ICLR 2026 Papers — Page 14
International Conference on Learning Representations · 5356 papers
DRIFT: Learning from Abundant User Dissatisfaction in Real-World Preference Learning
Yifan Wang (Purdue University), Qingkai Zeng (Nankai University)
Data-Centric LearningReinforcement Learning from Human FeedbackContrastive LearningText
🎯 What it does: Proposed an iterative preference learning framework called DRIFT, based on real user dissatisfaction feedback, using dissatisfied (DSAT) examples as negative samples and dynamically sampling positive samples for training.
DriftLite: Lightweight Drift Control for Inference-Time Scaling of Diffusion Models
Yinuo Ren (Stanford University), Jiequn Han (Flatiron Institute)
GenerationComputational EfficiencyDiffusion modelBiomedical DataPhysics RelatedStochastic Differential Equation
🎯 What it does: Propose DriftLite, a lightweight, training-free inference-time adaptation method that adjusts pre-trained diffusion models to adapt to new target distributions by dynamically controlling drift during inference.
DriveAgent-R1: Advancing VLM-based Autonomous Driving with Active Perception and Hybrid Thinking
Weicheng Zheng (Shanghai Qi Zhi Institute), Hang Zhao (Shanghai Qi Zhi Institute)
Autonomous DrivingLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIVision Language ModelVision-Language-Action ModelVideoTextMultimodalityRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Developed a 3B-parameter autonomous driving agent, DriveAgent-R1, capable of proactively invoking visual tools in high-level behavior planning and dynamically switching between text reasoning and tool reasoning based on scene complexity;
DriveMamba: Task-Centric Scalable State Space Model for Efficient End-to-End Autonomous Driving
Haisheng Su (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Autonomous DrivingMultimodality
🎯 What it does: Proposed a Mamba-based end-to-end autonomous driving framework called DriveMamba, integrating perspective correspondence learning, task relationship modeling, and long-term temporal fusion, achieving efficient scalability through sparse representation and linear-complexity state space models;
DriveVLA-W0: World Models Amplify Data Scaling Law in Autonomous Driving
Yingyan Li, Zhaoxiang Zhang
Autonomous DrivingMixture of ExpertsVision Language ModelVision-Language-Action ModelDiffusion modelWorld ModelVideoBenchmark
🎯 What it does: Provide dense self-supervised learning through future image prediction to address the lack of supervision in Vision-Language-Action (VLA) models; simultaneously achieve driving planning performance surpassing traditional BEV and multi-modal models under single front-view camera conditions.
DrivingGen: A Comprehensive Benchmark for Generative Video World Models in Autonomous Driving
Yang Zhou (University of Toronto), Steven L. Waslander (University of Toronto)
Autonomous DrivingTransformerVision Language ModelWorld ModelOptical FlowVideoBenchmark
🎯 What it does: Propose the DrivingGen benchmark to evaluate generative world models in the autonomous driving domain, containing diverse weather conditions, times, regions, and complex interaction scenarios;
Dropping Just a Handful of Preferences Can Change Top Large Language Model Rankings
Jenny Y. Huang (MIT-IBM Watson AI Lab), Tamara Broderick (MIT-IBM Watson AI Lab)
Large Language ModelTextBenchmark
🎯 What it does: Investigated the phenomenon where discarding an extremely small proportion (<0.01%) of user/AI preference data in ranking evaluation platforms using the Bradley-Terry model for large language models can lead to ranking inversions, and proposed a rapid detection method.
DRPO: Efficient Reasoning via Decoupled Reward Policy Optimization
Gang Li (Texas A&M University), Tianbao Yang (Texas A&M University)
OptimizationComputational EfficiencyLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes Decoupled Reward Policy Optimization (DRPO), a reinforcement learning framework that decouples positive and negative sample rewards and normalizes the length reward of positive samples, to train large reasoning models (LRMs) for efficient inference, significantly reducing redundant inference paths.
Drugging the Undruggable: Benchmarking and Modeling Fragment-Based Screening
Haichuan Tan (Tsinghua University), Yanyan Lan (Tsinghua University)
Drug DiscoveryGraph Neural NetworkTransformerLarge Language ModelContrastive LearningBiomedical DataBenchmark
🎯 What it does: Explored methods for fragment-level virtual screening on undruggable targets, proposing the FragBench benchmark and the FragCLIP tri-modal contrastive learning framework.
DrugTrail: Interpretable Drug Discovery via Structured Reasoning and Druggability‑Tailored Preference Optimization
Yurou Liu (Renmin University of China), Zheng Wang (Alibaba Cloud Computing)
OptimizationExplainability and InterpretabilityDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringBiomedical Data
🎯 What it does: Proposes the DRUGTRAIL framework, which generates explainable drug design reasoning trajectories using LLMs and produces molecules with high efficacy and accessibility through reward learning;
DrVoice: Parallel Speech-Text Voice Conversation Model via Dual-Resolution Speech Representations
Chao-Hong Tan (Alibaba Group), Jieping Ye (Alibaba Group)
GenerationTransformerLarge Language ModelFlow-based ModelTextMultimodalityAudio
🎯 What it does: Built an end-to-end parallel speech-text generation model called DRVOICE, achieving joint autoregressive generation of dual-resolution speech representations and a speech refinement head;
DSA: Efficient Inference For Video Generation Models via Distributed Sparse Attention
Shenggui Li (Nanyang Technological University), Tianwei Zhang (Agency for Science Technology and Research)
GenerationComputational EfficiencyTransformerVideo
🎯 What it does: Propose Distributed Sparse Attention (DSA), combining sparse attention with distributed inference to optimize the inference speed of video generation models.
DTO-KD: Dynamic Trade-off Optimization for Effective Knowledge Distillation
Zeeshan Hayder (Data61 / CSIRO), Richard Hartley (Data61 / CSIRO)
ClassificationObject DetectionKnowledge DistillationTransformerImage
🎯 What it does: Propose the DTO-KD framework, dynamically balancing task loss and distillation loss in knowledge distillation to achieve gradient-level multi-objective optimization.
DTP: Delta-Guided Two Stage Pruning for Mamba-based Multimodal Large Language Models
Seong Yeol Park (Sejong University), Yeong Hyeon Gu (Sejong University)
Computational EfficiencyLarge Language ModelMultimodality
🎯 What it does: Developed a Delta-guided Two Stage Pruning method tailored for the Mamba architecture multimodal large language models, achieving progressive pruning of visual tokens and significantly reducing inference costs.
Dual Distillation for Few-Shot Anomaly Detection
Le Dong, Lichao Mou (Technical University Of Munich)
Anomaly DetectionKnowledge DistillationConvolutional Neural NetworkImageBiomedical DataBenchmark
🎯 What it does: Proposes the D2 4FAD dual distillation framework, which achieves few-shot anomaly detection using a small number of normal reference images.
Dual Goal Representations
Seohong Park (University of California, Berkeley), Sergey Levine (University of California, Berkeley)
Reinforcement LearningContrastive LearningImageTabularBenchmark
🎯 What it does: Proposed and verified dual goal representations, which describe goals using a set of time distances from a goal state to all other states, achieving noise-robust and sufficiently informative representations.
Dual Optimistic Ascent (PI Control) is the Augmented Lagrangian Method in Disguise
Juan Ramirez (Mila Quebec AI Institute Universite De Montreal), Simon Lacoste-Julien (Mila Quebec AI Institute Universite De Montreal)
Optimization
🎯 What it does: Proves that double optimistic ascent (PI control) is equivalent to the augmented Lagrangian method within a single-step first-order optimization framework, thereby obtaining theoretical guarantees for the latter and explaining the effectiveness of optimistic gradients in constrained deep learning.
Dual Perspectives on Non-Contrastive Self-Supervised Learning
Jean Ponce (Ecole normale superieure/PSL New York University), Michael Arbel (Universite Grenoble Alpes)
OptimizationRepresentation LearningVideo
🎯 What it does: This paper deeply investigates two commonly used techniques, stop gradient (SG) and exponential moving average (EMA), in non-contrastive self-supervised learning from the dual perspectives of optimization and dynamics, proving that they do not optimize the original objective function but can avoid representation collapse during training, and characterizing their equilibrium points and stability under linear models.
Dual Randomized Smoothing: Beyond Global Noise Variance
Chenhao Sun (ETH Zürich), Martin Vechev (ETH Zürich)
ClassificationAdversarial AttackConvolutional Neural NetworkTransformerMixture of ExpertsDiffusion modelImage
🎯 What it does: Propose a Dual Random Smoothing (Dual RS) framework that allows using input-dependent noise variance during the random smoothing process, and prove robustness guarantees under local constant variance conditions.
Dual-Branch Representations with Dynamic Gated Fusion and Triple-Granularity Alignment for Deep Multi-View Clustering
Wenyuan Kong (Independent Researcher), Bing Li (China University of Labor Relations)
Representation LearningGraph Neural NetworkAuto EncoderContrastive LearningMultimodality
🎯 What it does: Propose a dual-branch deep multi-view clustering framework named DREAM, which separately extracts semantic (VAE) and structural (GCN) information, and achieves more discriminative clustering representations through dynamic gating fusion and triple-granularity alignment.
Dual-IPO: Dual-Iterative Preference Optimization for Text-to-Video Generation
Xiaomeng Yang (Shanghai Academy of AI for Science), Hao Li (Fudan University)
GenerationReinforcement LearningVision Language ModelDiffusion modelVideoTextMultimodalityChain-of-Thought
🎯 What it does: Developed the Dual-IPO framework, which significantly improves the quality of text-to-video generation and alignment with human preferences through alternating optimization of the reward model and video generation model.
Dual-Kernel Adapter: Expanding Spatial Horizons for Data-Constrained Medical Image Analysis
Ziquan Zhu (University of Leicester), Tianjin Huang (Mohamed bin Zayed University of Artificial Intelligence)
ClassificationSegmentationConvolutional Neural NetworkBiomedical Data
🎯 What it does: Systematic evaluation of adapter fine-tuning in low-data medical imaging tasks, and proposing Dual-Kernel Adapter (DKA) to expand the effective receptive field
Dual-objective Language Models: Training Efficiency Without Overfitting
David Samuel (University of Oslo), Lucas Georges Gabriel Charpentier (National Library of Norway)
Computational EfficiencyTransformerDiffusion modelText
🎯 What it does: Studied a dual-objective training method that simultaneously trains autoregressive and masked diffusion objectives within the same Transformer architecture to improve training efficiency and reduce overfitting.
Dual-Objective Reinforcement Learning with Novel Hamilton-Jacobi-Bellman Formulations
William Sharpless (University of California San Diego), Sylvia Lee Herbert (University of California San Diego)
Reinforcement LearningStochastic Differential Equation
🎯 What it does: Proposed two new multi-objective reinforcement learning tasks: Reach-Always-Avoid (RAA) and Reach-Reach (RR), along with their corresponding Hamilton-Jacobi-Bellman formulations and value functions.
Dual-Path Condition Alignment for Diffusion Transformers
Changhao Peng (Peking University), Wei Gao (Peking University)
GenerationTransformerDiffusion modelImage
🎯 What it does: Propose an unsupervised alignment method (DUPA) that does not require an external visual encoder. By performing multiple random noise samplings on the same image and processing them with a Decoupled Diffusion Transformer (DDT), the method enhances the performance of generative models by aligning conditional features obtained through different noise paths.
Dual-Robust Cross-Domain Offline Reinforcement Learning Against Dynamics Shifts
Zhongjian Qiao (City University of Hong Kong), Shuang Qiu (City University of Hong Kong)
Domain AdaptationReinforcement LearningBenchmark
🎯 What it does: Proposed a dual robust method for cross-domain offline reinforcement learning (training and testing) called DROCO, verifying its superior performance in dynamic shift scenarios.
Dual-Scale World Memory for LLM Agents towards Hard-Exploration Problems
Minsoo Kim (Seoul National University), seung-won hwang
TransformerLarge Language ModelReinforcement LearningAgentic AIWorld ModelText
🎯 What it does: Proposed and implemented the GLoW framework, leveraging dual-scale world memory (global trajectory frontier and local advantage reflection) to address the hard exploration challenge faced by LLM agents in text-based games;
Dual-Solver: A Generalized ODE Solver for Diffusion Models with Dual Prediction
Soochul Park (SteAI MODULABS), Yeon Ju Lee (Korea University)
GenerationDiffusion modelImageOrdinary Differential Equation
🎯 What it does: Propose a learnable dual solver (Dual-Solver) that achieves second-order numerical accuracy in prediction-correction sampling within diffusion models through learnable parameters.
Dual-Space Smoothness for Robust and Balanced LLM Unlearning
Han Yan (Chinese University of Hong Kong), Meng Jiang (University of Notre Dame)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: Proposed the PRISM framework for machine unlearning in large language models, enhancing robustness against leakage, relearning, and jailbreak attacks through dual-space smoothing (representation space and parameter space) and gradient conflict decoupling, achieving a balance between forgetting effectiveness, practicality, and privacy protection.
DualEdit: Mitigating Safety Fallback in LLM Backdoor Editing via Affirmation-Refusal Regulation
Houcheng Jiang (University Of Science And Technology Of China), Yang Deng (Singapore Management University)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: For safety-aligned large language models, we propose a dual-objective editing framework called DualEdit, designed to enhance positive responses and suppress rejection behaviors when injecting backdoors, thereby eliminating mid-process safety rollbacks.
DualMap: Enabling Both Cache Affinity and Load Balancing for Distributed LLM Serving
Ying Yuan (Huazhong University of Science and Technology), Zhou Yu (Huawei)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Propose DualMap, a dual-mapping scheduling strategy, to simultaneously achieve KV cache affinity and load balancing in distributed LLM services;
DualToken: Towards Unifying Visual Understanding and Generation with Dual Visual Vocabularies
Wei Song (Zhejiang University), Kaicheng Yu (Zhejiang University)
RecognitionGenerationTransformerLarge Language ModelVision Language ModelAuto EncoderContrastive LearningImageTextMultimodality
🎯 What it does: Propose the DualToken visual tokenizer, which uses two codebooks (pixel codebook and semantic codebook) to simultaneously support visual understanding and generation, and integrates it into a purely autoregressive large language model.
DUET: Distilled LLM Unlearning from an Efficiently Contextualized Teacher
Yisheng Zhong (George Mason University), Zhuangdi Zhu (George Mason University)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Proposes a distillation-based LLM lightweight 'forgetting' method called DUET, which utilizes the teacher model's contextual rejection behavior through Top-K prediction scores to fine-tune the student model, achieving precise deletion of undesirable knowledge.
DUET: Optimizing LLM Training Data Mixtures via Noisy Feedback from Unseen, Downstream Evaluation Tasks
Zhiliang Chen (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationData-Centric LearningTransformerLarge Language ModelText
🎯 What it does: Propose the DUET algorithm to optimize the mixture ratio of LLM training data through multi-round noisy feedback and Bayesian optimization, even when the evaluation task data is unknown.
DuPO: Enabling Reliable Self-Verification via Dual Preference Optimization
Shuaijie She (Nanjing University), Yuxuan Wang (ByteDance Seed)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningTextBenchmark
🎯 What it does: Proposes the DuPO framework, leveraging bidirectional learning and self-supervised rewards of LLMs for label-free optimization of model performance;
Durian: Dual Reference Image-Guided Portrait Animation with Attribute Transfer
Hyunsoo Cha (Seoul National University), Hanbyul Joo (Seoul National University)
Image TranslationImage HarmonizationGenerationPose EstimationTransformerVision Language ModelDiffusion modelContrastive LearningImageVideo
🎯 What it does: Propose a dual-reference diffusion network named Durian based on self-reconstruction training, which can generate animations and perform attribute transfer from a single portrait and one or more attribute reference images without requiring cross-identity paired data.
DVD-Quant: Data-free Video Diffusion Transformers Quantization
Zhiteng Li (Shanghai Jiao Tong University), Xiaokang Yang (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerDiffusion modelVideoBenchmark
🎯 What it does: Propose a data-free post-training quantization framework named DVD-Quant, specifically designed for video diffusion Transformers to achieve low-bit quantization such as W4A4;
DVLA-RL: Dual-Level Vision–Language Alignment with Reinforcement Learning Gating for Few-Shot Learning
Wenhao Li (Shandong University), Yilong Yin (Shandong University)
Meta LearningTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the DVLA-RL framework, achieving few-shot learning through dual-layer vision-language alignment and reinforcement learning gated attention.
Dyna-Mind: Learning to Simulate from Experience for Better AI Agents
Xiao Yu (Columbia University), Jianfeng Gao (Microsoft Research)
Robotic IntelligenceTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: Propose the Dyna-Mind two-stage framework, first using RESIM to generate simulated reasoning trajectories based on real-world experiences, then using DYNA-GRPO in online reinforcement learning to integrate intermediate state feedback and enhance the (V)LM agent's simulation and decision-making capabilities.
DynaGuard: A Dynamic Guardian Model With User-Defined Policies
Monte Hoover (University of Maryland), Tom Goldstein (University of Maryland)
Safty and PrivacyExplainability and InterpretabilityTransformerSupervised Fine-TuningTextBenchmarkChain-of-Thought
🎯 What it does: Developed the dynamic safeguard model DynaGuard and its corresponding dataset DynaBench, to support user-defined safeguard policies and provide explanations and fast inference;
Dynamic Chunking for End-to-End Hierarchical Sequence Modeling
Sukjun Hwang (Carnegie Mellon University), Albert Gu (Carnegie Mellon University)
Representation LearningTransformerTextBiomedical Data
🎯 What it does: Designed and implemented an end-to-end hierarchical network (H-Net) that achieves sequence modeling through a dynamic chunking mechanism without relying on any predefined tokenizer.
Dynamic Classifier-Free Diffusion Guidance via Online Feedback
Pinelopi Papalampidi (Google DeepMind), Aida Nematzadeh (Google DeepMind)
GenerationVision Language ModelDiffusion modelImageText
🎯 What it does: Propose a framework that dynamically schedules classifier-free guidance (CFG) scale in real-time during text-to-image diffusion generation; at each sampling step, the quality of the current noise latent variable is evaluated using online evaluators (e.g., CLIP, discriminators, human reward models, OCR, numerical reasoning), and the optimal CFG level is greedily selected to generate adaptive guidance curves for each prompt.
Dynamic Early Exit in Reasoning Models
Chenxu Yang (Institute of Information Engineering Chinese Academy of Sciences), Weiping Wang (Institute of Information Engineering Chinese Academy of Sciences)
Computational EfficiencyPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: Proposed a training-free dynamic early exit mechanism called DEER, allowing large-scale inference models to adaptively terminate thinking early and directly provide answers during chain-of-thought (CoT) generation based on their own confidence.
Dynamic Multi-sample Mixup with Gradient Exploration for Open-set Graph Anomaly Detection
Caiyang Yu (Sichuan University of China), Ziyue Qiao (Great Bay University)
Anomaly DetectionGraph Neural NetworkGraph
🎯 What it does: Propose a framework called DEMO for open-set graph anomaly detection, combining multi-sample Mixup, energy gradient adaptive weighting, and pseudo-label generation assisted by historical memory, enabling the detection of unknown anomalies under extremely limited label scenarios;
Dynamic Multimodal Activation Steering for Hallucination Mitigation in Large Vision-Language Models
Jianghao Yin (East China Normal University), Liang He (East China Normal University)
Explainability and InterpretabilityVision Language ModelMultimodalityRetrieval-Augmented Generation
🎯 What it does: Proposed a training-free dynamic multimodal activation modulation (DMAS) method, which alleviates hallucination problems by performing semantic context-aware activation vector interventions on the attention heads of large vision-language models (LVLMs) during inference.
Dynamic Novel View Synthesis in High Dynamic Range
Kaixuan Zhang (Nanjing University of Science and Technology), Xiatian Zhu (University of Surrey)
GenerationData SynthesisRecurrent Neural NetworkTransformerGaussian SplattingVideo
🎯 What it does: Proposed and implemented the HDR Dynamic Novel View Synthesis (HDR DNVS) task and the HDR-4DGS model, achieving temporally and spatially consistent HDR rendering through Gaussian Splatting combined with dynamic tone mapping.
Dynamic Reflections: Probing Video Representations with Text Alignment
Maks Ovsjanikov (Google DeepMind), Tengda Han (Google DeepMind)
ClassificationPose EstimationRetrievalRepresentation LearningTransformerVision Language ModelVideoTextMultimodality
🎯 What it does: This study explores cross-modal alignment between videos and text, significantly improving alignment by using multi-frame videos and multiple captions during testing, and constructing a zero-shot evaluation framework.
Dynamic Speculative Agent Planning
Yilin Guan (Johns Hopkins University), Wenyue Hua (University of California, Santa Barbara)
Computational EfficiencyTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: Proposed a dynamic speculative planning framework called DSP, which is driven by online reinforcement learning without prior preparation. It can dynamically predict and adjust speculative steps during the execution of LLM agents to achieve lossless acceleration and cost control.
Dynamic Weight Grafting: Localizing Finetuned Factual Knowledge in Transformers
Todd Nief (University of Chicago), Ari Holtzman (University of Chicago)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Developed a dynamic weight grafting method to locate how large language models retrieve new relational knowledge after fine-tuning.
Dynamic-dLLM: Dynamic Cache-Budget and Adaptive Parallel Decoding for Training-Free Acceleration of Diffusion LLM
Tianyi Wu (Harbin Institute of Technology), Zhuotao Tian (Harbin Institute of Technology)
Computational EfficiencyTransformerLarge Language ModelDiffusion modelText
🎯 What it does: Proposes a training-free framework called Dynamic-dLLM to accelerate inference for diffusion-based large language models (dLLMs).
Dynamical properties of dense associative memory
Kazushi Mimura (Hiroshima City University), Anthony CC Coolen
OptimizationPhysics Related
🎯 What it does: Perform a generating function analysis (GFA) under the system limit of large-scale dense associative memory networks (n≥3) for Krotov, providing exact dynamic solutions, including convergence time, size of the attractor basin, and characteristics of noise variance varying with overlap.
DynamicInfer: Runtime-Aware Sparse Offloading for LLMs Inference on a Consumer-Grade GPU
Zhui Zhu (Tsinghua University), Fan Dang (Beijing Jiaotong University)
OptimizationComputational EfficiencyTransformerText
🎯 What it does: Proposed a real-time neuron scheduling and sparse offloading framework called DynamicInfer, enabling large language models to perform efficient inference on consumer-grade GPUs, addressing the issues of low GPU utilization and high latency caused by static hot-cold neuron partitioning.
Dynamics-Predictive Sampling for Active RL Finetuning of Large Reasoning Models
Yixiu Mao (Tsinghua University), Xiangyang Ji (Tsinghua University)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Propose an online prediction method based on dynamic systems (DPS), which uses a hidden Markov model to perform Bayesian inference on the problem-solving state of prompts, enabling efficient selection of informative prompts without expensive rollouts during reinforcement learning fine-tuning of large language models.
Dyslexify: A Mechanistic Defense Against Typographic Attacks in CLIP
Lorenz Hufe (Fraunhofer Heinrich Hertz Institute), Wojciech Samek (Fraunhofer Heinrich Hertz Institute)
Explainability and InterpretabilityComputational EfficiencyAdversarial AttackTransformerVision Language ModelImageBiomedical Data
🎯 What it does: Propose a gradient-agnostic defense method called Dyslexify, which significantly enhances the model's robustness against layout attacks by identifying and ablating attention heads in the CLIP vision encoder that are specialized for processing text.
E²LoRA: Efficient and Effective Low-Rank Adaptation with Entropy-Guided Adaptive Sharing
Minglei Li (Fudan University), Tao Chen (Fudan University)
Computational EfficiencyImageTextMultimodality
🎯 What it does: Propose E²LoRA, a dual adaptive parameter sharing and rank allocation framework based on gradient entropy analysis;
e3: Learning to Explore Enables Extrapolation of Test-Time Compute for LLMs
Amrith Setlur (Carnegie Mellon University), Aviral Kumar (Carnegie Mellon University)
Computational EfficiencyTransformerSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes a training scheme named e3, which trains LLMs to extrapolate computational load during inference by exploring within the context (e.g., generate-validate-summarize);
EA3D: Event-Augmented 3D Diffusion for Generalizable Novel View Synthesis
Wangbo Yu (Peking University), Yonghong Tian (Peng Cheng Laboratory)
GenerationData SynthesisDiffusion modelNeural Radiance FieldImageVideoMultimodality
🎯 What it does: Propose the EA3D framework, combining event streams with sparse RGB images to achieve generalizable high-fidelity novel view synthesis.
EAMET: ROBUST MASSIVE MODEL EDITING VIA EMBEDDING ALIGNMENT OPTIMIZATION
Yanbo Dai (Hong Kong University of Science and Technology), Shuai Wang (Hong Kong University of Science and Technology)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a model editing method called EAMET, which improves the effectiveness and robustness of large-scale fact editing by aligning key embeddings with residual embeddings.
Early Signs of Steganographic Capabilities in Frontier LLMs
Artur Zolkowski (ML Alignment & Theory Scholars), David Lindner (Google DeepMind)
Safty and PrivacyTransformerPrompt EngineeringTextChain-of-Thought
🎯 What it does: This paper systematically evaluates the capabilities of state-of-the-art large language models in steganography, including the transmission of hidden information and hidden reasoning, and demonstrates their potential risks through case studies.
Earth-Agent: Unlocking the Full Landscape of Earth Observation with Agents
Peilin Feng (Shanghai Artificial Intelligence Laboratory), Weijia Li (Tsinghua University)
TransformerLarge Language ModelAgentic AIImageMultimodalityBenchmark
🎯 What it does: Proposed the Earth-Agent framework, integrating ReAct and MCP protocols with LLM agents, combining 104 specialized tools to support multi-step quantitative reasoning in three EO modes: RGB, spectral, and product;
EarthSE: A Benchmark Evaluating Earth Scientific Exploration Capability for Large Language Models
Wanghan Xu (Shanghai Jiao Tong University), LEI BAI
Large Language ModelTextBenchmarkPhysics RelatedChain-of-Thought
🎯 What it does: Created a multi-level evaluation benchmark called EarthSE, covering five fields of Earth science, 114 disciplines, and 11 fundamental tasks, including the basic question-answering dataset Earth-Iron, the professional question-answering dataset Earth-Silver, and the open-ended dialogue dataset Earth-Gold designed for scientific exploration.
Easier Painting Than Thinking: Can Text-to-Image Models Set the Stage, but Not Direct the Play?
Ouxiang Li (University of Science and Technology of China), Fuli Feng (Kling Team, Kuaishou Technology)
GenerationTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed a comprehensive evaluation benchmark named T2I-COREBENCH for systematically assessing the performance of text-to-image models in two core capabilities: composition and reasoning.
EAST: Early Action Prediction Sampling Strategy with Token Masking
Iva Sović (University of Zagreb), Marin Oršić (University of Zagreb)
RecognitionTransformerAuto EncoderVideo
🎯 What it does: Propose the EAST framework, which can predict video actions at any observation ratio by training a single model.
EasyCreator: Empowering 4D Creation through Video Inpainting
Yue Ma (Hong Kong University Of Science And Technology), Qifeng Chen (Hong Kong University Of Science And Technology)
GenerationData SynthesisDepth EstimationSupervised Fine-TuningDiffusion modelVideoPoint Cloud
🎯 What it does: Propose EasyCreator, which transforms 4D video creation into a video inpainting task, enabling multi-view generation and content editing for single-camera videos.
EasyTune: Efficient Step-Aware Fine-Tuning for Diffusion-Based Motion Generation
Xiaofeng Tan (Southeast University), Hongsong Wang (Southeast University)
GenerationSupervised Fine-TuningReinforcement LearningDiffusion modelTextMultimodalityStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Proposes an efficient fine-tuning framework called EasyTune based on step-wise differential rewards, which aligns the semantics of text-to-motion diffusion models and constructs a reward model through self-calibrated preference learning (SPL) without requiring human-labeled data.
ECHO: Toward Contextual Seq2Seq Paradigms in Large EEG Models
Chenyu Liu (Nanyang Technological University), Yi Ding (Nanyang Technological University)
ClassificationRepresentation LearningMeta LearningConvolutional Neural NetworkTransformerPrompt EngineeringBiomedical Data
🎯 What it does: Propose a decoder-based sequence-to-sequence framework called ECHO for large EEG models, capable of achieving context learning through support samples and directly inferring tasks and labels in multi-task settings;
Echo: Towards Advanced Audio Comprehension via Audio-Interleaved Reasoning
Daiqing Wu (Chinese Academy of Sciences), Yu ZHOU
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMultimodalityRetrieval-Augmented GenerationChain-of-ThoughtAudio
🎯 What it does: Proposed and implemented an audio-interleaved reasoning framework, constructing the Echo large audio language model. It employs a two-stage training approach (SFT + RL) combined with a structured data generation pipeline, aiming to enhance the model's fine-grained understanding and reasoning capabilities for complex audio.
Echoes as Anchors: Probabilistic Costs and Attention Refocusing in LLM Reasoning
Zhuoyuan Hao (Harbin Institute of Technology), Jing Li (Harbin Institute of Technology)
Explainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: Analyzes the phenomenon of large inference models spontaneously repeating prompts at the beginning of generation (Echo of Prompt), and proposes a probability framework based on rejection sampling, an attention reorientation mechanism, and two methods to improve inference performance;
EchoGen: Generating Visual Echoes in Any Scene via Feed-Forward Subject-Driven Auto-Regressive Model
Ruixiao Dong (University of Science and Technology of China), Houqiang Li (University of Science and Technology of China)
GenerationTransformerVision Language ModelAuto EncoderContrastive LearningMultimodality
🎯 What it does: Developed EchoGen, a feedforward topic-driven image generation framework based on a visual autoregressive model;
EchoMind: An Interrelated Multi-level Benchmark for Evaluating Empathetic Speech Language Models
Li Zhou (Chinese University of Hong Kong, Shenzhen), Haizhou Li (Chinese University of Hong Kong, Shenzhen)
Large Language ModelMultimodalityBenchmarkAudio
🎯 What it does: Propose EchoMind, a multi-level and interrelated emotional speech evaluation benchmark, to assess the emotional intelligence of speech language models in understanding, reasoning, and dialogue;
EchoMotion: Unified Human Video and Motion Generation via Dual-Modality Diffusion Transformer
Yuxiao Yang (Tsinghua University), Jieping Ye (Alibaba Group)
GenerationData SynthesisTransformerDiffusion modelVideoMultimodality
🎯 What it does: Propose the EchoMotion framework, which jointly models the joint distribution of videos and human motion to generate more realistic complex human action videos.
EdgeCape: Edge Weight Prediction For Category-Agnostic Pose Estimation
Or Hirschorn (Tel Aviv University), Shai Avidan (Tel Aviv University)
Pose EstimationGraph Neural NetworkTransformerImage
🎯 What it does: Propose EdgeCape, a category-agnostic pose estimation method based on predicted weighted pose-graphs, which can accurately locate keypoints under a small number of support images.
EDINET-Bench: Evaluating LLMs on Complex Financial Tasks using Japanese Financial Statements
Issa Sugiura (Sakana AI), David Ha (Sakana AI)
ClassificationTransformerLarge Language ModelPrompt EngineeringTextTabularBenchmarkFinance Related
🎯 What it does: Developed the open-source benchmark EDINET-Bench for Japan's financial sector, constructing an evaluation dataset with three tasks: accounting fraud detection, profit forecasting, and industry classification, and assessed the performance of multiple LLMs and traditional models in a zero-shot setting.
Edit-Based Flow Matching for Temporal Point Processes
David Lüdke, Stephan Günnemann
GenerationTransformerFlow-based ModelTime Series
🎯 What it does: Proposes EDITPP, which uses a continuous-time Markov chain model based on insert, delete, and replace edit operations to perform non-autoregressive joint generation and prediction for temporal point processes (TPP);
EditBench: Evaluating LLM Abilities to Perform Real-World Instructed Code Edits
Wayne Chi (Carnegie Mellon University), Chris Donahue (Carnegie Mellon University)
AI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: Developed a benchmark called EditBench, collecting real user code editing instructions and contexts from VS Code, constructing 540 multilingual, multi-programming language, real-world editing tasks, and using it to evaluate the editing capabilities of 40 LLMs.
EditLens: Quantifying the Extent of AI Editing in Text
Katherine Thai (Pangram Labs), Mohit Iyyer (University of Maryland)
ClassificationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposed a continuous quantification method for detecting AI editing intensity called EDITLENS, which can estimate the proportion of AI edits in text;
EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing
Keming Wu (University of Waterloo), Wenhu Chen (University of Waterloo)
GenerationReinforcement Learning from Human FeedbackTransformerVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Constructed a large-scale manually annotated preference dataset called EDITREWARD-DATA, trained a reward model specialized for instruction-driven image editing called EDITREWARD based on this dataset, and proposed a multi-path priority evaluation benchmark called EDITREWARD-BENCH;
EditScore: Unlocking Online RL for Image Editing via High-Fidelity Reward Modeling
Xin Luo (University of Science and Technology of China), Zheng Liu (Beijing Academy of Artificial Intelligence)
Image TranslationSupervised Fine-TuningReinforcement LearningVision Language ModelImageBenchmarkChain-of-Thought
🎯 What it does: Proposed an image editing reward model called EditScore and the corresponding benchmark EditReward-Bench, applying them to Best-of-N selection and online reinforcement learning, significantly enhancing the performance of editing models.
EditVerse: Unifying Image and Video Editing and Generation with In-Context Learning
Xuan Ju (Adobe Research), Qiang Xu (CUHK)
GenerationTransformerDiffusion modelAuto EncoderImageVideo
🎯 What it does: Developed a unified framework called EditVerse that enables a single model to simultaneously perform image and video editing and generation;
EdiVal-Agent: An Object-Centric Framework for Automated, Fine-Grained Evaluation of Multi-Turn Editing
Tianyu Chen (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)
Object DetectionAgentic AIPrompt EngineeringVision Language ModelImageTextBenchmark
🎯 What it does: Proposed EdiVal-Agent, an object-oriented automated evaluation framework for fine-grained assessment of multi-round instructive image editing;
EEPO: Exploration-Enhanced Policy Optimization via Sample-Then-Forget
Liang Chen (Chinese University of Hong Kong), Kam-Fai Wong (Chinese University of Hong Kong)
OptimizationLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: Under the RLVR framework, EEPO is proposed by introducing a temporary 'forgetting' step between two-stage sampling, forcing the model to skip already-occurred high-probability trajectories in subsequent sampling, thereby enhancing exploration.
Efficient Adversarial Attacks on High-dimensional Offline Bandits
Seyed Mohammad Hadi Hosseini (Sharif University of Technology), Mahdieh Soleymani Baghshah (Sharif University of Technology)
OptimizationAdversarial AttackImageBenchmark
🎯 What it does: Studied methods to attack reward models in offline Bandit evaluation by introducing small perturbations, causing Bandit algorithms to incorrectly select suboptimal arms.
Efficient Agent Training for Computer Use
Yanheng He (Shanghai Jiao Tong University), Pengfei Liu (Shanghai Jiao Tong University)
Data SynthesisReinforcement Learning from Human FeedbackTransformerLarge Language ModelAgentic AITextSequentialChain-of-Thought
🎯 What it does: Training a computer to act as an agent can achieve performance surpassing Claude 3.7 Sonnet using only 312 human trajectories expanded via AI synthesis.
Efficient algorithms for Incremental Metric Bipartite Matching
Ritesh Seth (IIIT Delhi), Syamantak Das (IIIT Delhi)
OptimizationComputational EfficiencyImageGraphTabular
🎯 What it does: Propose a deterministic incremental algorithm that maintains an approximate minimum cost bipartite matching in any metric space, supporting online point insertions with an amortized time per update of O~(n(1+δ log²(1/δ) log(nΔ))).
Efficient and Sharp Off-Policy Learning under Unobserved Confounding
Konstantin Hess (LMU Munich), Stefan Feuerriegel (LMU Munich)
Reinforcement LearningBiomedical Data
🎯 What it does: Proposed an efficient and sharp personalized off-policy learning method for unobserved confounding scenarios.
Efficient Approximate Posterior Sampling with Annealed Langevin Monte Carlo
Advait Parulekar (University of Texas at Austin), Sanjay Shakkottai (University of Texas at Austin)
Computational EfficiencyScore-based ModelStochastic Differential Equation
🎯 What it does: Proposed an approximate posterior sampling algorithm based on Annealed Langevin Monte Carlo, with theoretical guarantees of polynomial time complexity.
Efficient Audio-Visual Speech Separation with Discrete Lip Semantics and Multi-Scale Global-Local Attention
Kai Li (Tsinghua University), Xiaolin Hu (Tsinghua University)
Computational EfficiencyKnowledge DistillationRepresentation LearningConvolutional Neural NetworkTransformerAuto EncoderVideoMultimodalityAudio
🎯 What it does: Designed an efficient audio-visual speech separation model named Dolphin, which includes a lightweight DP-LipCoder video encoder and a single-iteration global-local attention (GLA) separator, achieving accurate separation of mixed speech.
Efficient Autoregressive Inference for Transformer Probabilistic Models
Conor Hassan (Aalto University), Luigi Acerbi (University of Helsinki)
Computational EfficiencyTransformerMultimodalityTabularTime Series
🎯 What it does: Proposed a causal autoregressive buffer, enabling Transformer-based probabilistic models to encode context in one pass and cache only the generated targets during subsequent autoregressive inference, achieving efficient joint sampling and density evaluation;
Efficient Best-of-Both-Worlds Algorithms for Contextual Combinatorial Semi-Bandits
Mengmeng Li (EPFL), Daniel Kuhn (EPFL)
Optimization
🎯 What it does: Proposed the optimal 'best two boundaries' algorithm for contextual composite semi-bandits, and accelerated the projection steps of FTRL/OSMD through one-dimensional root search.
Efficient Credal Prediction through Decalibration
Paul Hofman (LMU Munich), Eyke Hüllermeier (LMU Munich)
ClassificationAnomaly DetectionOptimizationExplainability and InterpretabilityComputational EfficiencyTransformerVision Language ModelImageTextBiomedical Data
🎯 What it does: Propose a post-processing, model-agnostic method that generates feasible probability intervals under a relative likelihood budget by applying controlled offset (decalibration) to the logits of a trained model, thereby constructing a credal set to express the model's epistemic uncertainty.
Efficient Degradation-agnostic Image Restoration via Channel-Wise Functional Decomposition and Manifold Regularization
Bin Ren (Mohamed bin Zayed University of Artificial Intelligence), Nicu Sebe (University of Trento)
RestorationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Proposed MIRAGE, an efficient degradation-agnostic image restoration framework;
Efficient Differentiable Contact Model with Long-range Influence
XIAOHAN YE (The University of Hong Kong), zherong pan
OptimizationComputational EfficiencyRobotic IntelligencePhysics Related
🎯 What it does: Proposed a contact model in differentiable rigid body simulations that maintains collision-free, smooth, non-adhesive, and non-vanishing gradients.
Efficient Discriminative Joint Encoders for Large Scale Vision-Language Reranking
Mitchell Keren Taraday (Ben-Gurion University of Negev), Chaim Baskin (Ben-Gurion University of Negev)
RetrievalCompressionComputational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Propose a multimodal retrieval re-ranking model EDJE capable of online efficient inference.
Efficient Ensemble Conditional Independence Test Framework for Causal Discovery
Zhengkang Guan (Zhejiang University), Kun Kuang (Zhejiang University)
Computational EfficiencyBiomedical Data
🎯 What it does: Propose a general framework for set conditional independence testing called E-CIT, which linearizes the computational complexity of original CIT and aggregates p-values using stable distributions.
Efficient Estimation of Kernel Surrogate Models for Task Attribution
Zhenshuo Zhang (Northeastern University), Hongyang R. Zhang (Northeastern University)
Explainability and InterpretabilityComputational EfficiencyImageText
🎯 What it does: Address the task attribution problem by proposing Kernel Surrogate Models (KERNELSM) for efficiently estimating the impact of training tasks on the performance of the target task, and achieving efficient learning without retraining through first-order gradient approximation.
Efficient Learning on Large Graphs using a Densifying Regularity Lemma
Jonathan Kouchly (Technion Israel Institute of Technology), Ron Levie (Technion Israel Institute of Technology)
Computational EfficiencyRepresentation LearningGraph Neural NetworkGraph
🎯 What it does: Designed a low-rank cross-block graph (IBG) to approximate any directed sparse graph, and based on this representation, proposed an IBG-NN with O(N) complexity for node classification, spatiotemporal graph prediction, and knowledge graph completion.
Efficient Message-Passing Transformer for Error Correcting Codes
Seong-Joon Park (POSTECH), Jong-Seon No (Seoul National University)
OptimizationComputational EfficiencyTransformerGraph
🎯 What it does: Designed an efficient message-passing Transformer (EfficientMPT) for error correction in linear block codes.
Efficient Morphology-Control Co-Design via Stackelberg Proximal Policy Optimization
Yanning Dai (King Abdullah University of Science and Technology), Jürgen Schmidhuber
OptimizationRobotic IntelligenceReinforcement LearningBenchmark
🎯 What it does: Proposed a morphology-control co-design framework based on Stackelberg game, named Stackelberg PPO
Efficient Multi-objective Prompt Optimization via Pure-exploration Bandits
Donghao Li (University of Virginia), Jing Yang (University of Virginia)
OptimizationComputational EfficiencyTransformerPrompt EngineeringText
🎯 What it does: This paper views the multi-objective prompt optimization problem as a pure exploration multi-objective Bandit problem, proposing algorithms GENSEC and GENPSI for identifying the best feasible prompt and Pareto set;
Efficient Multimodal Spatial Reasoning via Dynamic and Asymmetric Routing
Yixian Shen (University of Amsterdam), Anuj Pathania (University of Amsterdam)
Computational EfficiencyRepresentation LearningTransformerLarge Language ModelPrompt EngineeringVision Language ModelMultimodalityChain-of-Thought
🎯 What it does: To address the long-sequence computation and memory bottlenecks in multi-modal spatial reasoning, the DARE framework is proposed, which can dynamically prune low-value visual and text tokens at the network depth, reasoning steps, and modal levels.