π― What it does: Propose the D-AR framework, which reformulates the pixel-level diffusion process as a standard next-token prediction task; design an ordered diffusion tokenizer to map images into 1D discrete token sequences, then use the Llama decoder to generate these tokens, instantly decoding them into diffusion steps during generation to achieve streaming pixel generation.
π― What it does: Propose a mathematics reasoning framework DAG-MATH based on task-specific directed acyclic graphs (DAGs), and define the metrics of logical proximity and perfect reasoning rate.
Dancing in Chains: Strategic Persuasion in Academic Rebuttal via Theory of Mind
Zhitao He (Hong Kong University of Science and Technology), Yi R. Fung (Hong Kong University of Science and Technology)
CodeTransformerLarge Language ModelReinforcement LearningTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Proposes a rebuttal framework called RebuttalAgent based on Theory of Mind, achieving hierarchical modeling of reviewer intentions, strategy formulation, and evidence-driven response generation.
DARE-bench: Evaluating Modeling and Instruction Fidelity of LLMs in Data Science
Fan Shu (University of Houston), Feng Yan (University of Houston)
CodeData SynthesisData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTabularTime SeriesBenchmark
π― What it does: Built and released DARE-Bench, an executable and trainable multi-task data science benchmark for evaluating LLMs' performance in instruction following and machine learning modeling;
Data-Aware and Scalable Sensitivity Analysis for Decision Tree Ensembles
Namrita Varshney (Indian Institute of Technology Bombay), S. Akshay (Indian Institute of Technology Bombay)
CodeClassificationOptimizationExplainability and InterpretabilityImageTabularBenchmark
π― What it does: For decision tree ensemble models, this paper proposes a data-aware sensitivity analysis method to find feature subset examples near the training distribution that can alter model predictions.
π― What it does: Investigated how data memorized by the teacher model leaks to the student model when using soft labels in model distillation, i.e., whether the student can achieve non-random accuracy on unseen memorized samples.
DAVE: A VLM Vision Encoder for Document Understanding and Web Agents
Brandon Huang (MIT-IBM Watson AI Lab), Roei Herzig (MIT-IBM Watson AI Lab)
CodeRecognitionSegmentationTransformerSupervised Fine-TuningReinforcement LearningVision Language ModelVision-Language-Action ModelAuto EncoderImageMultimodalityBenchmark
π― What it does: DAVE is a visual encoder specifically designed for document and web understanding, employing a two-stage self-supervised and supervised autoregressive pre-training.
DaVinci: Reinforcing Visual-Structural Syntax in MLLMs for Generalized Scientific Diagram Parsing
Xingchen ZENG, Wei Zeng (Central South University)
CodeImage TranslationLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageMultimodality
π― What it does: This paper proposes DaVinci, a multimodal large language model based on a two-stage framework (supervised learning visual primitives + reinforcement learning structural relationships), for parsing scientific chart images into editable TikZ code.
π― What it does: Proposed a centralized core token aggregation-redistribution (CoTAR) module to replace the decentralized attention in Transformers, aiming to better capture channel dependencies in medical time series signals.
Decoding Open-Ended Information Seeking Goals from Eye Movements in Reading
Cfir Avraham Hadar (Technion Israel Institute of Technology), Yevgeni Berzak (Technion Israel Institute of Technology)
CodeRetrievalExplainability and InterpretabilityRepresentation LearningData-Centric LearningRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextMultimodality
π― What it does: The study automatically infers the specific information retrieval goals readers are pursuing while reading paragraphs by analyzing their eye movement trajectories, and proposes two decoding tasks: goal selection and goal reconstruction.
Decomposed Attention Fusion in MLLMs for Training-free Video Reasoning Segmentation
Su Ho Han (Yonsei University), Seon Joo Kim (Inha University)
CodeSegmentationTransformerPrompt EngineeringVision Language ModelVideoMultimodality
π― What it does: This paper proposes a training-free multi-modal large language model (MLLM) video reasoning segmentation framework named DecAF, which achieves fine-grained mask generation through attention convolution fusion and SAM2-guided refinement;
Decomposing Representation Space into Interpretable Subspaces with Unsupervised Learning
Xinting Huang, Michael Hahn (Saarland University)
CodeExplainability and InterpretabilityRepresentation LearningTransformerText
π― What it does: Propose an unsupervised neighborhood distance minimization method to decompose the representation space of neural networks into interpretable multidimensional subspaces.
Deconstructing Positional Information: From Attention Logits to Training Biases
Zihan Gu (Chinese Academy of Sciences), Yue Hu (Chinese Academy of Sciences)
CodeExplainability and InterpretabilityRepresentation LearningTransformerText
π― What it does: Investigated the mechanism of position encoding (PE) in Transformers, constructing a unified analytical framework based on Toeplitz matrices, distinguishing additive and multiplicative PE, and revealing RoPE's single-head deposition pattern through synthetic tasks and head ablation experiments.
π― What it does: Propose decomposing the DMD training objective into two parts: CFG Enhancement (CA) as the core engine for few-step generation, and Distribution Matching (DM) as a regularization shield for stable training, and design a decoupled denoising schedule to improve performance.
π― What it does: Proposes a theoretical framework for decoupling class labels from target concepts in machine learning, and defines three new class-level forgetting tasks (target mismatch, model mismatch, data mismatch) based on this framework.
Deep Hierarchical Learning with Nested Subspace Networks for Large Language Models
Paulius Rauba (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Propose Nested Subspace Networks (NSNs), which use adjustable-rank low-rank decomposition within linear layers, allowing a single model to dynamically switch between different ranks during inference based on computational budget, thus achieving a continuous performance-computation trade-off.
Yichao Fu (University Of California San Diego), Jiawei Zhao (Meta Ai)
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: Propose the DeepConf method, which utilizes the internal confidence of LLMs to dynamically filter low-quality reasoning trajectories, thereby enhancing reasoning efficiency and accuracy.
π― What it does: Proposes a completely gradient-agnostic deep analytical federated learning framework, DeepAFL, which utilizes residual blocks to achieve multi-layer representation learning.
π― What it does: Developed the DeepCompress framework, dynamically adjusting the Chain-of-Thought length of large-scale inference models, leveraging dual length rewards and model-aware difficulty to achieve higher accuracy and more efficient inference.
DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning
Ziwei Zheng (Xiaohongshu Inc), XingYu
CodeReinforcement LearningVision Language ModelMultimodalityChain-of-Thought
π― What it does: Train a visual-language model called DeepEyes that can proactively 'look at images' and make cropping decisions during reasoning, using end-to-end reinforcement learning without requiring pre-collected reasoning data.
CodeClassificationConvolutional Neural NetworkContrastive LearningTime Series
π― What it does: Proposed an end-to-end deep learning framework called DeepFRC for simultaneously accomplishing functional registration (alignment) and classification;
DeepMath-103K: A Large-Scale, Challenging, Decontaminated, and Verifiable Mathematical Dataset for Advancing Reasoning
Zhiwei He (Tencent), Dong Yu (Tencent)
CodeData-Centric LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: Proposed and released the DeepMath-103K dataset, and trained a series of LLMs (DeepMath series models) on this dataset, which can be further enhanced through RL and SFT, achieving state-of-the-art performance on mathematical and cross-domain reasoning tasks.
π― What it does: Propose DeepPrim, a deep physics-driven model for short-term weather forecasting by learning the primitive equations of the Earth's atmosphere.
π― What it does: This paper proposes the DeepRAG framework, modeling retrieval-augmented reasoning as a Markov decision process, enabling demand-based progressive retrieval and reasoning;
DeepResearch Bench: A Comprehensive Benchmark for Deep Research Agents
Mingxuan Du (University of Science and Technology of China), Zhendong Mao (University of Science and Technology of China)
CodeRetrievalLarge Language ModelTextBenchmark
π― What it does: Proposed DeepResearch Bench, a benchmark containing 100 PhD-level research tasks across 22 disciplines, to evaluate the report generation and information retrieval capabilities of Deep Research Agents (DRAs).
DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
Yuanpeng Zhang (Central South University), Lei Deng (Institute for Infocomm Research, A*STAR)
CodeDomain AdaptationDrug DiscoveryGraph Neural NetworkTransformerSupervised Fine-TuningAuto EncoderGraphTabularBiomedical Data
π― What it does: Propose the DeepSADR model, which constructs an interaction graph using drug substructures and gene function subsequences, and implements deep transfer learning to predict drug efficacy in cancer patients.
CodeOptimizationTransformerLarge Language ModelTextBenchmark
π― What it does: Built a LLM multi-agent system called DeepScientist, capable of autonomously completing the full scientific discovery process from conception to experimental validation on a monthly time scale, achieving results exceeding human state-of-the-art (SOTA) in three cutting-edge AI tasks.
π― What it does: Propose the DeepSearch framework, which directly embeds Monte Carlo Tree Search (MCTS) into the training loop of reinforcement learning with verifiable rewards (RLVR) to systematically enhance the model's reasoning capabilities through search.
π― What it does: Propose and implement a deep generative model called DeepWeightFlow based on Flow Matching, which directly generates complete and high-performance neural network weights in the weight space. It supports multiple architectures (MLP, ResNet, ViT, BERT) and can be scaled up to 100M parameters through Canonicalization (Git Re-Basin, TransFusion) and PCA.
Defending against Backdoor Attacks via Module Switching
Weijun Li (Macquarie University), Qiongkai Xu (Macquarie University)
CodeSafty and PrivacyAdversarial AttackData-Centric LearningNeural Architecture SearchImageText
π― What it does: This paper proposes a module switching defense (MSD) mechanism to counter backdoor attacks on deep learning models in post-training environments.
DefensiveKV: Taming the Fragility of KV Cache Eviction in LLM Inference
Yuan Feng (University of Science and Technology of China), Xike Xie (University of Science and Technology of China)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelAgentic AITextBenchmark
π― What it does: Proposed and implemented defensive aggregation for KV cache eviction, and built DefensiveKV and its hierarchical version Layer-DefensiveKV to significantly compress the cache while maintaining generation quality;
π― What it does: Redefine reference video segmentation (RVOS) as a continuous flow problem conditioned on text, and propose the end-to-end FlowRVS framework that directly transforms the video's implicit representation into masks.
Delta-XAI: A Unified Framework for Explaining Prediction Changes in Online Time Series Monitoring
Changhun Kim (AITRICS), Eunho Yang (AITRICS)
CodeAnomaly DetectionExplainability and InterpretabilityConvolutional Neural NetworkRecurrent Neural NetworkTransformerTime SeriesBiomedical DataElectronic Health Records
π― What it does: This paper proposes a unified framework Delta-XAI to explain the reasons behind prediction changes in online time series monitoring, and designs new evaluation metrics for this task;
π― What it does: Proposes Decoupled Momentum Optimization (DeMo), a framework that significantly reduces communication volume in distributed training while maintaining the same convergence performance as AdamW.
Demystifying and Enhancing the Efficiency of Large Language Model Based Search Agents
Tiannuo Yang (University Of Southern California), Willie Neiswanger (Nankai University)
CodeRetrievalOptimizationComputational EfficiencyLarge Language ModelTextBenchmarkRetrieval-Augmented Generation
π― What it does: This paper proposes an efficient inference framework called SearchAgent-X specifically designed for large language model (LLM) search agents, aiming to optimize the interaction process between alternating reasoning and retrieval.
π― What it does: This paper proposes a denoising neural re-ranker (DNR) for multi-stage recommendation systems, which enhances re-ranking performance by adding noise to and denoising the scores generated by the retriever.
DenseGRPO: From Sparse to Dense Reward for Flow Matching Model Alignment
Haoyou Deng (Huazhong University of Science and Technology), Nong Sang (Huazhong University of Science and Technology)
CodeGenerationReinforcement Learning from Human FeedbackReinforcement LearningScore-based ModelFlow-based ModelImageTextMultimodalityStochastic Differential EquationOrdinary Differential Equation
π― What it does: Propose DenseGRPO, which utilizes ODE inference to estimate dense rewards at each step and adaptively adjusts the noise intensity in SDE sampling, thereby achieving more precise reinforcement learning alignment in text-to-image flow matching models.
π― What it does: This paper proposes the HyperFedZero method, achieving seamless migration of the global model to non-participating clients without fine-tuning in federated learning, enabling zero-shot personalization.
Designing Time Series Experiments in A/B Testing with Transformer Reinforcement Learning
Xiangkun Wu (Zhejiang University), Chengchun Shi (London School of Economics and Political Science)
CodeOptimizationTransformerReinforcement LearningTime Series
π― What it does: Proposes a method for time series A/B test experiment design using Transformer-based reinforcement learning, aiming to minimize the mean squared error (MSE) of the average treatment effect (ATE) estimation.
Detecting Data Contamination from Reinforcement Learning Post-training for Large Language Models
Yongding Tao (Peking University), Ge Li (Peking University)
CodeAnomaly DetectionData-Centric LearningReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: This paper addresses the data pollution problem in large language models (LLM) during the reinforcement learning (RL) post-training phase, proposing a self-critique-based entropy similarity detection method called Self-Critique, and constructing the RL-MIA benchmark for systematic evaluation.
Detecting Data Contamination in LLMs via In-Context Learning
MichaΕ Zawalski (NVIDIA), Pablo Ribalta (NVIDIA)
CodeAnomaly DetectionData-Centric LearningTransformerLarge Language ModelPrompt EngineeringText
π― What it does: Propose the CoDeC method, which utilizes adding same-dataset samples in the context to detect training data contamination in large language models and quantifies the extent of contamination.
Lukas Eisenmann (Heidelberg University), Daniel Durstewitz (Heidelberg University)
CodeRecurrent Neural NetworkTime Series
π― What it does: Propose a semi-analytical algorithm based on ReLU RNN (PLRNN) to compute stable and unstable invariant manifolds, which can be used for partitioning basins of attraction, identifying similar/dissimilar orbits, and detecting chaos.
Detecting Misbehaviors of Large Vision-Language Models by Evidential Uncertainty Quantification
Tao Huang (State Key Laboratory of Advanced Rail Autonomous Operation), Liping Jing (Beijing Key Laboratory of Security and Privacy in Intelligent Transportation)
CodeAnomaly DetectionVision Language ModelMultimodalityBenchmark
π― What it does: By analyzing errors generated by large vision-language models (LVLM), we attribute them to internal conflicts and missing information, and propose a training-free evidence uncertainty quantification method based on the Dempster-Shafer theory.
π― What it does: To address the performance degradation in autonomous driving multi-modal fusion caused by temporal misalignment, this paper proposes a lightweight defense framework called AION to detect and suppress temporal misalignment attacks (TMA).
π― What it does: Propose a zero-shot multivariate time series anomaly detection method SPIE-AD based on physical dynamics model recovery and consistency reasoning, designed to identify unknown unknowns (U2) errors in autonomous systems.
π― What it does: Built upon Grounding DINO, the VIS-GDINO model was developed to support visual prompts, and further improved to DETR-ViP, enabling more accurate utilization of visual prompts for open-vocabulary object detection.
π― What it does: Proposed DGNet, a graph neural network leveraging the discrete Green's formula, for efficiently learning spatiotemporal PDE solutions under data-scarce conditions and achieving zero-shot generalization on unseen source terms.
π― What it does: Designed and released DHG-Bench, a comprehensive benchmark that evaluates the performance of 17 HNNs on 22 hypergraph datasets using a unified experimental protocol, covering node-level, edge-level, and graph-level tasks, and exploring effectiveness, efficiency, robustness, and fairness.
π― What it does: This paper proposes the DiaBlo method, which achieves parameter-efficient fine-tuning by only fine-tuning the diagonal blocks of the pre-trained model's weight matrix;
Diagnosing and Remedying Knowledge Deficiencies in LLMs via Label-free Curricular Meaningful Learning
Kai Xiong (Research Center for Social Computing and Interactive Robotics), Ting Liu (Research Center for Social Computing and Interactive Robotics)
CodeData SynthesisKnowledge DistillationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Automatically diagnose LLM knowledge defects through unlabeled user queries and improve them using curriculum-based meaningful learning
π― What it does: Designed and implemented an adaptive caching strategy called DiCache based on online shallow-layer probing to accelerate Diffusion model inference.
DiffAdapt: Difficulty-Adaptive Reasoning for Token-Efficient LLM Inference
Xiang Liu (Hong Kong University of Science and Technology), Eunsol Choi (New York University)
CodeComputational EfficiencyTransformerLarge Language ModelTextBenchmark
π― What it does: Analyze the overthinking phenomenon in LLM inference and propose DiffAdapt, a difficulty-based adaptive reasoning strategy framework, using a lightweight detector to dynamically select easy/normal/hard reasoning modes, reducing token consumption while improving accuracy
π― What it does: This paper proposes a differentiable graph lifting framework, DiffLift, which can automatically learn the optimal lifting to transform graphs into higher-order topological structures (such as hypergraphs, cell complexes, or simplicial complexes) during end-to-end training.
Differentiable Model Predictive Control on the GPU
Emre Adabag (Toyota Research Institute), Thomas Jonathan Lew (Toyota Research Institute)
CodeAutonomous DrivingOptimization
π― What it does: Developed a GPU-based differentiable model predictive control solver called DiffMPC, and verified its performance in tasks such as reinforcement learning, imitation learning, and car drifting.
π― What it does: Propose DiffMJX and CFD two techniques, using differentiable simulation to solve the problem of inaccurate gradients under hard contact and zero gradients for non-contact objects, and verify them in parameter identification and model predictive control (MPC).
DifficultyβDiversity Collaborative Filtering for Data-Efficient LLM Fine-Tuning
Long P. Hoang (Singapore University of Technology and Design), Wei Lu (Nanyang Technological University)
CodeData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: This paper proposes an automated method to select high-quality training subsets tailored for target LLMs, significantly reducing annotation costs while maintaining fine-tuning performance comparable to full-data training.
DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation
Shansan Gong (Apple), Yizhe Zhang (Apple)
CodeAI Code AssistantTransformerLarge Language ModelReinforcement LearningDiffusion modelText
π― What it does: Trained a 7B-parameter Masked Diffusion Language Model (DiffuCoder) for code generation, and conducted a systematic analysis of its decoding behavior;
π― What it does: Modify the object query initialization and training process of DETR to a denoising diffusion process, proposing DiffuDETR and DiffuDINO to achieve generation of precise boxes from noise.
DiffuGuard: How Intrinsic Safety is Lost and Found in Diffusion Large Language Models
Zherui Li (Beijing University of Posts and Telecommunications), Jiaheng Zhang (National University of Singapore)
CodeSafty and PrivacyLarge Language ModelDiffusion modelText
π― What it does: This paper conducts a two-dimensional (intra-step and inter-step) safety analysis of the iterative reasoning process in diffusion large language models (dLLMs), revealing safety biases caused by greedy remasking and the 'denoising path dependency' phenomenon. Based on this, it proposes an untrained defense framework called DIFFUGUARD, which includes two modules: stochastic annealing remasking and block-level audit and repair.
π― What it does: Proposed and theoretically analyzed a unified iterative proportional Markov fitting (IPMF) framework, revealing that bidirectional IMF is equivalent to alternating between IPF and IMF projection, and proving its exponential/weak convergence under Gaussian distribution and bounded support scenarios;
Diffusion Alignment as Variational Expectation-Maximization
Jaewoo Lee (KAIST), Jinkyoo Park (Omelet)
CodeGenerationReinforcement LearningDiffusion modelImageBiomedical Data
π― What it does: Designed and verified a diffusion model alignment framework DAV based on variational expectation maximization, aiming to maximize rewards while maintaining diversity in image generation and DNA sequence design.
π― What it does: Proposes two copula models based on diffusion and flowβclassification diffusion copula and reflection copula. Designs a normal OU diffusion process and reflection process that only forgets dependencies between variables while maintaining uniform margins, and achieves density estimation and sampling by learning dependencies during the memory process.
π― What it does: This paper proposes a KL-regularized reinforcement learning framework called SQDF, which directly utilizes differentiable rewards to fine-tune diffusion models via reparameterized policy gradients, avoiding over-optimization while maintaining sample diversity.
Diffusion Language Model Knows the Answer Before It Decodes
Pengxiang Li (Hong Kong Polytechnic University), Shiwei Liu (ELLIS Institute)
CodeComputational EfficiencyLarge Language ModelDiffusion modelText
π― What it does: Proposed a no-training fast decoding strategy called Prophet, leveraging the phenomenon that diffusion language models (DLMs) can converge early in the decoding process, enabling early submission decoding.
Diffusion Negative Preference Optimization Made Simple
Joshua Tian Jin Tee (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
CodeGenerationDiffusion modelImage
π― What it does: Propose a single-model Diff-SNPO for negative preference optimization, addressing computational costs and weakened contrast caused by dual-model training and weight merging.
Diffusion-DFL: Decision-focused Diffusion Models for Stochastic Optimization
Zihao Zhao (Georgia Institute of Technology), Kai Wang (Georgia Institute of Technology)
CodeOptimizationDiffusion modelScore-based ModelTabularFinance Related
π― What it does: This paper proposes embedding diffusion generative models into decision-focused learning (DFL) to learn the complete uncertain parameter distribution and optimize stochastic decisions in an end-to-end manner.
π― What it does: Propose the DiffusionBlocks framework, which splits residual networks such as Transformers into independently trainable blocks, utilizing diffusion theory to achieve gradient-free dependencies between blocks.
Directed Semi-Simplicial Learning with Applications to Brain Activity Decoding
Manuel Lecha (Istituto Italiano di Tecnologia), Claudio Battiloro (Harvard University)
CodeClassificationGraph Neural NetworkGraphBiomedical Data
π― What it does: Proposed and implemented Semi-Simplicial Neural Networks (SSNs) for semi-simplicial sets to learn directed high-order network structures.
Directional Textual Inversion for Personalized Text-to-Image Generation
Kunhee Kim (KAIST), Hyunjung Shim (KAIST)
CodeGenerationTransformerPrompt EngineeringVision Language ModelDiffusion modelImageText
π― What it does: Proposes the Directional Text Inversion (DTI) method to achieve personalization and improve prompt fidelity in text-to-image generation.
Disco: Densely-overlapping Cell Instance Segmentation via Adjacency-aware Collaborative Coloring
Rui Sun (Shanghai Academy of Artificial Intelligence for Science), Yuan Cheng (Shanghai Academy of Artificial Intelligence for Science)
CodeSegmentationConvolutional Neural NetworkBiomedical Data
π― What it does: Proposes Disco, a cell instance segmentation method based on graph coloring, to address the segmentation of overlapping cells in high-density tissues.
DiscoX: Benchmarking Discourse-Level Translation in Expert Domains
Xiying ZHAO (ByteDance), Wenhao Huang (ByteDance)
CodeLarge Language ModelAgentic AITextBenchmark
π― What it does: Proposed the DiscoX benchmark for evaluating expert-level ChineseβEnglish long dialogue-level translation, and designed a reference-free Metric-S evaluation system.
π― What it does: Propose a method based on discrete diffusion models for unordered set generation, applied to bundle construction in product bundling;
Discrete Latent Features Ablate Adversarial Attack: A Robust Prompt Tuning Framework for VLMs
Yang Chen (Southern University of Science and Technology), Yu Zhang (Hong Kong University of Science and Technology)
CodeClassificationAdversarial AttackPrompt EngineeringVision Language ModelAuto EncoderImageTextMultimodalityBenchmark
π― What it does: To address the adversarial robustness issue in vision-language models, this paper proposes DEFEAT, an adversarial training framework based on discrete latent features, which defends against adversarial samples by reconstructing image features through VQ-VAE discretization and combining Prompt training.
π― What it does: Proposes a framework based on discrete hierarchical variational autoencoders and latent diffusion models to generate realistic and physically plausible 3D human-human interactive motions from text prompts;
Disentangled Representation Learning for Parametric Partial Differential Equations
Ning Liu (Lehigh University), Yue Yu (Lehigh University)
CodeRepresentation LearningAuto EncoderBiomedical DataPhysics Related
π― What it does: Proposes DisentangO, a hyper neural operator architecture that can disentangle physical factors from neural operator parameters and simultaneously solve forward and inverse PDE problems.
DispViT: Direct Stereo Disparity Regression with a Single-Stream Vision Transformer
Tongfan Guan (Chinese University of Hong Kong), Yun-Hui Liu (Chinese University of Hong Kong)
CodeDepth EstimationTransformerImage
π― What it does: Propose DispViT, a direct disparity regression framework based on a single-stream Vision Transformer, which moves away from traditional explicit matching and cost volume construction. It employs lightweight designs such as shift-embedding tokenizer, hetero-initialization, and Disparity-Aware RoPE, and adds a refinement module after regression to achieve fine results.
π― What it does: Propose the DistDF framework, which aligns the conditional distributions of the predicted sequence and the label sequence by minimizing the joint distribution Wasserstein distance, replacing the traditional MSE training objective.
Distillation of Large Language Models via Concrete Score Matching
Yeongmin Kim (Korea Advanced Institute of Science and Technology), Il-chul Moon
CodeKnowledge DistillationTransformerLarge Language ModelScore-based ModelTextBenchmark
π― What it does: Propose a new large language model knowledge distillation method called Concrete Score Distillation (CSD), which directly aligns the difference between student and teacher logits through discrete score matching, solving the traditional issues of softmax smoothing and logit identity bias;
π― What it does: The paper proposes a layer selection method based on KL-divergence, distilling a pre-trained softmax Transformer into an efficient hybrid attention (softmax + linear) architecture.
Distribution-Aware Multi-Granularity Phase Coding: Towards Lower Conversion Error for Spike-Driven Large Language Models
Hanyuan Zheng (Jilin University), Bin Gu (Jilin University)
CodeComputational EfficiencyKnowledge DistillationSpiking Neural NetworkTransformerLarge Language ModelText
π― What it does: Developed a distribution-aware multi-granularity phase encoding and an ANN-to-SNN conversion paradigm based on this encoding for low-error, low-energy spike-driven large language models (LLMs).
π― What it does: Proposes a distribution-level machine learning forgetting framework that uses selective data deletion to eliminate the impact of a specific subgroup while preserving the statistical properties of other subgroups.
Distributionally Robust Optimization via Generative Ambiguity Modeling
JIAQI WEN, Jianyi Yang (University of Houston)
CodeOptimizationReinforcement LearningDiffusion modelAuto EncoderImageTime Series
π― What it does: This paper proposes a generative adversarial distribution robust optimization (GAS-DRO) framework based on generative models, utilizing diffusion models or VAEs to generate adversarial distributions and constructing a solvable ambiguity set through constrained reconstruction loss, providing complete algorithms and convergence proofs.
DIVA-GRPO: Enhancing Multimodal Reasoning through Difficulty-Adaptive Variant Advantage
Haowen Gao (State Key Laboratory of AI Safety Institute of Computing Technology Chinese Academy of Sciences), Xueqi Cheng (State Key Laboratory of AI Safety Institute of Computing Technology Chinese Academy of Sciences)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningMultimodalityBenchmark
π― What it does: Developed a GRPO framework named DIVA-GRPO based on dynamic difficulty adaptive variant advantage to enhance the long-chain reasoning capability of multimodal large language models.
π― What it does: Proposed two differentiable vector quantization methods, DiVeQ and SF-DiVeQ, and evaluated them on multiple tasks including image compression, image generation, and speech decoding.
π― What it does: Propose a zero-divergence neural network architecture (CENSURE) achieved through structured conservative field combinations, which can be directly applied to unsupervised image denoising;
Ruiqi Shi (Chinese University of Hong Kong), Sinno Jialin Pan (Chinese University of Hong Kong)
CodeGenerationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes OverRIDEβa decoding method based on iterative reweightingβwhich significantly enhances LLM output diversity by real-time capturing and suppressing already generated semantic patterns through fine-tuning a low-rank output head adapter during inference.
Diversity-Enhanced Reasoning for Subjective Questions
Yumeng Wang (Hong Kong University of Science and Technology), Yi R. Fung (Hong Kong University of Science and Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: Designed and trained the MultiRole-R1 framework, enhancing the performance of large reasoning models on subjective reasoning tasks through self-supervised synthesis of multi-role reasoning chains and diversity reward shaping.
Divide and Abstract: Autoformalization via Decomposition and Abstraction Learning
Marcus J. Min (University of Pennsylvania), Osbert Bastani (University of Pennsylvania)
CodeAI Code AssistantLarge Language ModelTextBenchmark
π― What it does: Proposes a zero-training automatic formalization framework DNA, which first learns generic abstractions from the corpus and then hierarchically decomposes sentences to generate formalizations.
Divide, Harmonize, Then Conquer It: Shooting Multi-Commodity Flow Problems with Multimodal Language Models
Xinyu Yuan (Zhejiang University), Wenzhi CHEN
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelTextMultimodalityGraph
π― What it does: The paper proposes a multi-commodity traffic allocation method called PRAM based on a multimodal large language model, achieving fast and high-quality allocation through partitioning and multi-agent reinforcement learning.
Tom Kempton (University of Manchester), Stuart Burrell (Visa Inc.)
CodeAnomaly DetectionExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Propose the DMAP (Distribution Map for Text) method, which maps text through a language model to sampling points within the 0~1 interval, jointly encoding word ranking and probability information for text statistics and analysis.
Do LLM Agents Know How to Ground, Recover, and Assess? Evaluating Epistemic Competence in Information-Seeking Agents
Jiaqi Shao (Duke Kunshan University), Bing Luo (Duke Kunshan University)
CodeRetrievalLarge Language ModelReinforcement LearningTextBenchmark
π― What it does: Propose SeekBench, a process-level evaluation framework for search agents in large language models, which uses expert-annotated trajectory data to measure the agents' knowledge retrieval and reasoning processes.