International Conference on Machine Learning Β· 722 papers
CEGA: A Cost-Effective Approach for Graph-Based Model Extraction and Acquisition
Zebin Wang (Harvard University), Yushun Dong (Florida State University)
CodeGraph Neural NetworkGraph
π― What it does: A framework for efficient model extraction and acquisition of graph neural networks under limited query budgets (CEGA) is proposed, which iteratively selects the most informative nodes for querying and trains an intermediate model to gradually approximate the target model.
Chuanhui Liu (Purdue University), Xiao Wang (Purdue University)
CodeTabularBiomedical Data
π― What it does: This paper proposes a latent variable survival model method based on censor-dependent variational inference (CDVI), aimed at addressing the bias issue of traditional VI under right-censored data.
CFP-Gen: Combinatorial Functional Protein Generation via Diffusion Language Models
Junbo Yin (King Abdullah University of Science and Technology), Xin Gao (King Abdullah University of Science and Technology)
CodeGenerationDrug DiscoveryTransformerDiffusion modelMultimodalityBiomedical Data
π― What it does: CFP-GEN is proposed, a multimodal functional protein generation model based on discrete diffusion, capable of simultaneously satisfying multiple constraints such as functional annotation, sequence, and structure to generate multifunctional proteins.
CFPT: Empowering Time Series Forecasting through Cross-Frequency Interaction and Periodic-Aware Timestamp Modeling
Feifei Kou (Beijing University of Posts and Telecommunications), Junping Du (Beijing University of Posts and Telecommunications)
CodeOptimizationConvolutional Neural NetworkTransformerTime Series
π― What it does: The CFPT framework is proposed, which achieves cross-frequency interaction and periodic-aware timestamp modeling through a dual-branch structure to enhance long-term sequence prediction performance.
Channel Normalization for Time Series Channel Identification
Seunghan Lee (KRAFTON), Kibok Lee (Yonsei University)
CodeTime Series
π― What it does: Proposes Channel Normalization (CN) and its adaptive (ACN) and prototype (PCN) variants to enhance channel identifiability (CID) in multi-channel time series models.
Clients Collaborate: Flexible Differentially Private Federated Learning with Guaranteed Improvement of Utility-Privacy Trade-off
Yuecheng Li (Sun Yat-sen University), Chuan Chen (Sun Yat-sen University)
CodeFederated LearningSafty and PrivacyImage
π― What it does: A new federated learning framework called FedCEO is proposed, aimed at balancing model utility and user privacy by enabling client collaboration.
π― What it does: By generating symbolic expression skeletons through reinforcement learning and incorporating physical constraints, a closed-form analytical solution for differential equations is obtained.
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A cross-layer orthogonal vector pruning method named CLOVER is proposed, which utilizes SVD to perform low-rank decomposition on the Query-Key and Value-Output matrices of each attention head to obtain orthogonal bases and singular values.
Clustering Items through Bandit Feedback: Finding the Right Feature out of Many
Maximilian Graf (University of Potsdam), Nicolas Verzelen (INRAE)
CodeOptimizationTabular
π― What it does: An adaptive two-group clustering method called BanditClustering is proposed, which uses bandit feedback to gradually select features that can distinguish the two classes and complete the clustering accordingly.
π― What it does: This paper proposes a positive feedback framework called ReSA based on self-supervised learning representations of clustering information, which guides contrastive learning by extracting high-quality clustering assignments at the encoding layer.
CMoS: Rethinking Time Series Prediction Through the Lens of Chunk-wise Spatial Correlations
Haotian Si (Computer Network Information Center, Chinese Academy of Sciences), Gaogang Xie (Computer Network Information Center, Chinese Academy of Sciences)
CodeMixture of ExpertsTime Series
π― What it does: This paper proposes a super lightweight multivariate time series forecasting model CMoS, which directly models the spatial correlations between different time segment blocks, replacing traditional shape embedding methods.
CoCoA-Mix: Confusion-and-Confidence-Aware Mixture Model for Context Optimization
Dasol Hong (Korea Advanced Institute of Science and Technology), Hyun Myung (Korea Advanced Institute of Science and Technology)
CodeOptimizationTransformerPrompt EngineeringVision Language ModelImage
π― What it does: The CoCoA-Mix framework is proposed, which enhances the specialization and generalization of VLM prompt tuning by combining mixed models with confusion-aware loss and confidence-aware weights.
π― What it does: Transform the code into executable functions and set the prediction task as a given function and query, predicting the natural language reasoning chain (CoT) for inputs or outputs.
CoDy: Counterfactual Explainers for Dynamic Graphs
Zhan Qu (TU Dresden), Michael FΓ€rber
CodeExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkReinforcement LearningGraph
π― What it does: Designed and implemented CoDyβa counterfactual explanation method for continuous-time dynamic graph neural networks (TGNN) that can automatically search for and provide minimized event subgraphs to explain the reasons behind model predictions.
CogReact: A Reinforced Framework to Model Human Cognitive Reaction Modulated by Dynamic Intervention
Songlin Xu (University of California San Diego), Xinyu Zhang (University of California San Diego)
CodeExplainability and InterpretabilityComputational EfficiencyRecurrent Neural NetworkReinforcement LearningDiffusion modelTabularTime SeriesSequential
π― What it does: The CogReact framework is proposed, which combines the drift-diffusion model with deep reinforcement learning to simulate the fine effects of dynamic environmental stimuli (such as time pressure) on human cognitive responses.
π― What it does: This paper proposes a progressive training strategy named CoTo, aimed at improving the issues of hierarchical gradient imbalance and susceptibility to local optima during the fine-tuning process of LoRA (Low-Rank Adaptation) models.
Complex Wavelet Mutual Information Loss: A Multi-Scale Loss Function for Semantic Segmentation
Renhao Lu (Cornell University)
CodeSegmentationConvolutional Neural NetworkImage
π― What it does: A multi-scale loss function based on Complex Waveform Mutual Information (CWMI) is proposed to enhance the pixel accuracy and structural consistency of semantic segmentation models.
π― What it does: A prompt-based domain incremental learning method called KA-Prompt is proposed to address the knowledge conflict issue caused by inconsistent alignment of prompt components across different domains.
Compressing tree ensembles through Level-wise Optimization and Pruning
Laurens Devos (KU Leuven), Jesse Davis (KU Leuven)
CodeCompressionOptimizationTabular
π― What it does: A hierarchical optimization and pruning algorithm named LOP has been developed, which can significantly compress existing decision tree ensembles (reducing the number of leaves) while maintaining prediction accuracy.
Evi Micha (University of Southern California), Vasilis Varsamis (University of Southern California)
Code
π― What it does: This paper studies whether it is possible to accurately compute the winners of various voting rules under improved feedback (local improvements made by users on initial candidates). It provides the learnability boundaries for positional scoring rules and Condorcet consistent rules, and validates their performance through experiments.
π― What it does: A concentration distribution learning (CDL) paradigm is proposed, which incorporates background concentration into the label distribution and designs a CDLLD model that can directly learn the concentration distribution from the LDL dataset.
Conditional Diffusion Model with Nonlinear Data Transformation for Time Series Forecasting
J Rishi (Indian Institute of Science), Deepak Subramani (Indian Institute of Science)
CodeGenerationOptimizationDiffusion modelTime Series
π― What it does: A time series forecasting framework named CN-Diff is proposed, which incorporates learnable nonlinear time transformations and conditions during the forward diffusion process, along with a corresponding variational training objective.
π― What it does: This study proposes a ConfDiff classification method based on consistency risk and consistency regularization (CRCR), which reduces the impact of noisy supervision by grouping confidence differences, thereby enhancing the performance of weakly supervised binary classifiers.
ConfPO: Exploiting Policy Model Confidence for Critical Token Selection in Preference Optimization
Hee Suk Yoon (Korea Advanced Institute of Science and Technology), Chang D. Yoo (Korea Advanced Institute of Science and Technology)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
π― What it does: This work proposes ConfPO, a preference optimization method that utilizes the model's own confidence for key token selection, updating only low-confidence (high information content) tokens, thereby significantly improving human preference consistency in alignment training.
π― What it does: A gradient-free deep reinforcement learning algorithm based on the Consensus-Based Optimization (CBO) framework is proposed to solve high-dimensional, finite-time stochastic optimal control problems.
π― What it does: A novel offline reinforcement learning algorithm named Constrained Exploitability Descent (CED) is proposed to find mixed strategy Nash equilibria under limited datasets.
Context is Key: A Benchmark for Forecasting with Essential Textual Information
Andrew Robert Williams, Alexandre Drouin (ServiceNow Research)
CodeLarge Language ModelPrompt EngineeringTextTime SeriesBenchmark
π― What it does: The 'Context is Key (CiK)' benchmark is proposed, specifically to evaluate the ability to utilize necessary natural language context in time series forecasting.
π― What it does: Designed and implemented an online reinforcement learning agent (OA) that learns a sparse shallow world model for online Follow-The-Leader (FTL) and uses planning and model predictive control (MPC) to achieve continuous multi-task learning under unified environmental dynamics.
π― What it does: A continuous semi-implicit model (CoSIM) is proposed, which transforms the hierarchical semi-implicit model into a continuous time framework to achieve rapid distillation and generation of multi-step diffusion models.
Contrastive Private Data Synthesis via Weighted Multi-PLM Fusion
Tianyuan Zou (Tsinghua University), Ya-Qin Zhang
CodeData SynthesisSafty and PrivacyLarge Language ModelPrompt EngineeringContrastive LearningGaussian SplattingText
π― What it does: Utilizing multi-model (PLM) collaboration to generate differential privacy synthetic data, addressing issues of private sample scarcity, synthetic noise, and model selection uncertainty.
Controlling Large Language Model with Latent Action
Chengxing Jia (Nanjing University), Yang Yu (Nanjing University)
CodeTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: This study investigates learning a compact latent action space in large language models and proposes the CoLA framework to enhance RL control and exploration capabilities.
Siyuan Duan (Sichuan University), Yuan Sun (Sichuan University)
CodeOptimizationComputational EfficiencyTabularPhysics Related
π― What it does: This paper proposes Cognitive Physics-Informed Neural Networks (CoPINN), which addresses the Unbalanced Prediction Problem in traditional PINNs through an adaptive training strategy that progresses from easy to difficult.
CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models
Qinsi Wang (Duke University), Yiran Chen (Duke University)
CodeOptimizationComputational EfficiencyTransformerVision Language ModelMultimodality
π― What it does: This paper proposes a framework called CoreMatching for joint adaptive sparse inference, which accelerates the inference of large-scale vision-language models by simultaneously sparsifying visual tokens and neurons.
π― What it does: A susceptibility risk framework based on causal counterfactuals, COSDA, is proposed to address the issues of domain shift and unknown category recognition in open set domain adaptation (OSDA).
CoSER: Coordinating LLM-Based Persona Simulation of Established Roles
Xintao Wang (Fudan University), Yanghua Xiao (Fudan University)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextRetrieval-Augmented Generation
π― What it does: This paper presents the COSER dataset, the corresponding LLaMA-3.1 base models (COSER-8B and COSER-70B), and a training and evaluation framework based on Given Contextual Abduction (GCA) to achieve realistic role-playing language agents for existing literary characters.
Counterfactual Effect Decomposition in Multi-Agent Sequential Decision Making
Stelios Triantafyllou (Max Planck Institute for Software Systems), Goran Radanovic (Max Planck Institute for Software Systems)
CodeExplainability and InterpretabilityReinforcement LearningAgentic AISequential
π― What it does: A method for counterfactual effect decomposition in multi-agent Markov decision processes is proposed, which can break down the impact of a single agent's action on the final outcome into contributions to subsequent agent behaviors and contributions to environmental state transitions.
Counting in Small Transformers: The Delicate Interplay between Attention and Feed-Forward Layers
Freya Behrens (Ecole Polytechnique Federale de Lausanne), Lenka Zdeborova (Ecole Polytechnique Federale de Lausanne)
CodeTransformerSequential
π― What it does: This study investigates the implementation mechanism of a single-layer Transformer model in counting tasks, comparing the roles of attention and feedforward layers under different architectures and hyperparameters.
π― What it does: This paper proposes the CoRE method, which achieves self-supervised learning of regional embeddings and latent space alignment across cities, enabling direct transfer of predictors in the absence of target city labels.
CSTrack: Enhancing RGB-X Tracking via Compact Spatiotemporal Features
Xiaokun Feng (University of Chinese Academy of Sciences), Kaiqi Huang (University of Chinese Academy of Sciences)
CodeObject TrackingTransformerMultimodality
π― What it does: A RGB-X visual tracker named CSTrack is proposed, which simplifies the network structure and enhances tracking robustness by merging RGB and auxiliary modalities (depth, thermal, event) into compact spatiotemporal features.
π― What it does: CTBENCH is proposed - a unified library and high-quality benchmark for fair and systematic evaluation and comparison of deterministic provably robust training methods under the Lβ norm.
Curse of High Dimensionality Issue in Transformer for Long Context Modeling
Shuhai Zhang (South China University of Technology), Mingkui Tan (South China University of Technology)
CodeComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Proposes Dynamic Group Attention (DGA), which reduces redundant attention computations in long sequences by aggregating unimportant tokens;
π― What it does: This paper proposes a curvature-aware graph attention mechanism for efficiently solving time-dependent partial differential equations (such as heat diffusion, p-Laplace diffusion, and wave equations) on discrete manifolds.
π― What it does: A curvature-based mixed curvature graph autoencoder (CurvGAD) is proposed, which detects graph structure, attributes, and geometric anomalies through two parallel pipelines: curvature equivariant and curvature invariant.
Customizing the Inductive Biases of Softmax Attention using Structured Matrices
Yilun Kuang (New York University), Andrew Gordon Wilson (New York University)
CodeTransformerLarge Language ModelTextTime Series
π― What it does: This paper addresses the low-rank bottleneck and the lack of distance-related computational bias by rewriting the Softmax attention through the integration of Block Tensor-Train with Multi-Level Low-Rank matrix embedded attention score functions.
π― What it does: A cutting and experience replay strategy named CUTER is proposed to address catastrophic forgetting, missing labels, and class imbalance issues in multi-label online continual learning.
CVE-Bench: A Benchmark for AI Agentsβ Ability to Exploit Real-World Web Application Vulnerabilities
Yuxuan Zhu (University of Illinois), Daniel Kang (University of Illinois)
CodeTransformerLarge Language ModelAgentic AIBenchmark
π― What it does: CVE-Bench was constructed, a security benchmark based on 40 high-risk CVEs from real-world web applications, and implemented a sandbox environment, reference exploits, and an automated evaluation system; experiments were conducted on three LLM agents (Cy-Agent, T-Agent, AutoGPT) in zero-day and one-day scenarios, reporting their success rates and costs.
D-Fusion: Direct Preference Optimization for Aligning Diffusion Models with Visually Consistent Samples
Zijing Hu (Zhejiang University), Kun Kuang (Zhejiang University)
CodeGenerationOptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningDiffusion modelImage
π― What it does: Proposes the D-Fusion method, which utilizes self-attention fusion to generate visually consistent samples that are similar to the original image but well-aligned with the text, and pairs them with the original samples for direct preference optimization (DPO) fine-tuning of the diffusion model, thereby improving the alignment quality between text and images.
DAMA: Data- and Model-aware Alignment of Multi-modal LLMs
Jinda Lu (University of Science and Technology of China), Xiangnan He (University of Science and Technology of China)
CodeOptimizationTransformerLarge Language ModelReinforcement LearningMultimodalityBenchmark
π― What it does: The DAMA method is proposed for direct preference optimization of multimodal large language models, dynamically adjusting Ξ² to balance the learning of easily distinguishable and difficult-to-distinguish samples.
π― What it does: A deep attention-based least squares reconstruction framework (DEAL) is proposed, which iteratively solves quadratic optimization and uses learnable filters and attention weights for adaptive regularization.
Decision-aware Training of Spatiotemporal Forecasting Models to Select a Top-K Subset of Sites for Intervention
Kyle Heuton (Tufts University), Michael C Hughes
CodeRecommendation SystemOptimizationReinforcement LearningTime Series
π― What it does: A spatiotemporal prediction model training and ranking method based on the Best Possible Rate (BPR) metric is proposed for selecting the optimal K sites for intervention under resource constraints.
Decomposition of Graphic Design with Unified Multimodal Model
Hui Nie (University of Chinese Academy of Sciences), Xinglong Wu
CodeRestorationGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningGenerative Adversarial NetworkImageMultimodality
π― What it does: This paper proposes the Layered Deconstruction of Graphic Design (LDGD) task and designs a unified multimodal model DeaM, which can decompose composite graphic designs into ordered RGB-A layers and corresponding metadata.
Deep Fuzzy Multi-view Learning for Reliable Classification
Siyuan Duan (Sichuan University), Peng Hu (Sichuan University)
CodeClassificationMultimodality
π― What it does: A deep multi-view learning framework based on fuzzy set theory, called FUML, is designed to achieve reliable classification and accurate uncertainty estimation in the presence of conflicting views.
Koen Minartz (Eindhoven University of Technology), Vlado Menkovski (Eindhoven University of Technology)
CodeGenerationData SynthesisConvolutional Neural NetworkReinforcement LearningImageBiomedical Data
π― What it does: Proposes NeuralCPM, which utilizes neural networks to parameterize the energy function of the cellular Potts model, allowing for direct learning of cellular dynamics from observational data.
DEFAME: Dynamic Evidence-based FAct-checking with Multimodal Experts
Tobias Braun (Technical University of Darmstadt), Anna Rohrbach (Technical University of Darmstadt)
CodeRetrievalExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringImageTextMultimodalityBenchmarkRetrieval-Augmented Generation
π― What it does: Designed DEFAME, a zero-shot dynamic evidence retrieval and reasoning framework based on multimodal LLM, achieving end-to-end multimodal fact-checking.
Delta Decompression for MoE-based LLMs Compression
Hao Gu (Hong Kong University of Science and Technology), Yike Guo (Hong Kong University of Science and Technology)
CodeCompressionTransformerLarge Language ModelMixture of ExpertsText
π― What it does: Designed and implemented the D2-MOE framework, achieving efficient compression without training by splitting the MoE expert weights into shared base weights and expert-specific Delta weights.
π― What it does: This study investigates the catastrophic forgetting mechanism of the two-stage incremental object detector Faster R-CNN and proposes the NSGP-RePRE framework to mitigate forgetting in the RoI Head classifier.
π― What it does: The DRE3 framework is proposed, utilizing the dequantized diffusion bridge (DDBI) and the dequantized SchrΓΆdinger bridge (DSBI) for robust density ratio estimation, addressing the density-chasm and support-chasm issues.
Differentiable Quadratic Optimization For the Maximum Independent Set Problem
Ismail Alkhouri, Alvaro Velasquez (University of Colorado)
CodeOptimizationGraph
π― What it does: A new continuous quadratic optimization framework pCQO-MIS based on maximum clique information is proposed to solve the maximum independent set problem.
π― What it does: This paper studies how to combine diffusion models with structural causal models (SCM) for high-fidelity, identity-preserving counterfactual image generation, and proposes three mechanisms: spatial, semantic, and dynamic semantic induction.
π― What it does: A PCA-based sampling correction method called PAS is proposed, which can improve the sampling quality of DPMs without significantly increasing training costs.
Directly Forecasting Belief for Reinforcement Learning with Delays
Qingyuan Wu (University of Southampton), Chao Huang (University of Southampton)
CodeTransformerReinforcement LearningSequential
π― What it does: To address the problem of reinforcement learning with observation delays, this paper proposes the Direct Forecasting Belief Transformer (DFBT) and its derivative method DFBT-SAC, which utilizes one-shot sequence prediction to directly reconstruct unobservable states, thereby avoiding the cumulative errors associated with recursive predictions.
Discovering Physics Laws of Dynamical Systems via Invariant Function Learning
Shurui Gui (Texas A&M University), Shuiwang Ji (Texas A&M University)
CodeOptimizationExplainability and InterpretabilityMeta LearningTime SeriesPhysics RelatedOrdinary Differential Equation
π― What it does: A distinguishable function learning framework based on causal analysis (DIF) is proposed to extract the essential invariant functions of dynamical systems from observational trajectories in different environments, thereby achieving the discovery of physical laws.
π― What it does: Construct an interpretable discrete neural algorithm interpreter that forces the network to execute algorithm steps on a finite set of predefined state combinations.
Discriminative Finetuning of Generative Large Language Models without Reward Models and Human Preference Data
Siqi Guo (Texas A&M University), Tianbao Yang (Texas A&M University)
CodeGenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A discriminative fine-tuning framework (DFT/DFT2) is proposed that does not rely on reward models or human preference data, shifting LLM from traditional generative fine-tuning to likelihood-based learning;
Discriminative Policy Optimization for Token-Level Reward Models
Hongzhan Chen (Sun Yat-sen University), Ting Yao (Tencent Inc)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerReinforcement LearningText
π― What it does: A Q-RM (Q-Function Reward Model) based on a discriminative strategy is proposed, which learns token-level rewards from preference data, avoiding conflicts between the generative model and reward modeling.
Liang Yang (Hebei University of Technology), Yuanfang Guo (Beihang University)
CodeDomain AdaptationGraph Neural NetworkGraph
π― What it does: The Disentangled Graph Spectral Domain Adaptation (DGSDA) framework is proposed, decoupling attribute alignment from topological alignment, and directly aligning spectral filters to achieve unsupervised graph domain adaptation.
Distillation of Discrete Diffusion through Dimensional Correlations
Satoshi Hayakawa (Sony Group Corporation), Yuki Mitsufuji (Sony AI)
CodeGenerationKnowledge DistillationMixture of ExpertsDiffusion modelImageText
π― What it does: This paper proposes a discrete diffusion model distillation framework based on dimension correlation, Di4C, which can compress a multi-step diffusion process into fewer sampling steps while capturing the correlation between dimensions through a mixture model.
π― What it does: This paper proposes a method to view bivariate causal structure models as dynamical systems, utilizing causal speed and score functions to achieve causal direction determination without simulation.
Distribution-aware Fairness Learning in Medical Image Segmentation From A Control-Theoretic Perspective
Yujin Oh (Massachusetts General Hospital), Quanzheng Li (Massachusetts General Hospital)
CodeSegmentationConvolutional Neural NetworkTransformerMixture of ExpertsImageBiomedical DataComputed TomographyOrdinary Differential Equation
π― What it does: A distribution-aware mixture of experts (dMoE) framework is proposed to enhance the fairness and robustness of medical image segmentation.
π― What it does: This paper combines inverse reinforcement learning with quality diversity (QD) methods, proposing an external behavior curiosity-based reward mechanism (EBC) that enhances the behavior diversity and performance of the IL model under limited demonstration data by rewarding unoccupied behavior space units.
Divide and Conquer: Grounding LLMs as Efficient Decision-Making Agents via Offline Hierarchical Reinforcement Learning
Zican Hu (Nanjing University), Yu Cheng (Chinese University of Hong Kong)
CodeTransformerLarge Language ModelReinforcement LearningAgentic AIText
π― What it does: This paper proposes a GLIDER framework for LLM agents based on offline hierarchical reinforcement learning, which can decompose large language models into high-level planning and low-level execution strategies, enabling efficient learning and deployment of long-term decision-making tasks.
π― What it does: A S-LDL framework is proposed, which automatically constructs sub-tasks (sub-label spaces) without prior knowledge to generate additional supervisory information, thereby enhancing the performance of label distribution learning (LDL).
Daniel Shao (Massachusetts Institute of Technology), Faisal Mahmood (Harvard University)
CodeClassificationDomain AdaptationRepresentation LearningTransformerBiomedical Data
π― What it does: Under the multi-instance learning (MIL) framework, transfer learning experiments were conducted on 19 computational pathology tasks, evaluating the transfer effects of 11 MIL architectures and 21 pre-training tasks. A 'foundation model' at the slide level based on pancancer supervised pre-training was proposed and validated, along with the provision of unified implementation and weight resources.
π― What it does: The DPCore method is proposed to achieve Continuous Testing Time-domain Adaptation (CTTA), efficiently adapting to the continuously changing target domain through visual prompts, a prompt core set, and dynamic updates.
DPO Meets PPO: Reinforced Token Optimization for RLHF
Han Zhong (Peking University), Liwei Wang (Peking University)
CodeOptimizationReinforcement Learning from Human FeedbackReinforcement LearningText
π― What it does: Re-modeling Reinforcement Learning with Human Feedback (RLHF) from the traditional sentence-level Bandit framework to a fine-grained token-level Markov Decision Process (MDP), and proposing the Reinforced Token Optimization (RTO) algorithm: first, using DPO to learn token-level rewards from offline preference data, and then applying PPO for reinforcement learning on that reward;
DragLoRA: Online Optimization of LoRA Adapters for Drag-based Image Editing in Diffusion Model
Siwei Xia (Shanghai Key Laboratory of Multidimensional Information Processing), Qingli Li (Key Laboratory of Advanced Theory and Application in Statistics and Data Science)
π― What it does: Implement drag-and-drop image editing on pre-trained diffusion models using the online optimization LoRA adapter framework called DragLoRA.
DreamDPO: Aligning Text-to-3D Generation with Human Preferences via Direct Preference Optimization
Zhenglin Zhou (Zhejiang University), Tat-Seng Chua (National University of Singapore)
CodeGenerationData SynthesisOptimizationLarge Language ModelReinforcement LearningNeural Radiance FieldTextPoint Cloud
π― What it does: Achieving text-to-3D content generation and alignment with human preferences through Direct Preference Optimization (DreamDPO), constructing online comparative samples and evaluating preferences using a reward model or large multimodal models, thereby updating 3D representations;
π― What it does: An end-to-end global routing model called DSBRouter is proposed, which can directly generate connected and low-overflow routing results.
Dynamic Similarity Graph Construction with Kernel Density Estimation
Steinar Laenen (University of Edinburgh), He Sun (University of Edinburgh)
CodeComputational EfficiencyImageTabular
π― What it does: An algorithm for dynamically maintaining kernel density estimation (KDE) and similarity graphs is proposed, and based on this, dynamic spectral clustering is implemented.
π― What it does: This study explores two main sequence generation mechanisms in short-term memory tasksβslow-point manifold and limit cycleβthrough large-scale training of RNNs, revealing their relationships with task design, learning rate, and delay duration.
e-GAI: e-value-based Generalized $\alpha$-Investing for Online False Discovery Rate Control
Yifan Zhang (Shanghai Jiao Tong University), Changliang Zou (Nankai University)
CodeAnomaly DetectionTime SeriesFinance Related
π― What it does: This paper proposes an online hypothesis testing framework based on e-values, called e-GAI, and designs two algorithms, e-LORD and e-SAFFRON, which can achieve online FDR control through e-values under any dependency structure. It also provides long-term improvement plans for the Ξ±-death problem with mem-e-LORD and mem-e-SAFFRON; additionally, the framework is extended to p-values that satisfy conditional super-uniformity. Its effectiveness is validated through theoretical proofs and simulations/real data experiments.
E-LDA: Toward Interpretable LDA Topic Models with Strong Guarantees in Logarithmic Parallel Time
Adam Breuer (Dartmouth College)
CodeOptimizationExplainability and InterpretabilityComputational EfficiencyText
π― What it does: This paper proposes a combination optimization-based LDA topic allocation algorithm E-LDA, which can approximately optimally solve the topic-word allocation problem under the constraint of document sparsity.
Earley-Driven Dynamic Pruning for Efficient Structured Decoding
Xintong Sun (Rice University), Shiwen Ni (Shenzhen Institutes of Advanced Technology)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelText
π― What it does: A constraint decoding engine named Formatron has been constructed, utilizing CFG and the Earley parser to achieve structured generation in large language models, while dynamically pruning invalid states during the inference process.
Edge-Colored Clustering in Hypergraphs: Beyond Minimizing Unsatisfied Edges
Alex Crane (University of Utah), Nate Veldt (Texas A&M University)
CodeOptimizationGraph
π― What it does: A clustering framework for edge-colored hypergraphs is proposed, and research is conducted from various objectives such as minimizing unsatisfied edges and maximizing satisfied edges.
EEG-Language Pretraining for Highly Label-Efficient Clinical Phenotyping
Sam Gijsen (Charite University Medicine Berlin), Kerstin Ritter (Hertie Institute for AI in Brain Health)
CodeClassificationRetrievalConvolutional Neural NetworkContrastive LearningMultimodalityTime SeriesBiomedical DataElectronic Health Records
π― What it does: This study proposes a multimodal pre-training model (ELM) based on clinical reports and EEG signals, achieving sub-unit level cross-modal alignment through the trimming of EEG time series, segmentation of reports, and the integration of multi-instance learning (MIL).
Efficient and Separate Authentication Image Steganography Network
Junchao Zhou (Harbin Institute of Technology), Guangming Lu (Harbin Institute of Technology)
CodeImage TranslationData SynthesisSafty and PrivacyComputational EfficiencyFlow-based ModelImage
π― What it does: An efficient and independently authenticable image steganography network (AIS) is proposed, which achieves multi-image steganography and decryption through a two-stage reversible network and supports independent authentication.
Suyuan Liu (National University of Defence Technology), Xinwang Liu (National University of Defence Technology)
CodeFederated LearningSafty and PrivacyComputational EfficiencyMultimodality
π― What it does: A federated incomplete multi-view clustering framework EFIMVC is proposed, which can handle missing views, has high communication efficiency, and maintains privacy.
Efficient Fine-Grained Guidance for Diffusion Model Based Symbolic Music Generation
Tingyu Zhu (University of California), Zeyu Zheng (University of California)
CodeGenerationDiffusion modelAudio
π― What it does: This paper proposes the Fine-Grained Guidance (FGG) method, which integrates fine-grained chord and rhythm constraints into the training and sampling of diffusion models, achieving high-precision generation of symbolic music.
π― What it does: A time-decomposed logits knowledge distillation framework is proposed to train deep SNNs, enabling them to maintain high accuracy at any time step without the need for retraining.
Efficient Multi-modal Long Context Learning for Training-free Adaptation
Zehong Ma (Peking University), Qi Tian (Huawei Inc)
CodeCompressionComputational EfficiencyTransformerLarge Language ModelMultimodality
π― What it does: A training-independent multimodal long-context learning framework EMLoC is proposed, which compresses long contexts into compact memory using chunk compression and hierarchical adaptive pruning.
CodeExplainability and InterpretabilityComputational EfficiencyTabularBiomedical DataFinance Related
π― What it does: A sparse regression framework for multi-output linear probability models, called NARD (Network Automatic Relevance Determination), is proposed. Based on this framework, three efficient variants (Sequential NARD, Surrogate NARD, Hybrid NARD) are designed to significantly reduce computational complexity while maintaining sparsity and modeling output correlation.
π― What it does: A method for unbiased estimation based on the Jacobian matrix sketch is proposed for efficiently calculating voxel-level noise variance in deep learning MRI reconstruction.
π― What it does: A Reweighted Score Matching (RSM) framework is proposed for training diffusion policies in online reinforcement learning, along with the implementation of two efficient algorithmsβDiffusion Policy Mirror Descent (DPMD) and Soft Diffusion Actor-Critic (SDAC);