi-Token Prediction via Embedding-Space Probing
Raghavv Goel (Qualcomm AI Research), Christopher Lott (Qualcomm AI Research)
GenerationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the ESP (Embedding-Space Probing) method, which uses mask tokens generated instantaneously in the embedding space for untrained multi-token prediction, thereby achieving parallel future token generation and validation.
Efficient Transformer Attention for SNNs via Hadamard Simplification
Tingting Jiang (Dalian University of Technology), Qi Xu (Dalian University of Technology)
ClassificationComputational EfficiencyConvolutional Neural NetworkSpiking Neural NetworkTransformerImage
🎯 What it does: Studied the implementation of Transformer attention in spiking neural networks and proposed two hardware-friendly attention mechanisms based on Hadamard simplification, named SSA and USSA.
Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speed
Yonggan Fu (NVIDIA), Pavlo Molchanov (NVIDIA)
Computational EfficiencyAI Code AssistantTransformerLarge Language ModelDiffusion modelScore-based ModelText
🎯 What it does: Continuously pretrain a pretrained autoregressive language model into an efficient diffusion language model, and propose block-level attention and position-related masking strategies.
Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion
Pengcheng Jiang (Google Research), Craig Boutilier (Google Research)
Data SynthesisRetrievalComputational EfficiencyReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringDiffusion modelScore-based ModelImageTextMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Proposes the R4T framework, which uses RL as a one-time target converter to generate synthetic training data suitable for fan output retrieval, and trains a lightweight diffusion retriever to achieve efficient set-based retrieval.
Efficient, Validation-Free Intrinsic Quality Estimation for Large-Scale Face Recognition Datasets
Zhichao Chen (DeepGlint), Ziyong Feng (DeepGlint)
RecognitionRepresentation LearningConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Proposed an unverified Intrinsic Quality (IQ) metric to estimate the trainability of large-scale face recognition datasets
Efficiently Learning Drifting Halfspaces with Massart Noise
Mingchen Ma (University of WisconsinMadison), Ilias Diakonikolas (University of WisconsinMadison)
ClassificationOptimizationComputational Efficiency
🎯 What it does: This paper studies online learning under drifting concepts, with the goal of learning a linear separator (halfspace) in the presence of Massart noise (labels are randomly flipped with a noise rate ≤ η), and proposes an efficient algorithm that provides error-controllable prediction models at each time step.
Efficiently Solving Discounted MDPs via Predictions with Unknown Prediction Errors
Lixing Lyu (National University of Singapore), Wang Chi Cheung (National University of Singapore)
OptimizationReinforcement LearningContrastive Learning
🎯 What it does: In infinite-horizon discounted Markov decision processes, an unknown-precision transition matrix prediction is introduced, and a corresponding learning algorithm is constructed.
Efficiently Training Time-to-First-Spike Spiking Neural Networks from Scratch
Kaiwei Che (Peking University), Yonghong Tian (Peking University)
ClassificationComputational EfficiencySpiking Neural NetworkSupervised Fine-TuningImage
🎯 What it does: Propose a complete framework for efficiently training time-to-first-spike (TTFS) spiking neural networks from scratch
EGG: An Expert-Guided Agent Framework for Kernel Generation
Yaochen Han (Beihang University), Yixiang Zhang (Beihang University)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmark
🎯 What it does: Propose an expert-guided multi-agent framework called EGG for automatically generating high-performance GPU kernels, and divide the generation process into two major stages: algorithm structure design and hardware-specific optimization;
Ego3S: Select, Strengthen, and Synchronize for Efficient Egocentric Reasoning
Shenshen Li (University of Electronic Science and Technology of China), Fumin Shen (University of Electronic Science and Technology of China)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelContrastive LearningVideoText
🎯 What it does: Propose the three-stage Ego3S framework to enhance the reasoning ability of LVLMs in first-person perspective scenarios.
EgoTactile: Learning Grasp Pressure for Everyday Objects from Egocentric Video
Yuan Zeng (Tsinghua University), Qingmin Liao (Tsinghua University)
Robotic IntelligenceTransformerDiffusion modelAuto EncoderContrastive LearningVideoMultimodalityBenchmark
🎯 What it does: Construct the EgoTactile benchmark and propose the EgoPressureDiff method, which uses a conditional diffusion framework to predict full-hand grasping pressure in first-person videos
EigenCache: Rethinking Diffusion Acceleration as Covariance-Optimal Forecasting and Submodular Information Allocation
Chenyang Xu (Xidian University), Hao Wang (Xidian University)
GenerationComputational EfficiencyMixture of ExpertsDiffusion modelContrastive LearningGaussian SplattingImageVideoText
🎯 What it does: Proposes a training-agnostic EigenCache framework that accelerates diffusion models through covariance-adaptive Kriging prediction and information-gain submodular scheduling for caching-prediction.
Eigenvectors of Experts are Training-free Non-collapsing Routers
Giang Do (Deakin University), Truyen Tran (Deakin University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelMixture of ExpertsContrastive LearningImageTextMultimodality
🎯 What it does: This study investigates the expert collapse problem in Sparse Mixture of Experts (SMoE), proposing to generate an untrained router SSMoE using the singular value decomposition (SVD) of the expert weight matrix, and designing an expert dropping strategy based on this to save GPU memory.
Ekka: Automated Diagnosis of Silent Errors in LLM Inference
Yile Gu (University of Washington), Baris Kasikci (University of Washington)
Anomaly DetectionExplainability and InterpretabilityComputational EfficiencyAI Code AssistantTransformerLarge Language ModelAgentic AIPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Propose EKKA, an automated method for diagnosing silent errors in LLM inference;
Elastic Attention: Test-time Adaptive Sparsity Ratios for Efficient Transformers
Zecheng Tang (Soochow University), Min Zhang (Soochow University)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Introduce a lightweight Attention Router into pre-trained large language models, enabling the model to dynamically assign full attention (FA) or sparse attention (SA) to each attention head based on the input during inference, thus achieving a variable sparse ratio.
Elastic Diffusion Transformer
Jiangshan Wang (CUHK), Chunchao Guo (Tencent)
GenerationComputational EfficiencyTransformerMixture of ExpertsDiffusion modelRectified FlowImagePoint CloudMesh
🎯 What it does: Propose Elastic Diffusion Transformer (E-DiT), which dynamically decides whether to skip a block and the width of the block's MLP by embedding a lightweight router in each DiT block, thereby achieving adaptive acceleration of the generation process.
ElicitR: Unlocking Latent Reasoning in Dense Retrievers via Generative Regularization
Fengyu Cai (Technical University of Darmstadt), Heinz Koeppl (Technical University of Darmstadt)
RetrievalTransformerLarge Language ModelGenerative Adversarial NetworkContrastive LearningText
🎯 What it does: Propose the ElicitR framework, which introduces generative regularization into contrastive learning to uncover the inherent reasoning capabilities of dense retrievers and improve reasoning retrieval performance.
Elign: Equivariant Diffusion Model Alignment from Foundational Machine Learned Force Fields
Yunyang Li (Yale University), Mark Gerstein (Yale University)
Drug DiscoveryGraph Neural NetworkTransformerReinforcement LearningDiffusion modelGraphTabular
🎯 What it does: Post-training E(3)-equivariant diffusion models using fundamental machine learning force fields (MLFF) to align the model to lower energy, more stable molecular conformations.
Eliminating Solution Bias in Differentially Private Optimization
DONGRUN LI, Zhiming Zheng (Beihang University)
OptimizationFederated LearningSafty and PrivacyContrastive LearningGaussian SplattingImageTextStochastic Differential Equation
🎯 What it does: Propose a new differential privacy stochastic optimization framework, DP-C4, which utilizes coupled clipping and threshold shifting to ensure that noise variance and clipping bias progressively vanish with iterations, thereby eliminating the solution bias of traditional DP algorithms.
Ellipsoidal Time Series Forecasting
Qilin Wang (Independent Researcher)
OptimizationExplainability and InterpretabilityFlow-based ModelContrastive LearningGaussian SplattingTime SeriesPhysics Related
🎯 What it does: Propose a Fern framework for elliptical time series prediction based on the Brenier theorem, treating prediction as optimal transport from a fixed Gaussian source to data-related ellipsoids, directly learning the symmetric positive definite Jacobian;
Embedding Hybrid Systems into Continuous Latent Vector Fields
Sangli Teng (University of California Berkeley), Koushil Sreenath (University of California Berkeley)
OptimizationComputational EfficiencyRepresentation LearningReinforcement Learning from Human FeedbackRecurrent Neural NetworkAuto EncoderContrastive LearningTime SeriesSequentialStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proves that any n-dimensional hybrid system can be embedded into an m-dimensional Euclidean space, and constructs continuous vector fields on the embedding map; subsequently, it proposes the CHyLL++ framework based on implicit neural ODEs, which learns the flow of hybrid systems from time series data.
Embedding Trust: Semantic Isotropy Predicts Nonfactuality in Long-Form Text Generation
Dhrupad Bhardwaj (New York University), Tim G. J. Rudner (University of Toronto)
GenerationExplainability and InterpretabilityTransformerLarge Language ModelContrastive LearningText
🎯 What it does: This paper proposes to predict the non-factuality of long text generation in large language models by utilizing the dispersion of text embeddings (semantic isotropy).
Embedding-perturbed Exploration Preference Optimization for Flow Models
Sujie Hu (Tsinghua University), Xiu Li (Tsinghua University)
OptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelFlow-based ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose the E2PO framework, which addresses the variance disappearance and reward hijacking issues in traditional group-based RL by injecting structured perturbations into the text embedding space to maintain intra-group discriminative variance when aligning flow models in reinforcement learning.
EMBGuard: Constructing Hazard-Aware Guardrails for Safe Planning in Embodied Agents
Dongwook Choi (Yonsei University), Jinyoung Yeo (Yonsei University)
Safty and PrivacyExplainability and InterpretabilityRobotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: Developed EMBGUARD — a multimodal safety guard for embodied agents, used to assess the risk of action-image pairs and provide natural language explanations.
Embodied Interpretability: Linking Causal Understanding to Generalization in Vision-Language-Action Models
Hanxin Zhang (University of Leicester), Zhou Daniel Hao (University of Leicester)
Explainability and InterpretabilityRobotic IntelligenceVision-Language-Action ModelMultimodalityBenchmark
🎯 What it does: Proposes an explanation method based on causal intervention—Interventional Significance Score (ISS) and Nuisance Mass Ratio (NMR), used to quantify the causal dependence of visual language action (VLA) models on visual information and the utilization of irrelevant features.
Embodied Task Planning via Graph-Informed Action Generation with Large Language Models
Xiang Li (Purdue University), Masood S. Mortazavi
Robotic IntelligenceReinforcement Learning from Human FeedbackGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringTextGraphBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposes the GiG (Graph-in-Graph) framework, which combines scene graphs with state transition graphs, uses GNN to encode the environment, and guides LLM to generate actions by retrieving similar experiences;
Embodied-DETR: End-to-End Temporal 3D Object Detection in Egocentric Views
Ziheng Ding (Fudan University), Rui Feng (Fudan University)
Object DetectionConvolutional Neural NetworkTransformerVideoPoint CloudBenchmark
🎯 What it does: Proposed an end-to-end temporal 3D object detection framework called Embodied-DETR for first-person perspective continuous RGB-D streams, and created a dedicated Embodied-Det benchmark.
EmBrace: A Collective Knowledge Fusion Framework Toward Unified EEG Foundation Models
Ziyu Jia (Beijing Key Laboratory of Brainnetome and Brain-Computer Interface, Institute of Automation, Chinese Academy of Sciences), Chenyu Liu (Nanyang Technological University)
Federated LearningComputational EfficiencyKnowledge DistillationRepresentation LearningMixture of ExpertsContrastive LearningBiomedical DataElectrocardiogram
🎯 What it does: Proposed a collective knowledge fusion framework called EmBrace, aiming to optimize the performance of multiple electroencephalogram-based foundation models (EFMs) through sample-aware knowledge fusion.
Emergence of Biased Consensus in Multi-Agent LLM Debates
Maya Okawa (Harvard University)
OptimizationFederated LearningExplainability and InterpretabilityComputational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringMixture of ExpertsDiffusion modelTextPhysics RelatedRetrieval-Augmented Generation
🎯 What it does: This study explores the phenomenon of collective bias consensus that emerges in multi-agent LLM debates through experimental and theoretical analysis.
Emergence of Exploration in Policy Gradient Reinforcement Learning via Retrying
Soichiro Nishimori (University of Tokyo), Yutaka Matsuo (University of Tokyo)
Reinforcement LearningTabularTime SeriesSequentialBenchmark
🎯 What it does: Proposes the ReMax objective based on retrying, which directly achieves exploration by maximizing the expected reward;
Emergence of Hierarchical Emotion Organization in Large Language Models
Maya Okawa (Harvard University), Hidenori Tanaka (NTT Research, Inc.)
RecognitionTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper constructs an emotional hierarchy tree by analyzing the logits of large language models (LLMs), evaluates the model's ability in emotional recognition, and explores the impact of different demographic characteristics on emotional recognition.
Emergent Alignment via Competition
Natalie Collina (University of Pennsylvania), Mirah Shi (University of Pennsylvania)
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackMixture of ExpertsTextTabular
🎯 What it does: This paper investigates whether competitive mechanisms can enable users to achieve utility comparable to that of a perfectly aligned model when interacting with multiple AI models of different biases; it proposes the market alignment hypothesis (user utility can be approximated as a non-negative convex combination of utilities from each model) and proves that under this hypothesis, users can achieve almost optimal gains in a multi-leader Stackelberg game.
Emergent Analogical Reasoning in Transformers
Gouki Minegishi (University of Tokyo), Yutaka Matsuo (University of Tokyo)
OptimizationExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: This paper proposes a synthetic task based on the concept of functors in category theory to evaluate the performance of Transformers in analogical reasoning and compositional reasoning, and systematically analyzes the impact of data characteristics, optimization settings, and model scale on the emergence of analogical reasoning in this task;
Emergent Communication Under Misinformation
Heeyoung Lee (Yonsei University), Kyungwoo Song (Yonsei University)
OptimizationRepresentation LearningRecurrent Neural NetworkReinforcement LearningTabular
🎯 What it does: Studies the impact of malicious intermediaries on the emergence of language, demonstrating that misinformation can promote the emergence of compositional language.
Emergent Visual Representations through Unsupervised Spiking Networks with Synaptic Pruning
Di Hong (Zhejiang University), Yueming Wang (Zhejiang University)
ClassificationImage TranslationRepresentation LearningConvolutional Neural NetworkSpiking Neural NetworkContrastive LearningImageBiomedical Data
🎯 What it does: A deep spiking neural network (SNN) based on unsupervised contrastive learning and activity-dependent synaptic pruning was constructed, generating visual representations highly aligned with the visual cortex of rodents and primates through adaptive learning and structural sparsification.
EMFormer: Efficient Multi-Scale Transformer for Accumulative Context Weather Forecasting
Hao Chen (Hong Kong University of Science and Technology), LEI BAI
OptimizationComputational EfficiencyTransformerSupervised Fine-TuningMixture of ExpertsContrastive LearningTime Series
🎯 What it does: Proposed an efficient multi-scale Transformer (EMFormer) for long-term weather forecasting, and introduced cumulative context fine-tuning and sine-weighted loss to improve the accuracy of long-term predictions.
Empirical Gaussian Processes
Jihao Andreas Lin (Meta), Eytan Bakshy (Meta)
Hyperparameter SearchData-Centric LearningTabularTime SeriesBenchmarkFinance RelatedPhysics Related
🎯 What it does: Propose Empirical Gaussian Processes (Empirical GP), a nonparametric method that estimates the GP prior mean and covariance directly from historical sample paths.
Empty Shelves or Lost Keys? Recall Is the Bottleneck for Parametric Factuality
Nitay Calderon (Google Research), Gal Yona (Google Research)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: A knowledge dissection framework is proposed by constructing the WikiProfile benchmark and conducting knowledge analysis at the encoding and recall levels for 13 LLMs.
EmWorld: Emotion World Model with Latent State Evolution for Scenario-Incremental Dynamic Facial Expression Recognition
ke Wang, Zhe Chen (La Trobe University)
RecognitionTransformerContrastive LearningWorld ModelOptical FlowImageVideo
🎯 What it does: Propose EmWorld, a dual-timescale emotional world model that explicitly models scene evolution and facial expression dynamics, achieving incremental dynamic facial expression recognition in scenes;
End-to-End Autoregressive Image Generation with 1D Semantic Tokenizer
Wenda Chu (ByteDance Seed), Qiushan Guo (ByteDance Seed)
GenerationTransformerVision-Language-Action ModelDiffusion modelAuto EncoderGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Proposed an end-to-end autoregressive image generation framework called EOSTok, which jointly trains a 1D visual tokenizer with an autoregressive generation model, directly feeding the generated results back into the tokenizer.
End-to-End Compression for Tabular Foundation Models
Guri Zabërgja, Josif Grabocka (Technical University of Nuremberg)
ClassificationData SynthesisCompressionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderContrastive LearningTabular
🎯 What it does: Developed an end-to-end compression model called TACO, which can compress the training set to less than 1% of its original size without significantly degrading prediction performance, thereby significantly reducing the inference complexity of the attention mechanism in Transformer-based tabular foundation models.
Endogenous Resistance to Activation Steering in Language Models
Alex McKenzie (AE Studio), Michael S. A. Graziano (Princeton University)
Explainability and InterpretabilityComputational EfficiencyRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringAuto EncoderText
🎯 What it does: Studied whether large language models can self-correct after interventions in the activation space, and introduced the concept of 'Endogenous Activation Resistance' (ESR).
EnerGS: Energy-Based Gaussian Splatting under Partial Geometric Priors
Rui Song (University of California Los Angeles), Jiaqi Ma (University of California Los Angeles)
Autonomous DrivingOptimizationComputational EfficiencyDiffusion modelScore-based ModelNeural Radiance FieldGaussian SplattingImagePoint Cloud
🎯 What it does: Proposed a 3D Gaussian Splatting method called EnerGS based on an energy field, which utilizes partial geometric priors (such as LiDAR) to construct a continuous geometric energy field to guide the distribution of Gaussian primitives, thereby achieving more stable and accurate view synthesis in large-scale outdoor scenes.
Energy-based Compositional Diffusion Planning
Tao Sun (Stanford University), Iro Armeni (Stanford University)
OptimizationRobotic IntelligenceDiffusion modelScore-based ModelPoint CloudTabularTime SeriesSequentialBenchmark
🎯 What it does: Propose Energy-based Compositional Diffuser (ECD), which achieves high-quality long-horizon trajectory planning by summing the bridge energy of local trajectory blocks to obtain global energy, and using its negative gradient as the conservative correction direction.
Energy-Structured Low-Rank Adaptation for Continual Learning
Longhua Li (Southeast University), Xin Geng (Southeast University)
ClassificationComputational EfficiencyKnowledge DistillationRepresentation LearningImage
🎯 What it does: Proposes the E2-LoRA framework, leveraging the low-rank energy concentration characteristics of task-induced output feature drift to achieve continual learning;
EngiAgent: Fully Connected Coordination of LLM Agents for Solving Open-ended Engineering Problems with Feasible Solutions
Xiyuan Zhou (Nanyang Technological University), Jinjin Gu (INSAIT Sofia University St Kliment Ohridski)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Proposed a multi-agent system called EngiAgent, equipped with a fully connected coordinator, focusing on achieving feasibility-priority full-process automation in engineering problem solving.
Enhanced Latent-Space Adversarial Training for Super-Resolution
Liangbin Xie (State Key Laboratory of Internet of Things for Smart City, University of Macau), Chao Dong (Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences)
Super ResolutionKnowledge DistillationConvolutional Neural NetworkTransformerDiffusion modelGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: Propose HYPIR++, a two-stage cascaded framework for real-world image super-resolution at large magnification factors (≥4×), improving upon the architecture and training approach of HYPIR.
Enhanced Multi-Instance Partial Label Learning via Average Gradient Outer Product
Nan Cao (Huazhong University of Science and Technology), Teng Zhang (Huazhong University of Science and Technology)
ClassificationAnomaly DetectionComputational EfficiencyRepresentation LearningTransformerDiffusion modelScore-based ModelFlow-based ModelAuto EncoderContrastive LearningImageBenchmarkAudio
🎯 What it does: Propose a multi-instance partial label learning framework called AGOPMIPL, which rescales the feature space using average gradient outer product (AGOP), enabling the localization of key instances and elimination of candidate label errors under noisy candidate labels.
Enhancing Affine Maximizer Auctions with Correlation-Aware Payment
Haoran Sun (Peking University), Xiaotie Deng (Peking University)
OptimizationFederated LearningReinforcement Learning from Human FeedbackTransformerReinforcement LearningPrompt EngineeringContrastive LearningTabularTime SeriesSequentialBenchmarkFinance RelatedChain-of-Thought
🎯 What it does: Proposed and optimized the Correlation-Aware Affine Maximizer Auction (CA-AMA), which maintains DSIC under value-correlated distributions and improves revenue by incorporating a correlation-aware payment term.
Enhancing Conformal Prediction via Class Similarity
Ariel Fargion (Bar-Ilan University), Tom Tirer (Bar-Ilan University)
ClassificationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkTransformerContrastive LearningImage
🎯 What it does: Propose a regularization method based on class similarity, which can improve the prediction set size and semantic consistency of any conformal prediction (CP) algorithm while maintaining coverage guarantees.
Enhancing Cross-subject Emotion Recognition via Heterogeneous Distribution Augmentation and Collaborative Learning
Wending Xiong (Wuhan University), Mang Ye (Wuhan University)
RecognitionData SynthesisDomain AdaptationGenerative Adversarial NetworkContrastive LearningMultimodalityTime SeriesBiomedical Data
🎯 What it does: Proposes the MixEmo framework, which enhances the generalization ability of cross-subject emotion recognition by separating and recombining the distribution of emotional data to generate unseen distributions, and collaboratively learning across multiple sub-distributions.