Conference on Neural Information Processing Systems Β· 1874 papers
Conjugated Semantic Pool Improves OOD Detection with Pre-trained Vision-Language Models
Mengyuan Chen (Institute of Automation, Chinese Academy of Sciences), Changsheng Xu (Institute of Automation, Chinese Academy of Sciences)
CodeAnomaly DetectionTransformerVision Language ModelImageText
π― What it does: A theoretical framework for zero-shot OOD detection is proposed, and based on this, a conjugated semantic pool (CSP) is designed to enhance performance.
Connecting the Dots: LLMs can Infer and Verbalize Latent Structure from Disparate Training Data
Johannes Treutlein (University of California Berkeley), Owain Evans
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This study explores whether large language models (LLMs) can infer hidden facts (referred to as inductive out-of-context reasoning, OOCR) by aggregating dispersed, implicit information from training data, and validates their 'connection point' ability through fine-tuning the model and evaluating it on five different tasks.
Connectivity-Driven Pseudo-Labeling Makes Stronger Cross-Domain Segmenters
Dong Zhao (Xidian University), Zhun Zhong (University of Nottingham)
CodeSegmentationDomain AdaptationImage
π― What it does: A pseudo-label generation framework based on semantic connectivity (SeCo) is proposed, which aggregates noisy pixels into connected regions for denoising and correction to enhance cross-domain semantic segmentation performance.
Consensus Learning with Deep Sets for Essential Matrix Estimation
Dror Moran (Weizmann Institute of Science), Ronen Basri (Weizmann Institute of Science)
CodeOptimizationComputational EfficiencyImage
π― What it does: A NACNet network based on Deep Sets has been designed and implemented for consistent learning from matching sets containing a large number of outliers and noise, first denoising inliers and then classifying, ultimately regressing the essential matrix through weighted DLT.
Consistency of Neural Causal Partial Identification
Jiyuan Tan (Stanford University), Vasilis Syrgkanis (Stanford University)
Code
π― What it does: A partial identification method based on Neural Causal Models (NCM) is proposed, and its consistency is proven under a general Structural Causal Model (SCM) that includes both continuous and categorical variables.
π― What it does: By introducing a learnable acceleration term, we propose Constant Acceleration Flow (CAF) to approximate ODE flows, improving the quality of single sampling.
ConStat: Performance-Based Contamination Detection in Large Language Models
Jasper Dekoninck (ETH Zurich), Martin Vechev (ETH Zurich)
CodeTransformerLarge Language ModelTextBenchmark
π― What it does: A performance-based detection method called CONSTAT has been designed to identify performance exaggeration caused by data contamination in benchmark tests of large language models.
π― What it does: A novel training-free guided diffusion method called Trust Sampling is proposed, which satisfies constraints through multi-step gradient optimization at each diffusion step.
Constrained Latent Action Policies for Model-Based Offline Reinforcement Learning
Marvin Alles (Volkswagen Group), Maximilian Karl (Volkswagen Group)
CodeReinforcement LearningTabularSequential
π― What it does: A model-based offline reinforcement learning method called C-LAP is proposed, which generates actions in the latent action space and maintains the policy within the dataset distribution through support constraints of the latent action distribution, addressing the value overestimation problem caused by model errors.
π― What it does: This paper studies the problem of sampling probability distributions under statistical constraints (such as expectation constraints) and proposes a Primal-Dual Langevin Monte Carlo (PD-LMC) algorithm without explicit integration;
π― What it does: This paper proposes an implicit field representation that can simultaneously estimate the occupancy rate, identity labels (ID), and contact fields of multiple human interactions, achieving precise reconstruction of occlusions and close-range interactions through a multi-view local-global feature module.
Context-Aware Testing: A New Paradigm for Model Testing with Large Language Models
Paulius Rauba (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeTransformerLarge Language ModelTabularBiomedical DataFinance Related
π― What it does: A Context-Aware Testing (CAT) framework is proposed, and the SMART Testing system is implemented, which uses large language models (LLMs) to generate potential model failure hypotheses, automate verification, and generate reports.
ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model
Yufei Wang (Nanyang Technological University), Bihan Wen (Nanyang Technological University)
CodeCompressionGaussian SplattingPoint Cloud
π― What it does: A compression framework based on 3D Gaussian Splatting (ContextGS) is proposed, utilizing a hierarchical anchor autoregressive model and hyper-prior features for efficient encoding.
Contextual Bilevel Reinforcement Learning for Incentive Alignment
Vinzenz Thoma (ETH AI Center), Yifan Hu (ETH Zurich)
CodeOptimizationReinforcement LearningTabular
π― What it does: This paper proposes the Contextual Bilevel Reinforcement Learning (CB-RL) framework, which studies the bilevel MDP problem in stochastic contexts and multi-follower environments, and presents a trajectory-based stochastic hypergradient descent algorithm (HPGD), further providing an accelerated version of RT-Q for controllable lower-level learning.
Contextual Linear Optimization with Bandit Feedback
Yichun Hu (Cornell University), Yanchen Wu (Tsinghua University)
CodeOptimizationReinforcement LearningTabular
π― What it does: This paper studies the contextual linear optimization (CLO) problem under bandit feedback, where only decision costs can be observed, and proposes an end-to-end Induced Empirical Risk Minimization (IERM) algorithm.
π― What it does: Proposes a continuous audio-visual sound separation task and designs the ContAV-Sep framework, allowing the model to maintain separation performance for old categories while learning new sound source categories.
π― What it does: This paper proposes a framework for incremental learning in the frequency domain called CLFD, which improves the efficiency and accuracy of continuous learning by utilizing frequency domain feature encoding and class-aware feature selection.
Xueying Bai (Stony Brook University), Niranjan Balasubramanian (Stony Brook University)
CodeTransformerSupervised Fine-TuningText
π― What it does: A global alignment method is proposed, utilizing pre-trained word representations as a foundation, and learning data representations through task-specific interpolation or low-rank adaptation (LoRA), thereby reducing cross-task gradient interference and catastrophic forgetting during continual learning.
Continual learning with the neural tangent ensemble
Ari S Benjamin, Kyle Daruwalla (Cold Spring Harbor Laboratory)
CodeConvolutional Neural NetworkMixture of ExpertsImage
π― What it does: Proposes to view a single neural network as a neural tangent ensemble (NTE) of parameter-level experts, and utilizes its posterior updates to achieve continual learning, avoiding catastrophic forgetting;
π― What it does: A Continuous Contrastive Learning (CCL) framework is proposed, which utilizes reliable and continuous pseudo-labels to improve representation learning in long-tail semi-supervised learning, addressing the issues of pseudo-label confirmation bias and inconsistency in the label distribution of unlabeled data.
Continuous Heatmap Regression for Pose Estimation via Implicit Neural Representation
Shengxiang Hu (Nanjing University of Science and Technology), Jin Wang (Nantong University)
CodePose EstimationImage
π― What it does: This paper studies a continuous heatmap regression model called NerPE based on implicit neural representations for 2D human pose estimation.
π― What it does: A CutSSL framework based on continuous non-convex quadratic programming is proposed to directly obtain integer label assignments in graph-structured semi-supervised learning, addressing the degradation problem of Laplacian learning under low label rates and class imbalance.
π― What it does: A continuous product graph neural network called CITRUS is proposed, based on tensor partial differential equations, for joint learning and spatiotemporal prediction of multi-domain graph data.
π― What it does: This work proposes the Continuous Time Domain Generalization (CTDG) framework Koodos, which addresses the limitation of traditional TDG that can only handle discrete time points, allowing for generalization of the model at any continuous time point.
π― What it does: This paper presents Symile, a contrastive learning framework for an arbitrary number of modalities that captures higher-order joint information while maintaining the simplicity of CLIP, generating modality-specific representations that can be directly used for zero-shot retrieval and downstream tasks.
Sam Hawke (University of North Carolina), Didong Li (University of North Carolina)
CodeAnomaly DetectionContrastive LearningImageBiomedical Data
π― What it does: This paper proposes a hypothesis testing method to determine whether there is unique information between the foreground group and the background group, and provides an estimator for quantifying the unique dimensions of the foreground;
Controlling Counterfactual Harm in Decision Support Systems Based on Prediction Sets
Eleni Straitouri (Max Planck Institute for Software Systems), Manuel Gomez Rodriguez (Max Planck Institute for Software Systems)
CodeClassificationImage
π― What it does: This study investigates causal counterfactual harm in decision support systems based on predictive sets and proposes a theoretical and algorithmic framework to control its occurrence frequency.
Controlling Multiple Errors Simultaneously with a PAC-Bayes Bound
Reuben Adams (University College London), Benjamin Guedj (University College London)
CodeOptimizationImageBiomedical Data
π― What it does: A PAC-Bayes generalization bound that can simultaneously control the distribution of multiple types of errors is proposed, and it is transformed into a differentiable training objective.
π― What it does: A novel continuous-time neural network called ControlSynth Neural ODEs (CSODEs) is proposed for modeling highly nonlinear and scalable dynamical systems, ensuring convergence through solvable linear inequalities.
Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents
Giorgio Piatti (ETH ZΓΌrich), Rada Mihalcea (University of Michigan)
CodeTransformerLarge Language ModelAgentic AIText
π― What it does: Created the GOVSIM platform to simulate cooperation and governance among multiple agents in public resource sharing using large language models, and to assess its sustainability.
π― What it does: A Federated Hardware Prompt Learning (FedHP) framework is proposed and implemented for collaborative training of multi-hardware configuration snapshot compression imaging (SCI) systems without sharing hardware configurations and measurement data, significantly improving reconstruction performance under different optical encoders.
CorDA: Context-Oriented Decomposition Adaptation of Large Language Models for Task-Aware Parameter-Efficient Fine-tuning
Yibo Yang (King Abdullah University of Science and Technology), Bernard Ghanem (King Abdullah University of Science and Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: A parameter-efficient fine-tuning method called CorDA based on context-oriented singular value decomposition is proposed, which allows for flexible switching between maintaining world knowledge and enhancing downstream task performance.
π― What it does: A compression scheme for satellite images called COSMIC is proposed, which uses a lightweight encoder to compress images at the satellite end and employs a conditional diffusion model at the ground end to compensate for the details missing from the encoder, achieving efficient and high-quality image compression and reconstruction.
Cost-aware Bayesian Optimization via the Pandora's Box Gittins Index
Qian Xie (Cornell University), Alexander Terenin (Cornell University)
CodeOptimizationTabular
π― What it does: This paper proposes a novel acquisition function PBGI based on the Pandora's Box Gittins index for making sampling decisions in Bayesian optimization considering evaluation costs, and validates its effectiveness on multi-dimensional, multi-modal problems.
CodeClassificationExplainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkAuto EncoderImage
π― What it does: This paper proposes a biologically interpretable Counter-Current Learning (CCL) framework, which consists of a dual network structure formed by a feedforward network and a corresponding feedback network. Learning is achieved by backpropagating information, local loss, and gradient separation between the two networks.
Counterfactual Fairness by Combining Factual and Counterfactual Predictions
Zeyu Zhou (Purdue University), David I. Inouye (Purdue University)
CodeTabular
π― What it does: Theoretical analysis of the trade-off between adversarial bias and predictive performance is conducted, proposing the optimal counterfactual fairness (CF) predictor, and providing error bounds and a practical plug-in algorithm PCF in the absence of complete causal knowledge.
Coupled Mamba: Enhanced Multimodal Fusion with Coupled State Space Model
Wenbing Li (Huazhong University of Science and Technology), Wei Yang (Huazhong University of Science and Technology)
CodeClassificationTransformerMultimodality
π― What it does: The Coupled Mamba model is proposed, utilizing the state chain in state space models to achieve the coupling and fusion of multimodal information, enhancing the performance of multimodal sentiment analysis and classification tasks.
COVE: Unleashing the Diffusion Feature Correspondence for Consistent Video Editing
Jiangshan Wang (Tsinghua University), Xiu Li (Tsinghua University)
CodeGenerationData SynthesisDiffusion modelVideo
π― What it does: Utilizing the inherent feature similarity in diffusion models, a cross-frame token correspondence is established to achieve temporal consistency and high-quality generation during video editing.
CriticEval: Evaluating Large-scale Language Model as Critic
Tian Lan (Beijing Institute of Technology), Xian-Ling Mao (Beijing Institute of Technology)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: The CRITICEVAL benchmark has been constructed and released, systematically evaluating the criticism ability of large models across four dimensions (feedback, comparison, correction, meta-feedback) covering nine task scenarios.
π― What it does: The CRONOS algorithm is proposed, which utilizes convex optimization to train two-layer ReLU networks and is extended to arbitrary depth networks through CRONOS-AM, achieving efficient training on large-scale datasets.
π― What it does: A multi-device collaborative terminal adaptation (CoLA) framework is proposed, achieving collaborative learning and knowledge sharing across devices and terminals with online adaptation.
π― What it does: This paper proposes a Cross-Modal Representation Flattening (CMRF) method aimed at addressing the issues of modality competition and single-modal flattening differences in multi-modal domain generalization.
π― What it does: A universal perturbation attack method called CMPS is proposed for cross-modal person re-identification (RGB and infrared/thermal imaging) systems, which can jointly optimize perturbations across different modalities to confuse retrieval results.
Cross-model Control: Improving Multiple Large Language Models in One-time Training
Jiayi Wu (East China Normal University), Ming Gao (East China Normal University)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes a cross-model control method (Cross-model Control, CMC) that enhances various large language models (LLMs) with a single training session. It achieves this by training a portable small 'Delta model' to adjust the output logits of different LLMs, enabling multi-model instruction following and forgetting tasks.
Cross-Scale Self-Supervised Blind Image Deblurring via Implicit Neural Representation
Tianjing Zhang (National University of Singapore), Hui Ji (South China University of Technology)
CodeRestorationContrastive LearningImage
π― What it does: A self-supervised blind image deblurring method that does not require real images is proposed, utilizing cross-scale consistency constraints to achieve simultaneous estimation of the image and the blur kernel.
π― What it does: A cross-video identity association pre-training framework CION is proposed, which learns cross-video identity invariance from a large number of unlabeled video portraits using progressive multi-level denoising and identity-guided self-distillation.
π― What it does: By randomly partitioning the dataset and sharing routing vectors, multiple sparse adjacency graphs (CSPG) are constructed. During queries, a two-stage (fast approach + cross-graph expansion) greedy Beam Search is employed, significantly reducing distance calculations and search steps, thereby accelerating nearest neighbor retrieval.
CultureLLM: Incorporating Cultural Differences into Large Language Models
CHENG LI, Xing Xie (Microsoft Research)
CodeClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: This paper proposes CultureLLM, a method for fine-tuning large language models (LLMs) with cultural awareness based on a small sample from the World Values Survey (WVS) and generating training data through semantic data augmentation.
CulturePark: Boosting Cross-cultural Understanding in Large Language Models
CHENG LI, Jindong Wang (William and Mary)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
π― What it does: Developed the CulturePark multi-agent platform, using LLM to generate cross-cultural dialogue data, and fine-tuning culture-specific LLMs with this data for tasks such as content moderation, cultural alignment, and cultural education.
CuMo: Scaling Multimodal LLM with Co-Upcycled Mixture-of-Experts
Jiachen Li (Georgia Tech and UIUC), Longyin Wen (ByteDance Inc)
CodeRecognitionData-Centric LearningTransformerLarge Language ModelMixture of ExpertsVision Language ModelImageTextMultimodality
π― What it does: Incorporating a sparse mixture of experts (MoE) module into multimodal large language models, expanding parameters from both the visual encoder and the visual-language connector, and employing a co-upcycling strategy for expert initialization, along with a three-stage training process (pre-training MLP, pre-fine-tuning, visual instruction tuning) and auxiliary load balancing loss, enhances the model's performance on visual tasks.
Curvature Clues: Decoding Deep Learning Privacy with Input Loss Curvature
Deepak Ravikumar (Purdue University), Kaushik Roy (Purdue University)
CodeSafty and PrivacyAdversarial AttackConvolutional Neural NetworkImage
π― What it does: This paper proposes a curvature-based theoretical framework and black-box membership inference attack by studying the distinguishability of training and testing samples based on input loss curvature, verifying its superiority.
Customized Multiple Clustering via Multi-Modal Subspace Proxy Learning
Jiawei Yao (University of Washington), Juhua Hu (University of Washington)
CodeRepresentation LearningTransformerLarge Language ModelContrastive LearningImageMultimodality
π― What it does: Multi-Sub is proposed, a customizable multi-clustering method based on multi-modal subspace proxy learning, which can automatically generate corresponding clustering results based on user interests (such as color, variety, etc.);
Customizing Language Models with Instance-wise LoRA for Sequential Recommendation
Xiaoyu Kong (University of Science and Technology of China), Xiangnan He (University of Science and Technology of China)
CodeRecommendation SystemTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsSequential
π― What it does: For the sequence recommendation task, an Instance-wise LoRA (iLoRA) framework is proposed to dynamically generate LoRA parameters for each user sequence during the fine-tuning process of LLM, addressing the negative transfer problem caused by uniform LoRA.
π― What it does: Designed and trained a 3D video VAE (CV-VAE) compatible with existing image VAEs (such as Stable Diffusion), achieving a continuous latent space for spatiotemporal compression that can be directly used for video generation models.
D-MiSo: Editing Dynamic 3D Scenes using Multi-Gaussians Soup
Joanna Waczynska, PrzemysΕaw Spurek (IDEAS National Centre for Research and Development)
CodeGaussian SplattingPoint Cloud
π― What it does: A dynamic 3D scene editing framework D-MiSo based on Gaussian Splatting is proposed, utilizing multi-Gaussian (core + sub-Gaussian) and triangle soup structures to achieve local editability and dynamic control of objects at any moment.
π― What it does: This paper proposes an unsupervised representation learning framework D2R2 based on diffusion models to address the few-shot classification problem in tabular data.
DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection
Haochen Li (Intelligent Software Research Center, Institute of Software, Chinese Academy of Sciences), Ling Li (Intelligent Software Research Center, Institute of Software, Chinese Academy of Sciences)
CodeObject DetectionDomain AdaptationTransformerVision Language ModelImage
π― What it does: Based on the Visual Language Model (VLM), a Domain-Aware Adapter (DA-Ada) is proposed to address the Domain Adaptation Object Detection (DAOD) problem.
DAGER: Exact Gradient Inversion for Large Language Models
Ivo Petrov (INSAIT Sofia University St Kliment Ohridski), Martin Vechev (ETH Zurich)
CodeFederated LearningAdversarial AttackTransformerLarge Language ModelText
π― What it does: This paper studies a gradient inversion attack for large language models called DAGER, which can accurately recover complete batches of text from shared gradients in federated learning.
DALD: Improving Logits-based Detector without Logits from Black-box LLMs
Cong Zeng (MBZUAI), Dongkuan Xu (North Carolina State University)
CodeClassificationRecognitionTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: The DALD framework is proposed, which achieves more accurate text source identification by aligning the distribution of the proxy model in black-box LLM detection.
π― What it does: Designed and implemented DapperFL, a federated learning framework for heterogeneous edge devices and domain shift, utilizing model fusion pruning and domain adaptive regularization to achieve personalized compression and domain generalization.
DARG: Dynamic Evaluation of Large Language Models via Adaptive Reasoning Graph
Zhehao Zhang (Dartmouth College), Diyi Yang (Stanford University)
CodeLarge Language ModelPrompt EngineeringMixture of ExpertsTextChain-of-Thought
π― What it does: Proposes the DARG framework, which utilizes the construction and perturbation of inference graphs to dynamically expand existing evaluation benchmarks, generating test samples with controllable complexity while maintaining language diversity and label correctness.
DarkSAM: Fooling Segment Anything Model to Segment Nothing
Ziqi Zhou (Huazhong University of Science and Technology), Hai Jin (Huazhong University of Science and Technology)
CodeSegmentationAdversarial AttackImage
π― What it does: We propose DarkSAM, which achieves the first universal adversarial perturbation attack on the Segment Anything Model (SAM) and its variants, rendering the model unable to correctly segment under any prompt.
DARNet: Dual Attention Refinement Network with Spatiotemporal Construction for Auditory Attention Detection
Sheng Yan (Anhui University), Zhao Lv (Anhui University)
CodeClassificationRecognitionConvolutional Neural NetworkTime SeriesBiomedical Data
π― What it does: A dual-attention refinement-based spatiotemporal construction network (DARNet) is proposed for EEG signal auditory attention detection.
DART-Math: Difficulty-Aware Rejection Tuning for Mathematical Problem-Solving
Yuxuan Tong (Tsinghua University), Junxian He (Hong Kong University of Science and Technology)
CodeGenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextChain-of-Thought
π― What it does: This paper studies the performance of large language models on mathematical reasoning tasks and proposes a difficulty-aware rejection tuning method called DART, which is used to construct more challenging synthetic training data and enhance the model's reasoning capabilities.
π― What it does: A distribution-level data valuation method based on Maximum Mean Discrepancy (MMD) suitable for the Huber model is proposed, which can directly assess and compare the data distribution value of different suppliers from limited samples.
Data Mixture Inference Attack: BPE Tokenizers Reveal Training Data Compositions
Jonathan Hayase (University of Washington), Noah A. Smith (University of Washington)
CodeLarge Language ModelText
π― What it does: By analyzing the merge list of the BPE tokenizer, the mixed proportions of its training data in terms of language, code, and domain are inferred.
Data-Driven Discovery of Dynamical Systems in Pharmacology using Large Language Models
Samuel Holt (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
CodeDrug DiscoveryTransformerLarge Language ModelBiomedical DataOrdinary Differential Equation
π― What it does: A framework named Data-Driven Discovery (D3) is proposed, which utilizes large language models (LLM) to iteratively discover and optimize interpretable dynamic system models through a model generation, feature acquisition, and evaluation loop, primarily used for pharmacokinetic (PK) modeling in pharmacology.
Alaia Solko-Breslin (University of Pennsylvania), Eric Wong (University of Pennsylvania)
CodeData-Centric LearningReinforcement Learning from Human FeedbackReinforcement LearningTextBenchmark
π― What it does: Proposes the ISED algorithm, which supports end-to-end learning of a 'neural program' composed of any black-box program (Python code or GPT-4 API calls) and neural networks;
Data-Efficient Operator Learning via Unsupervised Pretraining and In-Context Learning
Wuyang Chen (Simon Fraser University), Michael W. Mahoney (International Computer Science Institute)
CodeTransformerAuto EncoderTabularTime SeriesPhysics Related
π― What it does: Enhance the performance of PDE neural operators in data-scarce and OOD scenarios through unsupervised pre-training and context learning methods.
Data-faithful Feature Attribution: Mitigating Unobservable Confounders via Instrumental Variables
Qiheng Sun (Zhejiang University), Jinfei Liu (Zhejiang University)
CodeExplainability and InterpretabilityTabular
π― What it does: A method is proposed to achieve feature attribution of the data generation process using instrumental variables to train unconfounded models.
Dataset Decomposition: Faster LLM Training with Variable Sequence Length Curriculum
Hadi Pouransari (Apple), Oncel Tuzel (Apple)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes a Dataset Decomposition method that splits the original text collection into fixed-length subsequence buckets and uses variable sequence length (VSL) and length-based learning curves for sampling during training;
π― What it does: In the federated learning framework, multiple backdoors are implanted in the diffusion model using combination triggers (ComboTs) to achieve data stealing of a large amount of private image data from local clients.
π― What it does: A method called DC-Gaussian is proposed to generate high-fidelity new perspective renderings from in-vehicle dashcam videos, achieving the separation and elimination of occlusions (such as mirrors, phone mounts, stains, etc.) without reflections and obstructions.
DDK: Distilling Domain Knowledge for Efficient Large Language Models
Jiaheng Liu (Taobao and Tmall Group of Alibaba), Bo Zheng (Taobao and Tmall Group of Alibaba)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
π― What it does: A dynamic distillation framework DDK based on domain knowledge is proposed for knowledge transfer between large language models (teachers) and small language models (students); by dynamically adjusting the domain mixing ratio of the distillation dataset, the student gains more training in domains with significant performance gaps.
DDN: Dual-domain Dynamic Normalization for Non-stationary Time Series Forecasting
Tao Dai (Shenzhen University), Zexuan Zhu (Tsinghua University)
CodeAnomaly DetectionOptimizationTransformerTime Series
π― What it does: To address the non-stationarity problem in time series forecasting, a Dual-Domain Dynamic Normalization (DDN) module is proposed, which performs sliding normalization and denormalization on both the time domain and frequency domain sides, thereby achieving dynamic capture and correction of distribution drift.
π― What it does: This study investigates the impact of synthetic image pollution on online continual learning and proposes a method called ESRM to mitigate this effect.
Decompose, Analyze and Rethink: Solving Intricate Problems with Human-like Reasoning Cycle
Shangzi Xue (University of Science and Technology of China), Qi Liu (University of Science and Technology of China)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
π― What it does: The DeAR framework is proposed, which constructs a reasoning tree within large language models and generates, verifies, and updates the reasoning process layer by layer through the Decompose-Analyze-Rethink cycle, thereby achieving a more human-like solution to complex problems.
Decomposed Prompt Decision Transformer for Efficient Unseen Task Generalization
Hongling Zheng (Wuhan University), Dacheng Tao (Nanyang Technological University)
CodeMeta LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringSequential
π― What it does: A multi-task offline reinforcement learning framework DPDT based on Prompt is proposed, which uses pre-trained language model parameters for initialization and achieves efficient generalization to unseen tasks through prompt decomposition (cross-task and task-specific) and test-time alignment (TTA).
Decomposing and Interpreting Image Representations via Text in ViTs Beyond CLIP
Sriram Balasubramanian (University of Maryland), Soheil Feizi (University of Maryland)
CodeRetrievalExplainability and InterpretabilityTransformerContrastive LearningImageText
π― What it does: A general framework is proposed to automatically decompose the final representation of Vision Transformers into the contributions of various components (such as attention heads and MLPs) and map these contributions to the CLIP space through linear mapping for text interpretation. A continuous scoring function is designed to evaluate the importance of components for different image features.
π― What it does: A gradient optimization analysis of KL divergence loss is conducted, and an improved IKL loss is proposed to enhance adversarial robustness and knowledge distillation performance.
Deep Bayesian Active Learning for Preference Modeling in Large Language Models
Luckeciano Carvalho Melo, Yarin Gal (University of Oxford)
CodeRecommendation SystemOptimizationTransformerLarge Language ModelPrompt EngineeringText
π― What it does: A new Bayesian active learning method, BAL-PM, is proposed for efficiently collecting human preference labels in large-scale language models, significantly reducing the amount of required annotations.
π― What it does: A Deep Correlated Prompting method is designed to fine-tune large multimodal Transformers (CLIP) with learnable prompts to maintain model robustness in the absence of any modality.
π― What it does: This paper studies the application of Deep Equilibrium Networks (DEQ) to neural algorithm reasoning, directly solving the equilibrium points of graph neural networks to complete algorithm execution without needing to know the number of steps in the algorithm.
Deep Graph Neural Networks via Posteriori-Sampling-based Node-Adaptative Residual Module
Jingbo Zhou (Westlake University), Stan Z. Li (Emory University)
CodeClassificationGraph Neural NetworkGraph
π― What it does: This paper proposes a posterior sampling node adaptive residual module (PSNR) to alleviate the over-smoothing problem that occurs in deep Graph Neural Networks (GNNs);
Takanori Maehara (Roku), Hoang NT (University of Tokyo)
CodeGraph Neural NetworkGraph
π― What it does: This paper proposes a Deep Homomorphism Network (DHN) based on graph homomorphism, which achieves the expression of graph structures through a learnable nonlinear transformation and aggregation of the graph homomorphism numbers of a predefined set of subgraph patterns P;
π― What it does: This paper systematically compares the performance and generalization ability of classical optimization registration methods with deep learning registration (DLIR) methods in medical image registration, focusing on the differences between unsupervised and supervised training as well as domain transfer.
π― What it does: An incremental deep policy gradient algorithm called AVG is proposed, which does not require experience replay, target networks, or batch updates, enabling real-time learning on resource-constrained robots.
DeepDRK: Deep Dependency Regularized Knockoff for Feature Selection
Hongyu Shen (University of Illinois at Urbana Champaign), Zhizhen Zhao (University of Illinois at Urbana Champaign)
CodeTransformerGenerative Adversarial NetworkBiomedical Data
π― What it does: This paper proposes DeepDRK, a deep knockoff generation framework that utilizes Transformer and multi-source adversarial training, and subsequently enhances retrieval effectiveness through regularization perturbations.
DeepITE: Designing Variational Graph Autoencoders for Intervention Target Estimation
Hongyuan Tao (Ant Group), Jianguo Li (Ant Group)
CodeGraph Neural NetworkAuto EncoderGraph
π― What it does: This paper proposes DeepITE, a framework based on Variational Graph Autoencoders (VGAE) for collaboratively learning multiple causal graphs and various intervention targets, allowing for the rapid identification of new intervention targets without the need for retraining during inference.
π― What it does: By integrating Eulerian grid features with Lagrangian particle tracking, DeepLag is proposed to achieve autoregressive prediction of fluids.
DeeR-VLA: Dynamic Inference of Multimodal Large Language Models for Efficient Robot Execution
Yang Yue (Tsinghua University), Gao Huang (Tsinghua University)
CodeComputational EfficiencyRobotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelTextMultimodalityBenchmark
π― What it does: A dynamic early stopping multimodal large language model (DeeR) is designed, which adaptively adjusts the model size based on the robot operation scenario through multi-level intermediate exits, significantly reducing computational and memory requirements.
π― What it does: A robust machine unlearning framework called AdvUnlearn based on adversarial training (AT) is proposed to safely erase harmful concepts in diffusion models and resist adversarial prompt attacks.