ICLR 2025 Papers — Page 10
International Conference on Learning Representations · 3704 papers
Diversity Empowers Intelligence: Integrating Expertise of Software Engineering Agents
Kexun Zhang (Salesforce AI Research), Caiming Xiong (Salesforce AI Research)
OptimizationAI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: A framework called DEI (Diversity Empowered Intelligence) is proposed to integrate and collaborate multiple Software Engineering (SWE) agents, thereby enhancing the automatic repair rate of GitHub issues.
Diversity-Rewarded CFG Distillation
Geoffrey Cideron (Google DeepMind), Alexandre Rame
GenerationKnowledge DistillationTransformerReinforcement LearningContrastive LearningTextAudio
🎯 What it does: Online CFG distillation is performed on the text-to-music generation model, and a diversity reward based on RL is added to fine-tune the model, allowing it to maintain high quality and enhance diversity without using CFG; subsequently, dynamic control of the quality-diversity trade-off is achieved through weight linear interpolation.
Divide and Translate: Compositional First-Order Logic Translation and Verification for Complex Logical Reasoning
Hyun Ryu (Korea Advanced Institute of Science and Technology), Eunho Yang (Korea Advanced Institute of Science and Technology)
Large Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A neural-symbolic framework named CLOVER is proposed, which utilizes LLM to decompose sentences through logical dependency structures and gradually translate them into first-order logic formulas, followed by verification using a SAT solver, thereby achieving complex logical reasoning.
DLEFT-MKC: Dynamic Late Fusion Multiple Kernel Clustering with Robust Tensor Learning via Min-Max Optimization
Yi Zhang (National University of Defense Technology), En Zhu (National University of Defense Technology)
OptimizationTabular
🎯 What it does: A multi-kernel clustering method DLEFT-MKC based on dynamic late fusion and tensor learning is proposed, utilizing min-max optimization to achieve clustering consistency and robustness;
Do as I do (Safely): Mitigating Task-Specific Fine-tuning Risks in Large Language Models
Francisco Eiras (University of Oxford), Adel Bibi (Dynamo AI)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper studies the impact of task-specific data fine-tuning on the security of large language models and proposes a hybrid strategy based on secure data corpus rewriting (Paraphrase) to mitigate attacks.
Do as We Do, Not as You Think: the Conformity of Large Language Models
Zhiyuan Weng (Zhejiang University), Wenguan Wang (Zhejiang University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This study investigates the conformity behavior of large language models in multi-agent collaboration scenarios and evaluates its extent by constructing the BENCHFORM benchmark.
Do Contemporary Causal Inference Models Capture Real-World Heterogeneity? Findings from a Large-Scale Benchmark
Haining Yu (Amazon), Yizhou Sun (University of California Los Angeles)
TabularBenchmark
🎯 What it does: A large-scale benchmark evaluation of 16 mainstream CATE estimation models on 12 real RCT datasets is conducted, using observational sampling and a newly defined statistic Q (and its unbiased estimator ˆQ) to assess the mean squared error of the models and rank them without the need for counterfactual ground truth.
Do Deep Neural Network Solutions Form a Star Domain?
Ankit Sonthalia (Tübingen AI Center Universität Tübingen), Seong Joon Oh (Tübingen AI Center Universität Tübingen)
OptimizationSupervised Fine-TuningImage
🎯 What it does: The Star Domain Conjecture is proposed, suggesting that the solution set of deep neural networks forms a star domain rather than a simple convex set, and introduces the Starlight algorithm to find star models.
Do Egocentric Video-Language Models Truly Understand Hand-Object Interactions?
Boshen Xu (Renmin University of China), Qin Jin (Renmin University of China)
RecognitionRetrievalDomain AdaptationTransformerLarge Language ModelContrastive LearningVideoTextBenchmark
🎯 What it does: This study investigates the limitations of first-person perspective video-text pre-training in understanding human-object interactions (HOI) and proposes the EgoHOIBench benchmark to evaluate the model's grasp of fine-grained semantic variations in HOI. Additionally, it introduces the asynchronous contrastive learning objective EgoNCE++, which generates hard negative samples for V2T using LLM or vocabulary, while maintaining noun-centric clustering on the T2V side to enhance the model's ability to distinguish verbs.
Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models
Javier Ferrando (Universitat Politècnica de Catalunya), Neel Nanda (ETH Zürich)
TransformerLarge Language ModelSupervised Fine-TuningAuto EncoderText
🎯 What it does: This paper utilizes Sparse Autoencoders (SAE) to identify linear directions in the representation space of large language models (Gemma 2 2B/9B, Llama 3.1 8B) that can distinguish between known and unknown entities, and verifies that these directions have a causal impact on knowledge rejection and hallucination behaviors in chat models; it also discovers an 'unknown' direction that can differentiate between the model's correct and incorrect responses to queries, indicating internal uncertainty signals within the model.
Do Large Language Models Truly Understand Geometric Structures?
Xiaofeng Wang (Shanghai Jiao Tong University), Rui Wang (Shanghai Jiao Tong University)
RecognitionTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: The GeomRel dataset is proposed, focusing on geometric relationship recognition, evaluating LLM's understanding of geometric structures, and revealing its limitations through experiments.
Do LLM Agents Have Regret? A Case Study in Online Learning and Games
Chanwoo Park (Massachusetts Institute of Technology), Kaiqing Zhang (University of Maryland)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This paper explores the regret performance of pre-trained large language models (LLMs) in online learning and repeated game environments through experimental and theoretical analysis, verifying whether they can behave as no-regret learners. Based on this, an unsupervised Regret-Loss training method is proposed to further enhance the no-regret behavior of LLMs.
Do LLMs ``know'' internally when they follow instructions?
Juyeon Heo (University of Cambridge), Jaya Narain (Apple)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Identify a single dimension related to instruction following within LLMs and enhance adherence rates through representation engineering.
Do LLMs estimate uncertainty well in instruction-following?
Juyeon Heo (University of Cambridge), Jaya Narain (Apple)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: A systematic evaluation of the ability of large language models to estimate uncertainty in instruction-following tasks was conducted, and a dual-version (Controlled and Realistic) controlled benchmark was proposed;
Do LLMs have Consistent Values?
Naama Rozen (Tel Aviv University), Ella Daniel (Tel Aviv University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: Using the Personal Values Questionnaire (PVQ-RR) under various prompts and temperature settings, six types of LLMs generated answers, and a quantitative analysis was conducted on their value priorities and the correlations between values, testing their consistency with human value structures.
Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs
Siyan Zhao (University of California Los Angeles), Kaixiang Lin (Amazon)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: A benchmark called PREFEVAL is proposed to evaluate the ability of large language models to recognize, remember, and follow user preferences in multi-turn dialogues.
Do Mice Grok? Glimpses of Hidden Progress in Sensory Cortex
Tanishq Kumar (Harvard University), Samuel J. Gershman (Harvard University)
Representation LearningSupervised Fine-TuningTime Series
🎯 What it does: The study investigates the continuous representation learning in the olfactory cortex of rodents even after behavioral performance has peaked, and explains this phenomenon through reanalysis of experimental data and the construction of a simple neural network model.
Do Stochastic, Feel Noiseless: Stable Stochastic Optimization via a Double Momentum Mechanism
Tehila Dahan (Technion), Kfir Yehuda Levy
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A dual momentum mechanism is proposed, combining Anytime-SGD and STORM, resulting in a stochastic gradient descent algorithm with stable convergence rates that is insensitive to learning rates, along with its accelerated version.
Do Vision & Language Decoders use Images and Text equally? How Self-consistent are their Explanations?
Letitia Parcalabescu (Heidelberg University), Anette Frank (Heidelberg University)
Explainability and InterpretabilityTransformerVision Language ModelTextMultimodality
🎯 What it does: This paper studies the differences in the utilization of image and text information by visual language model (VLM) decoders when answering and generating explanations, and evaluates their self-consistency.
Do Vision-Language Models Represent Space and How? Evaluating Spatial Frame of Reference under Ambiguities
Zheyuan Zhang (University of Michigan), Ziqiao Ma (University of Michigan)
TransformerPrompt EngineeringVision Language ModelImageText
🎯 What it does: Proposes the COMFORT evaluation framework, which uses synthetic 3D scenes and multilingual prompts to systematically assess the framework bias and consistency of visual-language models in spatial relationship reasoning.
Do WGANs succeed because they minimize the Wasserstein Distance? Lessons from Discrete Generators
Ariel Elnekave (Hebrew University of Jerusalem), Yair Weiss (Hebrew University of Jerusalem)
GenerationData SynthesisOptimizationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This paper systematically studies the essence of WGAN's success by constructing a discrete GAN model and utilizing the computable Wasserstein distance. It proves that when the discriminator is a convolutional network, WGAN actually optimizes the block-level Wasserstein distance rather than the entire image-level distance.
Do You Keep an Eye on What I Ask? Mitigating Multimodal Hallucination via Attention-Guided Ensemble Decoding
Yeongjae Cho (Seoul National University), Sungzoon Cho (Seoul National University)
RecognitionObject DetectionTransformerVision Language ModelImageMultimodality
🎯 What it does: Proposes two untrained decoding strategies, Ensemble Decoding (ED) and FastED, which utilize input images segmented into sub-images and combine attention-weighted logit sets to suppress object hallucinations in LVLM; simultaneously introduces adaptive confidence constraints to calibrate logits in ED.
Dobi-SVD: Differentiable SVD for LLM Compression and Some New Perspectives
Wang Qinsi, Chenfeng Xu (University of California Berkeley)
CompressionOptimizationLarge Language ModelText
🎯 What it does: SVD compression of large language models (LLM) is performed, proposing the Dobi-SVD method with differentiable truncation, IPCA weight updates, and quantization remapping.
DocMIA: Document-Level Membership Inference Attacks against DocVQA Models
Khanh Nguyen (Computer Vision Center, Universitat Autonoma de Barcelona), Dimosthenis Karatzas (Computer Vision Center, Universitat Autonoma de Barcelona)
OptimizationKnowledge DistillationAdversarial AttackTransformerSupervised Fine-TuningMultimodality
🎯 What it does: This study investigates document-level membership inference attacks on the DocVQA model, designing both white-box and black-box attack methods without auxiliary data.
DOCS: Quantifying Weight Similarity for Deeper Insights into Large Language Models
Zeping Min (Alibaba Group Hupan Laboratory Chinese Academy of Sciences), Xinshang Wang (Alibaba Group)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes and evaluates the DOCS metric, which quantifies the similarity between weight matrices of large language models, and reveals similarity patterns among adjacent layers and layer clusters through experiments.
Does Refusal Training in LLMs Generalize to the Past Tense?
Maksym Andriushchenko (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper demonstrates the shortcomings of current LLM refusal training in tense generalization by rewriting harmful requests into the past tense and conducts a systematic evaluation of multiple mainstream LLMs.
Does Safety Training of LLMs Generalize to Semantically Related Natural Prompts?
Sravanti Addepalli (Google DeepMind), Prateek Jain (Google DeepMind)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A new security assessment method is proposed, utilizing 'Response Guided Question Augmentation (ReG-QA)' to generate semantically relevant natural questions on aligned LLMs, detecting the model's robustness in natural scenarios.
Does SGD really happen in tiny subspaces?
Minhak Song (Korea Advanced Institute of Science and Technology), Chulhee Yun (Korea Advanced Institute of Science and Technology)
OptimizationConvolutional Neural NetworkTransformerTabular
🎯 What it does: This study investigates the alignment phenomenon between gradients and Hessian subspaces during the training of deep networks, exploring whether it is possible to train the network solely within the subspace and validating its effectiveness.
Does Spatial Cognition Emerge in Frontier Models?
Santhosh Kumar Ramakrishnan (Apple), Vladlen Koltun
Large Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A benchmark called SPACE is proposed and implemented for systematically evaluating the spatial cognitive abilities of cutting-edge models.
Does Training with Synthetic Data Truly Protect Privacy?
Yunpeng Zhao (National University of Singapore), Jie Zhang (ETH Zurich)
Data SynthesisSafty and PrivacyGenerative Adversarial NetworkImage
🎯 What it does: Systematically evaluate the privacy leakage situations of four methods using synthetic data for training.
DoF: A Diffusion Factorization Framework for Offline Multi-Agent Reinforcement Learning
Chao Li (Xiamen University), Siqi Shen (Xiamen University)
Reinforcement LearningDiffusion model
🎯 What it does: Designed and implemented the DoF (Diffusion Factorization Framework), which decomposes collective diffusion models into individual diffusion models through noise and data factorization, achieving offline multi-agent reinforcement learning.
Domain Guidance: A Simple Transfer Approach for a Pre-trained Diffusion Model
Jincheng Zhong (Tsinghua University), Mingsheng Long (Tsinghua University)
GenerationDomain AdaptationDiffusion modelImage
🎯 What it does: A domain-guided (DoG) transfer learning method is proposed, which fine-tunes a pre-trained diffusion model on the target domain and guides generation during the sampling process.
Don't flatten, tokenize! Unlocking the key to SoftMoE's efficacy in deep RL
Ghada Sokar (Google DeepMind), Pablo Samuel Castro (Google DeepMind)
Convolutional Neural NetworkReinforcement LearningMixture of Experts
🎯 What it does: This study investigates the performance of SoftMoE in deep reinforcement learning and reveals the key role of tokenization on the encoder output.
DON’T STOP ME NOW: EMBEDDING BASED SCHEDULING FOR LLMS
Rana Shahout (Harvard University), Michael Mitzenmacher (Harvard University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: During the inference process of LLM, iterative output length prediction is performed using the embedding information from the internal layers of the model, combined with a limited preemption SRPT scheduling strategy that considers the memory usage of KV cache, in order to reduce the average latency and first response time of interactive LLMs.
Don't Take Things Out of Context: Attention Intervention for Enhancing Chain-of-Thought Reasoning in Large Language Models
Shaotian Yan (Alibaba Cloud Computing), Jieping Ye (Alibaba Cloud Computing)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: This study addresses the reasoning bias problem caused by individual vocabulary in few-shot chain-of-thought (CoT) reasoning and proposes a targeted attention intervention method (FAI) to suppress interference.
DOPL: Direct Online Preference Learning for Restless Bandits with Preference Feedback
GUOJUN XIONG, Srinivas Shakkottai (Texas A&M University)
Recommendation SystemReinforcement Learning
🎯 What it does: This paper proposes the infinite-horizon RMAB model PREF-RMAB, which can only observe preference feedback without scalar rewards, and designs an online algorithm DOPL that can explore, collect preference information, and make direct decisions in an unknown environment, theoretically proving its sublinear regret.
DOTS: Learning to Reason Dynamically in LLMs via Optimal Reasoning Trajectories Search
Murong Yue (George Mason University), Dong Yu (Tencent AI Lab)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: The DOTS method is proposed, allowing LLMs to autonomously plan the optimal reasoning path before answering; by defining atomic reasoning actions, searching for the best action trajectory, and training external or internal planners through supervised fine-tuning, the reasoning ability of LLMs is enhanced.
Doubly Optimal Policy Evaluation for Reinforcement Learning
Shuze Liu, Shangtong Zhang (University of Virginia)
OptimizationReinforcement LearningTabular
🎯 What it does: A dual optimal strategy evaluation method is designed, which includes both an optimal data collection strategy (behavior policy) and an optimal baseline function, capable of significantly reducing variance while maintaining unbiasedness.
Doubly robust identification of treatment effects from multiple environments
Piersilvio De Bartolomeis (ETH Zurich), Fanny Yang (University of Michigan)
Tabular
🎯 What it does: A method named RAMEN is proposed, which utilizes multiple heterogeneous data sources to achieve unbiased estimation of treatment effects (ATE) without a complete causal graph by identifying nodes with invariant conditional distributions.
DPaI: Differentiable Pruning at Initialization with Node-Path Balance Principle
Lichuan Xiang (University of Warwick), Hongkai Wen (University of Warwick)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A differentiable initialization pruning method based on node-path balance, DPaI, is proposed, which directly obtains a sparse subnetwork during the network initialization phase.
DPLM-2: A Multimodal Diffusion Protein Language Model
Xinyou Wang (Nanjing University), Quanquan Gu (ByteDance Research)
Protein Structure PredictionTransformerLarge Language ModelDiffusion modelMultimodalityBiomedical Data
🎯 What it does: DPLM-2 is proposed, a multimodal diffusion language model that can simultaneously model protein sequences and structures, achieving seamless co-generation, folding, reverse folding, and motif scaffold design;
Drama: Mamba-Enabled Model-Based Reinforcement Learning Is Sample and Parameter Efficient
Wenlong Wang (Trinity College Dublin), Vinny Cahill (Trinity College Dublin)
Reinforcement LearningWorld ModelSequentialBenchmark
🎯 What it does: This paper proposes Drama, which utilizes the Mamba-2 state space model to construct a model-based RL world model with 7M parameters, achieving high sample and parameter efficiency on Atari100k.
Draw-and-Understand: Leveraging Visual Prompts to Enable MLLMs to Comprehend What You Want
Weifeng Lin (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
ClassificationObject DetectionSegmentationRetrievalTransformerLarge Language ModelPrompt EngineeringVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the Draw-and-Understand framework, seamlessly integrating visual prompt understanding capabilities into existing multimodal large language models (MLLMs). It supports point, box, and free-form prompts through a Visual Prompt Encoder (VPE) while maintaining original image-level perception.
Dream to Manipulate: Compositional World Models Empowering Robot Imitation Learning with Imagination
Leonardo Barcellona (University of Padua), Efstratios Gavves (University of Amsterdam)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningGaussian SplattingWorld ModelImage
🎯 What it does: The DREMA world model is proposed, utilizing high-resolution Gaussian Splatting and the PyBullet physics engine to construct controllable object-level digital twins, enabling robots to generate valid demonstrations through imagination with minimal demonstrations for imitation learning.
DreamBench++: A Human-Aligned Benchmark for Personalized Image Generation
Yuang Peng (Tsinghua University), Shu-Tao Xia (Tsinghua University)
GenerationTransformerLarge Language ModelPrompt EngineeringImageMultimodalityBenchmarkChain-of-Thought
🎯 What it does: The DREAMBENCH++ benchmark is proposed, utilizing GPT-4o for automated evaluation of the concept retention and prompt following of personalized image generation models.
DreamCatalyst: Fast and High-Quality 3D Editing via Controlling Editability and Identity Preservation
Jiwook Kim (Korea Advanced Institute of Science and Technology), Hyunjung Shim (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisComputational EfficiencyDiffusion modelScore-based ModelNeural Radiance FieldGaussian SplattingPoint CloudMeshStochastic Differential Equation
🎯 What it does: A 3D editing framework based on SDS called DreamCatalyst is proposed, which enables fast and high-quality text-driven editing on NeRF and 3D Gaussian Splatting (3DGS) scenes.
DreamDistribution: Learning Prompt Distribution for Diverse In-distribution Generation
Brian Nlong Zhao (University of Southern California), Yunhao Ge (University of Southern California)
GenerationData SynthesisPrompt EngineeringDiffusion modelImage
🎯 What it does: This paper proposes a method called DreamDistribution based on prompt distribution learning, which enables pre-trained text-to-image diffusion models to learn the distribution of visual attributes from a small number of reference images, thereby generating diverse new instances that are consistent with the reference images.
Dreamweaver: Learning Compositional World Models from Pixels
Junyeob Baek (Korea Advanced Institute of Science and Technology), Sungjin Ahn (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisRecurrent Neural NetworkTransformerWorld ModelVideo
🎯 What it does: Learn composable world models from unsupervised video data, automatically extracting and representing static properties of objects (color, shape) and dynamic properties (motion direction, speed, dance patterns), and generating new future videos based on these abstract concepts.
DRESSing Up LLM: Efficient Stylized Question-Answering via Style Subspace Editing
Xinyu Ma (Peking University), Yasha Wang (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The DRESS method is proposed, utilizing untrained representation editing techniques to enable large language models to output answers in a specified style when responding to questions.
DriveTransformer: Unified Transformer for Scalable End-to-End Autonomous Driving
Xiaosong Jia (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Autonomous DrivingTransformerPoint Cloud
🎯 What it does: A unified Transformer framework called DriveTransformer is designed, integrating perception, prediction, mapping, and planning for E2E autonomous driving tasks, achieving task parallelism, sparse representation, and streaming processing.
DRL: Decomposed Representation Learning for Tabular Anomaly Detection
Hangting Ye (Jilin University), Yi Chang (Jilin University)
Anomaly DetectionRepresentation LearningTabularFinance Related
🎯 What it does: A framework for anomaly detection in tabular data based on Decomposed Representation Learning (DRL) is proposed, which maps the representations of positive samples to a constrained latent space using fixed orthogonal basis vectors, and enhances the distinction between positive and anomalous classes through decomposition loss and separation constraints.
DRoC: Elevating Large Language Models for Complex Vehicle Routing via Decomposed Retrieval of Constraints
Xia Jiang (Eindhoven University of Technology), Yingqian Zhang (Southeast University)
OptimizationTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: The DRoC framework is proposed, which helps large language models (LLMs) generate more accurate solution programs by decomposing vehicle routing problem (VRP) constraints and retrieving external knowledge.
Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization
Taishi Nakamura (Institute of Science Tokyo), Jun Suzuki (Tohoku University)
Mixture of ExpertsText
🎯 What it does: A new MoE model construction method called Drop-Upcycling is proposed, which utilizes pre-trained dense model initialization and partially reinitializes parameters in the expert FFN to achieve a balance between knowledge transfer and expert specialization.
DRoP: Distributionally Robust Data Pruning
Artem M Vysogorets, Julia Kempe (New York University)
ClassificationData-Centric LearningConvolutional Neural NetworkImage
🎯 What it does: A data pruning method based on distribution robustness, DRoP, is proposed. It significantly reduces classification bias while ensuring average performance by assigning pruning ratios to each category based on the validation set error and randomly pruning within each category.
DS-LLM: Leveraging Dynamical Systems to Enhance Both Training and Inference of Large Language Models
Ruibing Song (University of Rochester), Tong Geng (Pacific Northwestern National Laboratory)
TransformerLarge Language ModelText
🎯 What it does: The DS-LLM framework is proposed, which maps large language models to dynamic system (DS) machines based on current dynamics, achieving efficient execution of training and inference.
DSBench: How Far Are Data Science Agents from Becoming Data Science Experts?
Liqiang Jing (University of Texas at Dallas), Dong Yu (Tencent AI Lab)
Large Language ModelAgentic AIMultimodalityTabularBenchmarkFinance Related
🎯 What it does: Proposed and implemented the DSBench data science benchmark, collecting 466 data analysis tasks from ModelOff and 74 data modeling tasks from Kaggle, designed a complete evaluation workflow, and made the data and code publicly available;
DSPO: Direct Score Preference Optimization for Diffusion Model Alignment
Huaisheng Zhu (Pennsylvania State University), Vasant G Honavar (Pennsylvania State University)
GenerationOptimizationReinforcement Learning from Human FeedbackDiffusion modelImageText
🎯 What it does: This paper proposes and implements an algorithm called DSPO, which directly uses score matching for fine-tuning text-to-image diffusion models, aiming to make the generated images more aligned with human preferences.
Dual Process Learning: Controlling Use of In-Context vs. In-Weights Strategies with Weight Forgetting
Suraj Anand (Brown University), Ellie Pavlick (Brown University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study implements a dual-process learning method that coexists structured context learning and weight learning, utilizing active forgetting and temporary forgetting techniques to maintain the model's flexible adaptation to new and old vocabulary.
Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces
DiJia Su (Meta AI), Qinqing Zheng (Meta AI)
TransformerLarge Language ModelSequential
🎯 What it does: Proposes Dualformer, a Transformer model that can switch between fast and slow thinking modes during inference;
DUALFormer: Dual Graph Transformer
Jiaming Zhuo (Hebei University of Technology), Liang Yang (Hebei University of Technology)
ClassificationGraph Neural NetworkTransformerGraph
🎯 What it does: A dual-dimensional graph Transformer named DUALFormer is proposed for node classification tasks.
DUET: Decentralized Bilevel Optimization without Lower-Level Strong Convexity
Zhen Qin (Ohio State University), Jia Liu (Ohio State University)
OptimizationMeta LearningImage
🎯 What it does: This paper proposes a single-loop distributed dual-layer optimization algorithm called DUET, which addresses the reliance of traditional methods on the strong convexity of the lower layer (LLSC) and enables collaborative optimization in the absence of LLSC and under heterogeneous data conditions.
DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Heads
Guangxuan Xiao (Massachusetts Institute of Technology), Song Han (Massachusetts Institute of Technology)
RetrievalCompressionOptimizationComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: The DuoAttention framework is proposed for inference in long-context large language models, which reduces memory and latency by identifying retrieval heads and streaming heads, using full KV caching only for retrieval heads.
Duoduo CLIP: Efficient 3D Understanding with Multi-View Images
Han-Hung Lee (Simon Fraser University), Angel X Chang
ClassificationRetrievalTransformerContrastive LearningImagePoint Cloud
🎯 What it does: The Duoduo CLIP model is proposed, which learns 3D shape encoding from multi-view images and aligns it with text.
Durable Quantization Conditioned Misalignment Attack on Large Language Models
Peiran Dong (Hong Kong Polytechnic University), Song Guo (Hong Kong University of Science and Technology)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a Quantized Misalignment Attack (Q-Misalign), which achieves a jailbreak attack on quantized LLMs by embedding potential misalignments in a full-precision model and activating them after model quantization.
DyCAST: Learning Dynamic Causal Structure from Time Series
Yue Cheng (Beijing Jiaotong University), Zhanxing Zhu (University of Southampton)
Time SeriesOrdinary Differential Equation
🎯 What it does: This paper proposes DyCAST, a framework based on constrained neural ODEs for learning time-evolving causal structures from time series;
DynAlign: Unsupervised Dynamic Taxonomy Alignment for Cross-Domain Segmentation
Han Sun (École Polytechnique Fédérale de Lausanne), Olga Fink (École Polytechnique Fédérale de Lausanne)
SegmentationDomain AdaptationAutonomous DrivingLarge Language ModelImage
🎯 What it does: The DynAlign framework is proposed, achieving automatic alignment of different label spaces and fine-grained segmentation in unsupervised cross-domain segmentation.
DynaMath: A Dynamic Visual Benchmark for Evaluating Mathematical Reasoning Robustness of Vision Language Models
Chengke Zou (University of California Berkeley), Huan Zhang (University of Illinois Urbana-Champaign)
TransformerPrompt EngineeringVision Language ModelTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes and evaluates a dynamic visual mathematics benchmark, DYNAMATH, to test the mathematical reasoning robustness of visual language models under different visual and textual variants.
Dynamic Assortment Selection and Pricing with Censored Preference Feedback
Jung-hun Kim (Seoul National University), Min-hwan Oh (Seoul National University)
Recommendation SystemOptimizationTabular
🎯 What it does: A dynamic multi-product selection and pricing framework based on the truncated multinomial logit (C-MNL) model is proposed, addressing the issue of buyers filtering and only providing sparse feedback on purchases under price decisions.
Dynamic Contrastive Skill Learning with State-Transition Based Skill Clustering and Dynamic Length Adjustment
Jinwoo Choi (Seoul National University), Seung-Woo Seo (Seoul National University)
Robotic IntelligenceReinforcement Learning from Human FeedbackAuto EncoderContrastive LearningSequential
🎯 What it does: This paper proposes a dynamic skill learning framework based on contrastive learning, DCSL, which can learn variable-length, semantically consistent skills from unlabeled offline data.
Dynamic Diffusion Transformer
Wangbo Zhao (National University of Singapore), Yang You (National University of Singapore)
GenerationComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: This paper proposes a Dynamic Diffusion Transformer (DyDiT), which significantly reduces the computational redundancy of DiT by dynamically adjusting the model width and token computation in both time steps and spatial dimensions.
Dynamic Gaussians Mesh: Consistent Mesh Reconstruction from Dynamic Scenes
Isabella Liu (University of California San Diego), Xiaolong Wang (University of California San Diego)
GenerationData SynthesisNeural Radiance FieldVideoPoint CloudMesh
🎯 What it does: By combining 3D Gaussian point clouds with deformable fields, DG-Mesh can reconstruct temporally consistent, high-quality meshes from dynamic videos and track vertex motion, supporting topological changes.
Dynamic Loss-Based Sample Reweighting for Improved Large Language Model Pretraining
Daouda Sow, Yingbin Liang (Ohio State University)
TransformerLarge Language ModelText
🎯 What it does: Proposes an online instance-level loss-based sample reweighting method for large-scale language model pre-training;
Dynamic Low-Rank Sparse Adaptation for Large Language Models
Weizhong Huang (Xiamen University), Rongrong Ji (Xiamen University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Designed and implemented LoSA, a fine-tuning method capable of dynamic low-rank sparse adaptation on sparse LLMs.
Dynamic Mixture of Experts: An Auto-Tuning Approach for Efficient Transformer Models
Yongxin Guo (Chinese University of Hong Kong), Tao Lin (Westlake University)
OptimizationComputational EfficiencyTransformerMixture of ExpertsMultimodality
🎯 What it does: A dynamic mixture of experts (DYNMOE) framework is proposed, which can automatically determine the number of experts and the number of experts to be activated for each token during the training process.
Dynamic Modeling of Patients, Modalities and Tasks via Multi-modal Multi-task Mixture of Experts
Chenwei Wu (University of Michigan), Liyue Shen (University of Michigan)
ClassificationSegmentationTransformerMixture of ExpertsMultimodalityBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes M4oE, a multimodal multi-task mixture of experts framework that enables sample-adaptive dynamic modality fusion and task-specific modality fusion.
Dynamic Multimodal Evaluation with Flexible Complexity by Vision-Language Bootstrapping
Yue Yang (Shanghai AI Laboratory), Wenqi Shao (University of Hong Kong)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: This paper proposes a dynamic multimodal evaluation protocol called Vision-Language Bootstrapping (VLB), which generates new evaluation samples through adaptive enhancement of images and text, aiming to reduce data contamination and increase evaluation complexity.
Dynamic Negative Guidance of Diffusion Models
Felix Koulischer (Ghent University), Luca Ambrogioni (Radboud University)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A Dynamic Negative Guidance method is proposed for negative prompts in diffusion models, which can dynamically adjust the guidance strength during the generation process based on posterior probabilities, and suppress unwanted features without the need for additional training.
Dynamic Neural Fortresses: An Adaptive Shield for Model Extraction Defense
Siyu Luan (University of Copenhagen), Dacheng Tao (Shenzhen Campus of Sun Yat-sen University)
Safty and PrivacyComputational EfficiencyConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: Proposes Dynamic Neural Fortresses (DNF), which allows attack queries to randomly exit early in the model, reducing computational load and suppressing information leakage.
Dynamic Sparse Training versus Dense Training: The Unexpected Winner in Image Corruption Robustness
Boqian Wu (University of Twente), Elena Mocanu (University of Twente)
ClassificationRecognitionConvolutional Neural NetworkImageVideo
🎯 What it does: This study investigates the robustness performance of Dynamic Sparse Training (DST) on image and video classification tasks at different levels of sparsity, comparing it with traditional Dense Training.
Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification
Wenxuan Huang (East China Normal University), Shaohui Lin (Key Laboratory of Advanced Theory and Application in Statistics and Data Science)
GenerationComputational EfficiencyTransformerLarge Language ModelVision Language ModelTextMultimodality
🎯 What it does: The Dynamic-LLaVA framework is proposed to achieve dynamic visual-language context sparsification for multi-modal large language models (MLLM), significantly reducing computational and memory overhead during inference.
Dynamic-SUPERB Phase-2: A Collaboratively Expanding Benchmark for Measuring the Capabilities of Spoken Language Models with 180 Tasks
Chien-yu Huang (National Taiwan University), Hung-yi Lee (National Taiwan University)
ClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningMultimodalityBenchmarkAudio
🎯 What it does: Released and maintained the Dynamic-SUPERB Phase-2 evaluation benchmark, which includes 180 instruction-based tasks covering speech, music, and general audio, and established a task classification system and automated evaluation process; simultaneously conducted a systematic evaluation of various public models on this benchmark.
Dynamical Diffusion: Learning Temporal Dynamics with Diffusion Models
Xingzhuo Guo (Tsinghua University), Mingsheng Long (Tsinghua University)
Diffusion modelVideoTime Series
🎯 What it does: A new framework called Dynamical Diffusion (DyDiff) is proposed for learning temporal dynamics, particularly applying diffusion models in time prediction tasks.
DynamicCity: Large-Scale 4D Occupancy Generation from Dynamic Scenes
Hengwei Bian (Shanghai AI Laboratory), Ziwei Liu (Nanyang Technological University)
GenerationData SynthesisAutonomous DrivingTransformerDiffusion modelAuto EncoderPoint Cloud
🎯 What it does: The DynamicCity framework is proposed to achieve large-scale, high-quality 4D (spatiotemporal) occupancy scene generation, supporting various conditional controls and post-processing.
DynaPrompt: Dynamic Test-Time Prompt Tuning
Zehao Xiao (AIM Lab, University of Amsterdam), Cees G. M. Snoek (AIM Lab, University of Amsterdam)
Domain AdaptationTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: This paper proposes Dynamic Prompt Tuning (DynaPrompt), which enhances the zero-shot generalization performance of the CLIP model under distribution shifts by adaptively selecting, updating, and appending prompts through an online prompt buffer.
DynFrs: An Efficient Framework for Machine Unlearning in Random Forest
Shurong Wang (Zhejiang University), Meng Zhang (Zhejiang University)
ClassificationComputational EfficiencyTabular
🎯 What it does: The DYNFRS framework is proposed, achieving efficient random forest machine forgetting.
Dysca: A Dynamic and Scalable Benchmark for Evaluating Perception Ability of LVLMs
Jie Zhang (Chinese Academy of Sciences), Xilin Chen (Chinese Academy of Sciences)
Data SynthesisAdversarial AttackTransformerVision Language ModelDiffusion modelImageTextMultimodalityBenchmark
🎯 What it does: A dynamic and scalable benchmark named Dysca has been developed to evaluate the perceptual capabilities of large visual language models (LVLMs), which includes 20 perceptual sub-tasks, 4 types of image scenes (Clean, Corruption, Print-Attacking, Adversarial-Attacking), and 3 types of question formats (multiple choice, true/false, open-ended), generating a total of 617K visual-text question-answer pairs.
E(3)-equivariant models cannot learn chirality: Field-based molecular generation
Alexandru Dumitrescu (Aalto University), Harri Lähdesmäki (Aalto University)
GenerationDrug DiscoveryGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A field-based molecular generation model (FMG) is proposed, demonstrating that traditional E(3) equivariant models fail to capture the issue of molecular chirality (chiral asymmetry).
E(n) Equivariant Topological Neural Networks
Claudio Battiloro (Harvard University), Francesca Dominici (Harvard University)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes E(n)-Equivariant Topological Neural Networks (ETNN), a message-passing network that implements E(n) (rotations, reflections, and translations in Euclidean space) equivariance on combinatorial complexes, capable of utilizing both geometric features and higher-order interactions simultaneously.
Eagle: Exploring The Design Space for Multimodal LLMs with Mixture of Encoders
Min Shi (Georgia Tech), Guilin Liu (NVIDIA)
TransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: The system evaluates and constructs a multi-modal large language model (MLLM) framework called Eagle, which combines multiple visual encoders and enhances the model's perceptual capabilities by gradually adding visual experts and pre-alignment techniques.
Earlier Tokens Contribute More: Learning Direct Preference Optimization From Temporal Decay Perspective
Ruichen Shao (Meituan Inc), Peng Li (Institute of Software, Chinese Academy of Sciences)
Recommendation SystemOptimizationTransformerSupervised Fine-TuningReinforcement LearningText
🎯 What it does: An improved method (D2PO) is proposed by incorporating a time decay factor γ into Direct Preference Optimization (DPO), allowing early tokens to have a greater weight during the alignment process through dynamic weighting, thereby reducing length bias and enhancing the model's alignment effectiveness.
Easing Training Process of Rectified Flow Models Via Lengthening Inter-Path Distance
Xu Shifeng, Adams Wai-Kin Kong
GenerationComputational EfficiencyDiffusion modelRectified FlowImage
🎯 What it does: A method is proposed to accelerate the training of Rectified Flow and Diffusion models by extending the distance between noise and sample paths.
EC-Diffuser: Multi-Object Manipulation via Entity-Centric Behavior Generation
Carl Qi (University of Texas at Austin), Amy Zhang (Meta AI)
Robotic IntelligenceTransformerDiffusion modelImage
🎯 What it does: This paper proposes EC-Diffuser, a behavior cloning method based on diffusion models, which uses unsupervised object-centric representations (Deep Latent Particles, DLP) and an entity-centric Transformer to generate future particle states and actions, enabling multi-object manipulation.
EC-DIT: Scaling Diffusion Transformers with Adaptive Expert-Choice Routing
Haotian Sun (Georgia Institute of Technology), Nan Du (Apple AI/ML)
GenerationTransformerMixture of ExpertsDiffusion modelRectified FlowImageText
🎯 What it does: Proposes EC-DIT, a sparse Mixture-of-Experts (MoE) model using expert-choice routing in the Diffusion Transformer (DiT);
ECD: A Machine Learning Benchmark for Predicting Enhanced-Precision Electronic Charge Density in Crystalline Inorganic Materials
Pin Chen (Sun Yat-sen University), Yutong Lu (Sun Yat-sen University)
Graph Neural NetworkTabularBenchmarkPhysics Related
🎯 What it does: This paper constructs a public dataset ECD containing 140,646 PBE-precision crystal structures and 7,147 HSE high-precision electronic charge density data, and predicts charge density through machine learning methods, significantly accelerating DFT calculations.
ECHOPulse: ECG Controlled Echocardio-gram Video Generation
Yiwei Li (Massachusetts General Hospital and Harvard Medical School), Xiang Li (Massachusetts General Hospital and Harvard Medical School)
GenerationData SynthesisTransformerGenerative Adversarial NetworkVideoMultimodalityTime SeriesBiomedical DataUltrasoundElectrocardiogram
🎯 What it does: A cardiac ultrasound (ECHO) video generation model called ECHOPulse has been developed based on ECG time series signals, capable of quickly generating high-quality, ECG-synchronized cardiac ultrasound videos without expert annotations.
EcoFace: Audio-Visual Emotional Co-Disentanglement Speech-Driven 3D Talking Face Generation
Jiajian Xie (Zhejiang University), Fei Wu (Zhejiang University)
GenerationContrastive LearningVideoAudio
🎯 What it does: The EcoFace framework is proposed to achieve voice-driven 3D facial animation, enhancing the realism of expressions and lip synchronization quality through audio-visual collaborative emotional decoupling.
econSG: Efficient and Multi-view Consistent Open-Vocabulary 3D Semantic Gaussians
Can Zhang (National University of Singapore), Gim Hee Lee (National University of Singapore)
Object DetectionSegmentationComputational EfficiencyAuto EncoderGaussian SplattingPoint Cloud
🎯 What it does: The econSG model is proposed, achieving an efficient and multi-view consistent open vocabulary three-dimensional semantic Gaussian field. The CRR method integrates OpenSeg and SAM to obtain fine semantic features and accelerates training through a low-dimensional 3D context space.
Edge Prompt Tuning for Graph Neural Networks
Xingbo Fu (University of Virginia), Jundong Li (University of Virginia)
ClassificationGraph Neural NetworkPrompt EngineeringGraph
🎯 What it does: Proposes two edge-based graph prompt tuning methods, EdgePrompt and EdgePrompt+, for adapting downstream tasks on pre-trained GNN models without the need for fine-tuning.
Edge-aware Image Smoothing with Relative Wavelet Domain Representation
Huiqing QI, Fang Li (East China Normal University)
RestorationImage
🎯 What it does: A mutual guidance edge-aware image smoothing model based on relative wavelet domain representation has been designed and implemented.
EdgeRunner: Auto-regressive Auto-encoder for Artistic Mesh Generation
Jiaxiang Tang (Peking University), Qinsheng Zhang (NVIDIA Research)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderPoint CloudMesh
🎯 What it does: This paper proposes EdgeRunner, which combines grid-oriented efficient tessellation, an autoregressive autoencoder (ArAE), and a latent diffusion model to generate high-quality and diverse artistic meshes (up to 4000 faces and 512 3D resolution) from point clouds or single-view images.