ICLR 2024 Papers with AI Summaries
International Conference on Learning Representations · 2260 papers
→ ICLR 2024 papers with code (1064)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 2260 ICLR 2024 papers by keyword, author or institution —
free trial on arXivSub.
"What Data Benefits My Classifier?" Enhancing Model Performance and Interpretability through Influence-Based Data Selection
Anshuman Chhabra (University of California), Hongfu Liu (Brandeis University)
ClassificationExplainability and InterpretabilityImageTextTabular
🎯 What it does: This paper uses a method that combines influence functions and decision trees to evaluate the impact of training samples in the feature space on the utility, fairness, and robustness of classifiers. Based on this, a general data pruning algorithm is proposed that can enhance performance while keeping the model unchanged.
#InsTag: Instruction Tagging for Analyzing Supervised Fine-tuning of Large Language Models
Keming Lu (Alibaba DAMO Academy), Jingren Zhou (Alibaba DAMO Academy)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes an open instruction labeling method based on ChatGPT called INSTAG, which performs fine-grained labeling of instructions in publicly available SFT datasets, quantifying the diversity and complexity of the instructions, and utilizes these labels for efficient data sampling to train a TAGLM model with significant alignment effects.
$\alpha$TC-VAE: On the relationship between Disentanglement and Diversity
Cristian Meo (Delft University of Technology), Justin Dauwels (Delft University of Technology)
GenerationData SynthesisReinforcement LearningAuto EncoderImage
🎯 What it does: A new variational autoencoder, α-TCVAE, is proposed, which achieves better separation and information retention by optimizing the lower bound of the joint total correlation.
$\infty$-Diff: Infinite Resolution Diffusion with Subsampled Mollified States
Sam Bond-Taylor (Durham University), Chris G. Willcocks (Durham University)
GenerationData SynthesisDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes a diffusion model defined on infinite-dimensional Hilbert space, ∞-Diff, and trains it by learning continuous functions that can generate high-quality images at any resolution through random sampling of subset coordinates.
$\mathbb{D}^2$ Pruning: Message Passing for Balancing Diversity & Difficulty in Data Pruning
Adyasha Maharana (University of North Carolina), Mohit Bansal (University of North Carolina)
Data-Centric LearningGraph Neural NetworkImageText
🎯 What it does: This paper proposes a core set (coreset) selection method based on graph message passing called D2PRUNING, which aims to balance data diversity and difficulty.
$\mathcal{B}$-Coder: Value-Based Deep Reinforcement Learning for Program Synthesis
Zishun Yu (University of Illinois Chicago), Hongxia Yang (ByteDance Inc.)
GenerationAI Code AssistantTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study presents B-Coder, a value function-based deep reinforcement learning framework for generating accurate executable programs from natural language descriptions.
$\pi$2vec: Policy Representation with Successor Features
Gianluca Scarpellini (Istituto Italiano di Tecnologia), Misha Denil (Google DeepMind)
Robotic IntelligenceTransformerReinforcement LearningContrastive LearningMultimodality
🎯 What it does: A π2 vec method is proposed, which vectorizes the black-box policy using successful features and a base model on offline data, thereby achieving efficient policy evaluation and selection.
$\texttt{NAISR}$: A 3D Neural Additive Model for Interpretable Shape Representation
Yining Jiao (University of North Carolina at Chapel Hill), Marc Niethammer (University of North Carolina at Chapel Hill)
Explainability and InterpretabilityRepresentation LearningPoint CloudMeshAlzheimer's Disease
🎯 What it does: This paper proposes NAISR, an interpretable 3D shape representation method based on deep implicit functions, which can simultaneously consider the effects of multidimensional covariates on shape.
$t^3$-Variational Autoencoder: Learning Heavy-tailed Data with Student's t and Power Divergence
Juno Kim (University of Tokyo), Joong-Ho Won (Seoul National University)
GenerationData SynthesisAuto EncoderImage
🎯 What it does: This paper studies a variational autoencoder using Student's t-distribution, called t^3 VAE, to alleviate over-regularization and better model heavy-tailed data.
3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining
Siming Yan (University of Texas at Austin), Qixing Huang (Peking University)
ClassificationObject DetectionSegmentationRepresentation LearningTransformerAuto EncoderPoint Cloud
🎯 What it does: In the self-supervised pre-training of 3D point clouds, MaskFeat3D is proposed, which predicts the surface normal vectors and curvature variations of masked points under the Masked Autoencoder (MAE) framework, rather than recovering point coordinates.
3D Reconstruction with Generalizable Neural Fields using Scene Priors
Yang Fu (University of California), Sifei Liu (NVIDIA)
GenerationDepth EstimationNeural Radiance FieldPoint CloudMesh
🎯 What it does: A generalizable Neural Field Scene Prior Network (NFP) is proposed, capable of achieving fast and high-quality 3D scene reconstruction from single-view RGB-D input, and supports novel view synthesis from a single image.
3D-Aware Hypothesis & Verification for Generalizable Relative Object Pose Estimation
Chen Zhao (École Polytechnique Fédérale de Lausanne), Mathieu Salzmann (École Polytechnique Fédérale de Lausanne)
Pose EstimationTransformerImage
🎯 What it does: A hypothesis-validation based 3D perception framework is proposed to estimate the relative pose of unseen objects in a query image with only a single reference view.
A 2-Dimensional State Space Layer for Spatial Inductive Bias
Ethan Baron (Tel Aviv University), Lior Wolf (Tel Aviv University)
Convolutional Neural NetworkTransformerImage
🎯 What it does: A two-dimensional state space layer (2D-SSM) is proposed, which can be inserted before and after different modules of visual Transformers and CNNs to help the model adaptively learn two-dimensional spatial features.
A Benchmark for Learning to Translate a New Language from One Grammar Book
Garrett Tanzer (Google), Luke Melas-Kyriazi (Stanford University)
TransformerLarge Language ModelPrompt EngineeringTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: Created the MTOB benchmark to evaluate the ability of large language models to learn and translate a new language from a Kalamang grammar book.
A Benchmark Study on Calibration
Linwei Tao (University of Sydney), Chang Xu (University of Sydney)
ClassificationNeural Architecture SearchConvolutional Neural NetworkTransformerImageBenchmark
🎯 What it does: A calibration benchmark dataset consisting of 117,702 CNN models and 11 types of Transformers was constructed based on the NATS-Bench NAS search space. The dataset was evaluated on 102 calibration metrics across CIFAR-10, CIFAR-100, and ImageNet-16-120, systematically studying the calibration properties of models, metric reliability, post-processing methods, the impact of bin size, and the influence of architectural design on calibration.
A Black-box Approach for Non-stationary Multi-agent Reinforcement Learning
Haozhe Jiang (Institute for Interdisciplinary Information Sciences Tsinghua University), Simon Shaolei Du
Reinforcement Learning
🎯 What it does: This paper proposes a black-box method for learning equilibria in non-stationary multi-agent reinforcement learning (MARL) and provides upper bounds on sublinear dynamic equilibrium loss under known or unknown non-stationarity metrics.
A Branching Decoder for Set Generation
Zixian Huang (Nanjing University), Gong Cheng (Nanjing University)
GenerationTransformerSupervised Fine-TuningText
🎯 What it does: A branching decoder is proposed, capable of generating multiple text sequences in parallel during a single decoding process, avoiding the concatenation order bias of traditional sequential decoders.
A Characterization Theorem for Equivariant Networks with Point-wise Activations
Marco Pacini (Fondazione Bruno Kessler), Gabriele Santin (University of Venice)
🎯 What it does: This paper proposes and proves a complete representation theorem to describe all representations and activation combinations that equivariant networks can accommodate under pointwise activation functions.
A Cognitive Model for Learning Abstract Relational Structures from Memory-based Decision-Making Tasks
Haruo Hosoya (ATR International)
Representation LearningReinforcement LearningImage
🎯 What it does: A memory-based cognitive model (ARDMO) is proposed to learn abstract relational structures from decision-making tasks and achieve cross-domain generalization.
A Data-Driven Measure of Relative Uncertainty for Misclassification Detection
Eduardo Dadalto Câmara Gomes (Universite Paris-Saclay), Pablo Piantanida (Quebec AI Institute)
ClassificationAnomaly DetectionConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A relative uncertainty measure (REL-U) based on positive and negative sample learning is proposed for misclassification detection.
A differentiable brain simulator bridging brain simulation and brain-inspired computing
Chaoming Wang (Peking University), Si Wu (Peking University)
Spiking Neural NetworkReinforcement LearningTime SeriesBiomedical Data
🎯 What it does: BrainPy has been developed—a differentiable brain simulator based on JAX, integrating sparse, event-driven, and JIT connection technologies, supporting scalable simulation and training from discrete synapses to multi-scale brain networks.
A Differentially Private Clustering Algorithm for Well-Clustered Graphs
Weiqiang He (University of Science and Technology of China), Pan Peng (University of Science and Technology of China)
OptimizationSafty and PrivacyGraph Neural NetworkGaussian SplattingGraph
🎯 What it does: A differential privacy clustering algorithm for well-clustered graphs is proposed, which can approximately recover the true clustering structure under the premise of (ϵ,δ)-DP.
A Discretization Framework for Robust Contextual Stochastic Optimization
Rares C Cristian, Georgia Perakis (Massachusetts Institute of Technology)
OptimizationTabularTime Series
🎯 What it does: This paper proposes a robust contextual stochastic optimization framework based on feasible domain discretization, which directly learns the probability of the optimal decision subset from the data, thereby providing robust decisions.
A Dynamical View of the Question of Why
Mehdi Fatemi (Microsoft Research), Sindhu C. M. Gowda
Reinforcement LearningTime SeriesSequential
🎯 What it does: A causal inference framework based on dynamic processes is proposed, which uses reinforcement learning to directly estimate the expected grit and reachability of events from observation sequences, and quantifies the causal contributions of each state/action component through decomposition formulas.
A Fast and Provable Algorithm for Sparse Phase Retrieval
Jian-Feng CAI, Jiaxi Ying (Hong Kong University of Science and Technology)
OptimizationTabular
🎯 What it does: A second-order hard thresholding algorithm is proposed for sparse phase recovery;
A Flexible Generative Model for Heterogeneous Tabular EHR with Missing Modality
Huan He (University of Pennsylvania), Joyce Ho
GenerationData SynthesisRecurrent Neural NetworkDiffusion modelTabularTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a generative framework named FLEXGEN-EHR, which can simultaneously handle the static features and time series features of tabular electronic health records (EHR), and can still synthesize high-quality data even when a certain modality is missing;
A Foundation Model for Error Correction Codes
Yoni Choukroun (Tel Aviv University), Lior Wolf (Tel Aviv University)
Transformer
🎯 What it does: Designed and trained the first universal foundational model FECCT that can perform zero-shot or fine-tuning for any code length and any code family, achieving efficient decoding on various linear block codes.
A Framework for Inference Inspired by Human Memory Mechanisms
Xiangyu Zeng (University of Electronic Science and Technology of China), Zhicheng Zhang (University of Electronic Science and Technology of China)
TransformerText
🎯 What it does: Proposes the PMI framework: a three-module structure of perception, memory, and reasoning, constructing a dual-layer structure of working memory (WM) and long-term memory (LTM), and achieving information updating and integration through competitive writing, outer product association, and content retrieval.
A General Framework for User-Guided Bayesian Optimization
Carl Hvarfner (Lund University), Luigi Nardi (Lund University)
OptimizationHyperparameter SearchTabularBenchmark
🎯 What it does: A Bayesian Optimization framework named ColaBO is proposed, allowing users to inject prior knowledge about function properties (such as optimal location, optimal value, or preference relations) into the posterior distribution of the Gaussian Process, thereby directly reflecting user experience in sampling and acquisition functions.
A Good Learner can Teach Better: Teacher-Student Collaborative Knowledge Distillation
Ayan Sengupta (Indian Institute of Technology), Tanmoy Chakraborty (Indian Institute of Technology)
Knowledge DistillationMeta LearningTransformerReinforcement LearningTextBenchmark
🎯 What it does: A Meta-Policy knowledge distillation framework called MPDistil has been developed, allowing teachers and students to collaborate and compete during the distillation process, thereby enhancing the performance of both.
A Graph is Worth 1-bit Spikes: When Graph Contrastive Learning Meets Spiking Neural Networks
Jintang Li (Sun Yat-sen University), Liang Chen (Sun Yat-sen University)
Computational EfficiencyRepresentation LearningGraph Neural NetworkSpiking Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes SPIKEGCL, a framework that utilizes Spiking Neural Networks (SNN) for graph contrastive learning, capable of learning 1-bit sparse node representations and supporting large-scale unsupervised graph learning.
A Hard-to-Beat Baseline for Training-free CLIP-based Adaptation
Zhengbo Wang (University of Science and Technology of China), Tieniu Tan (Chinese Academy of Sciences)
ClassificationDomain AdaptationVision Language ModelContrastive LearningImage
🎯 What it does: This paper proposes a training-free adaptation method for CLIP, which directly estimates class means and covariances from a small number of samples using Gaussian Discriminant Analysis (GDA), constructs a linear classifier, and integrates it with the original zero-shot classifier of CLIP.
A Hierarchical Bayesian Model for Few-Shot Meta Learning
Minyoung Kim (Samsung AI Center), Timothy Hospedales
Meta LearningImageStochastic Differential Equation
🎯 What it does: A hierarchical Bayesian model is proposed for few-shot meta-learning, designing fully parametric high-level global variables and local variables for each task, and providing a variational inference algorithm with a closed-form solution;
A Lie Group Approach to Riemannian Batch Normalization
Ziheng Chen (University of Trento), Nicu Sebe (University of Trento)
Time SeriesSequentialBenchmark
🎯 What it does: A unified Lie group batch normalization framework, LieBN, is proposed, which can control the mean and variance on any left-invariant Lie group, and implement batch normalization of various metrics on SPD manifolds through parameterized deformed Lie groups.
A Lightweight Method for Tackling Unknown Participation Statistics in Federated Averaging
Shiqiang Wang (IBM T. J. Watson Research Center), Mingyue Ji (University of Utah)
Federated LearningImage
🎯 What it does: A lightweight method called FedAU is proposed, which adaptively adjusts the aggregation weights based on each client's participation history to address the bias problem caused by unknown participation statistics in federated learning.
A Linear Algebraic Framework for Counterfactual Generation
Jong-Hoon Ahn (Otsuka Pharmaceutical Development and Commercialization), Akshay Vashist (Otsuka Pharmaceutical Development and Commercialization)
GenerationData SynthesisOptimizationTabularTime Series
🎯 What it does: A framework based on linear algebra is proposed to generate counterfactual longitudinal data that perfectly matches the pre-treatment period, and this data is used to estimate individualized treatment effects.
A Multi-Level Framework for Accelerating Training Transformer Models
Longwei Zou (Tsinghua University), Yangdong Deng (Tsinghua University)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A multi-level training framework is proposed, which significantly reduces the training cost of Transformer models by utilizing the coarse and fine tuning of model width and depth.
A Mutual Information Perspective on Federated Contrastive Learning
Christos Louizos (Qualcomm AI Research), Denis Korzhenkov (Qualcomm AI Research)
Federated LearningRepresentation LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: The paper addresses the unsupervised and semi-supervised pre-training problem in federated learning by comparing the similarity of features and proposes a Federated SimCLR method that combines SimCLR with user verification (UV).
A Neural Framework for Generalized Causal Sensitivity Analysis
Dennis Frauen (LMU Munich), Mihaela van der Schaar (University of Cambridge)
Flow-based ModelTabularBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes NEURALCSA, a general causal sensitivity analysis framework based on neural networks, which can provide effective upper and lower bounds under different sensitivity models, types of treatments (discrete/continuous), and causal queries (such as CATE, distribution effects, multiple outcomes).
A Newborn Embodied Turing Test for Comparing Object Segmentation Across Animals and Machines
Manju Garimella, Samantha Marie Waters Wood
Object DetectionSegmentationConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: This study proposes the 'Neonatal Embodiment Turing Test' (NETT) for comparing newborn animals and machine learning algorithms under the same environment and tasks, using the single object segmentation task of neonatal chicks and artificial intelligence agents as an experimental example.
A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models
Haoran Xu (Johns Hopkins University), Hany Hassan Awadalla (Microsoft)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Two-stage fine-tuning of LLaMA-2: first, fine-tune with multilingual monolingual data to enhance cross-lingual capabilities, and then fine-tune with a small amount of high-quality parallel data to obtain the ALMA model, which significantly improves performance on multilingual translation tasks.
A path-norm toolkit for modern networks: consequences, promises and challenges
Antoine Gonon (University of Lyon), Rémi Gribonval (University of Lyon)
Convolutional Neural NetworkImage
🎯 What it does: A set of path norm tools suitable for general DAG ReLU networks is proposed, along with a universal generalization bound based on these tools.
A Plug-and-Play Image Registration Network
Junhao Hu (Washington University in St Louis), Ulugbek Kamilov
OptimizationConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a plugin-based prior (PnP) deformable image registration network called PIRATE, and further fine-tunes it using a deep equilibrium model (DEQ) to create PIRATE+, achieving a collaborative optimization of data consistency constraints in the registration field and learned CNN priors.
A Poincaré Inequality and Consistency Results for Signal Sampling on Large Graphs
Thien Le (Massachusetts Institute of Technology), Stefanie Jegelka
ClassificationOptimizationGraph Neural NetworkGraph
🎯 What it does: A signal sampling theory based on graphons is proposed, providing the Poincaré inequality and consistency proof, and designing a scalable sampling algorithm;
A Policy Gradient Method for Confounded POMDPs
Mao Hong (Johns Hopkins University), Yanxun Xu (Johns Hopkins University)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper proposes a policy gradient method for confounded POMDPs in an offline setting. It first achieves non-parametric identification of the gradient through a bridge function, then uses a max-min estimator to obtain the gradient estimate, and finally performs gradient ascent optimization on offline data.
A Precise Characterization of SGD Stability Using Loss Surface Geometry
Gregory Dexter (Purdue University), Rajiv Khanna (Purdue University)
Optimization
🎯 What it does: A precise characterization of the linear stability of stochastic gradient descent (SGD) near optimal points is conducted, proposing and utilizing the Hessian consistency measure σ to determine the convergence or divergence of SGD.
A Primal-Dual Approach to Solving Variational Inequalities with General Constraints
Tatjana Chavdarova (University of California), Michael Jordan
OptimizationGenerative Adversarial NetworkImage
🎯 What it does: A new primal-dual method (ACVI) is proposed and analyzed for solving variational inequalities (VI) with general constraints, and it is improved into two variants: I-ACVI for approximately solving subproblems and P-ACVI for simple inequality constraints.
A Probabilistic Framework for Modular Continual Learning
Lazar Valkov (MIT IBM Watson AI Lab), Charles Sutton (University of Edinburgh)
SequentialBenchmark
🎯 What it does: A modular continual learning framework PICLE is proposed, which uses probabilistic models for rapid evaluation of module combinations.
A Progressive Training Framework for Spiking Neural Networks with Learnable Multi-hierarchical Model
Zecheng Hao (Peking University), Tiejun Huang (Peking University)
Spiking Neural NetworkImage
🎯 What it does: A learnable multi-level (LM-H) neuron model has been designed, and a training framework based on progressive STBP and an efficient hybrid/time-slicing training scheme have been proposed to address the shortcomings of traditional LIF models in deep gradient computation and historical information extraction.
A Quadratic Synchronization Rule for Distributed Deep Learning
Xinran Gu (Institute for Interdisciplinary Information Sciences Tsinghua University), Longbo Huang (Institute for Interdisciplinary Information Sciences Tsinghua University)
OptimizationComputational EfficiencyConvolutional Neural NetworkTransformerImage
🎯 What it does: A method called Quadratic Synchronization Rule (QSR) is proposed for dynamically adjusting the gradient synchronization period in distributed deep learning to reduce communication overhead and improve final test accuracy.
A Real-World WebAgent with Planning, Long Context Understanding, and Program Synthesis
Izzeddin Gur (Google DeepMind), Aleksandra Faust (Google DeepMind)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AIText
🎯 What it does: Designed and implemented a WebAgent based on LLM, achieving planning, long HTML understanding, and program synthesis, completing real website command execution.
A Recipe for Improved Certifiable Robustness
Kai Hu (Carnegie Mellon University), Matt Fredrikson (Carnegie Mellon University)
ClassificationGenerationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper systematically explores the design space of Lipschitz constrained training, proposing large-scale residual dense layers (CHORD) and Cholesky orthogonalization techniques, combined with generative model data augmentation, significantly improving the deterministic robust accuracy (VRA).
A representation-learning game for classes of prediction tasks
Neria Uzan (Technion Israel Institute of Technology), Nir Weinberger (Technion Israel Institute of Technology)
Representation LearningImage
🎯 What it does: A game theory-based framework is proposed to learn low-dimensional representations using prior knowledge of existing prediction task classes; optimal representations and minimized maximum regret are provided under pure and mixed strategies in the case of linear mean squared error (MSE); a general iterative algorithm is designed to approximate the optimal mixed representation using random representations and multiplier weight updates.
A Restoration Network as an Implicit Prior
Yuyang Hu (Washington University in St. Louis), Ulugbek Kamilov (Washington University in St. Louis)
RestorationSuper ResolutionImage
🎯 What it does: A general Deep Recovery Prior (DRP) method is proposed to solve image inversion problems (such as deblurring and super-resolution) by utilizing pre-trained recovery networks (e.g., super-resolution models) as implicit priors.
A ROBUST DIFFERENTIAL NEURAL ODE OPTIMIZER
Panagiotis Theodoropoulos (Georgia Institute of Technology), Evangelos Theodorou
OptimizationAdversarial AttackImageOrdinary Differential Equation
🎯 What it does: This paper studies a robust neural ODE optimizer GTSONO based on min-max second-order differential dynamic programming to enhance the model's robustness against adversarial attacks.
A Semantic Invariant Robust Watermark for Large Language Models
Aiwei Liu (Tsinghua University), Lijie Wen (Tsinghua University)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: A robust watermarking algorithm based on contextual semantic invariance is proposed, which uses an embedding model to generate watermark logits, maintaining detection accuracy under semantic-preserving attacks such as text rewriting and synonym replacement.
A Simple and Effective Pruning Approach for Large Language Models
Mingjie Sun (Carnegie Mellon University), J Zico Kolter
TransformerLarge Language ModelText
🎯 What it does: A pruning method for large language models (LLM) called Wanda is proposed, which evaluates importance by multiplying the absolute value of weights with the L2 norm of the corresponding input activations for each output dimension, directly setting low-importance weights to zero without the need for retraining.
A Simple and Scalable Representation for Graph Generation
Yunhui Jang (Pohang University of Science and Technology), Sungsoo Ahn (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisRecurrent Neural NetworkGraph Neural NetworkPoint CloudGraph
🎯 What it does: A graph generation representation method based on edge lists, named GEEL, is proposed, and scalable graph generation is achieved through the serialization of edge lists.
A Simple Interpretable Transformer for Fine-Grained Image Classification and Analysis
DIPANJYOTI PAUL, Wei-Lun Chao
ClassificationExplainability and InterpretabilityTransformerImage
🎯 What it does: An interpretable Transformer classifier called INTR is proposed, which allows each category to actively search for itself in the image using class-specific queries, thereby naturally generating attention heatmaps during inference.
A Simple Romance Between Multi-Exit Vision Transformer and Token Reduction
Dongyang Liu (Key Lab of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences), Xilin CHEN
ClassificationComputational EfficiencyKnowledge DistillationTransformerImage
🎯 What it does: This paper proposes the Multi-Exit Token Reduction (METR) method, which enhances the [CLS] attention weights as a token importance assessment metric in Vision Transformers by adding multiple early exit heads and self-distillation, thereby achieving more effective token pruning.
A Stable, Fast, and Fully Automatic Learning Algorithm for Predictive Coding Networks
Tommaso Salvatori (VERSES AI Research Lab), Thomas Lukasiewicz (Vienna University of Technology)
ClassificationGenerationTransformerLarge Language ModelImageText
🎯 What it does: The Incremental Prediction Coding (iPC) algorithm is proposed, which simultaneously updates weights and infers hidden layers, eliminating external control signals and significantly improving training efficiency and stability.
A Statistical Analysis of Wasserstein Autoencoders for Intrinsically Low-dimensional Data
Saptarshi Chakraborty (University of California Berkeley), Peter Bartlett
GenerationData SynthesisAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: The statistical properties of the Wasserstein Autoencoder (WAE) on low-dimensional structured data are theoretically analyzed, deriving an upper bound for the expected excess risk and proving that this upper bound is related to the intrinsic dimension of the data (Minkowski dimension), rather than the dimension of the high-dimensional feature space.
A Study of Bayesian Neural Network Surrogates for Bayesian Optimization
Yucen Lily Li (New York University), Andrew Gordon Wilson (New York University)
OptimizationRepresentation LearningReinforcement LearningTabular
🎯 What it does: The system evaluates various Bayesian Neural Networks (BNN) as surrogate models for Bayesian optimization and compares them with traditional Gaussian Process (GP) benchmarks, covering a range of scenarios from low-dimensional synthetic functions to high-dimensional real-world tasks.
A Sublinear Adversarial Training Algorithm
Yeqi Gao (Tsinghua University), Yitan Wang (Yale University)
OptimizationComputational EfficiencyAdversarial Attack
🎯 What it does: A sublinear adversarial training algorithm for two-layer neural networks is proposed, utilizing the shifted ReLU activation function to activate only a sublinear number of neurons in each iteration, and accelerating forward/backward propagation through half-space retrieval data structures.
A Symmetry-Aware Exploration of Bayesian Neural Network Posteriors
Olivier Laurent (Paris-Saclay University), Gianni Franchi (ENSTA Paris)
ClassificationRecognitionOptimizationImage
🎯 What it does: Systematically explore and evaluate the structure and quality of the posterior distribution of deep Bayesian neural networks (BNN), focusing on the impact of weight space permutation and scale symmetry on the posterior, and based on this, propose methods to assess the quality of posterior approximations.
A Topological Perspective on Demystifying GNN-Based Link Prediction Performance
Yu Wang (Vanderbilt University), Tyler Derr (Vanderbilt University)
Graph Neural NetworkGraph
🎯 What it does: This study investigates the performance differences of different nodes in the link prediction (LP) task using GNNs, proposing the Topological Concentration (TC) metric to measure the relationship between local topology of nodes and LP performance, and explores the computation, approximation, and impact of TC on performance.
A Unified and General Framework for Continual Learning
Zhenyi Wang (University of Maryland), Heng Huang (JD Explore Academy)
ClassificationImage
🎯 What it does: A unified continuous learning (CL) framework is proposed, and within this framework, a 'refresh learning' (unlearn–relearn) plugin is designed to enhance the performance of existing CL methods.
A Unified Framework for Bayesian Optimization under Contextual Uncertainty
Sebastian Shenghong Tay (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
OptimizationTabularTime Series
🎯 What it does: This paper proposes a unified framework for handling Bayesian optimization under context uncertainty (BOCU) and presents a general Thompson sampling algorithm capable of optimizing any uncertain objective.
A Unified Sampling Framework for Solver Searching of Diffusion Probabilistic Models
Enshu Liu (Tsinghua University), Yu Wang (Tsinghua University)
GenerationData SynthesisDiffusion modelImageOrdinary Differential Equation
🎯 What it does: A Unified Sampling Framework (USF) is designed, and the sampling quality of the diffusion model is optimized through an automatic search method (S3) with very few NFE.
A unique M-pattern for micro-expression spotting in long videos
Jinxuan Wang (South China University of Technology), Tong Zhang (South China University of Technology)
RecognitionObject DetectionOptical FlowVideo
🎯 What it does: Proposes a block-wise Mean Directional Motion Optical Flow (MDMO) feature based on the skipk-frame strategy, combined with the M-pattern and a complete set of micro-expression localization rules, utilizing pixelmatch to dynamically update the reference frame for facial alignment.
A Variational Framework for Estimating Continuous Treatment Effects with Measurement Error
Erdun Gao (University of Melbourne), Mingming Gong (Mohamed bin Zayed University of Artificial Intelligence)
Tabular
🎯 What it does: The paper proposes a variational framework to estimate the average dose response function (ADRF) for continuous treatment amounts and can handle measurement errors in treatment variables.
A Variational Perspective on Solving Inverse Problems with Diffusion Models
Morteza Mardani (NVIDIA Inc), Arash Vahdat (NVIDIA Inc)
RestorationDiffusion modelImage
🎯 What it does: A backward sampling method based on variational inference for diffusion models (RED-diff) is proposed for general inverse problems.
A Versatile Causal Discovery Framework to Allow Causally-Related Hidden Variables
Xinshuai Dong (Carnegie Mellon University), Kun Zhang (Mohamed bin Zayed University of Artificial Intelligence)
Tabular
🎯 What it does: A variable causal discovery framework RLCD is proposed, which utilizes the rank information of the covariance matrix to identify causal structures containing interrelated hidden variables.
Abstractors and relational cross-attention: An inductive bias for explicit relational reasoning in Transformers
Awni Altabaa (Yale University), John Lafferty (Yale University)
ClassificationGenerationOptimizationTransformerSequential
🎯 What it does: An extended module of Transformer called Abstractor is proposed, which achieves explicit relational reasoning through relational cross-attention, enabling decoupling between object features and relational information, thereby improving sample efficiency in relational reasoning tasks.
Accelerated Convergence of Stochastic Heavy Ball Method under Anisotropic Gradient Noise
Rui Pan (Hong Kong University of Science and Technology), Tong Zhang (University of Illinois Urbana-Champaign)
OptimizationImageTabular
🎯 What it does: This paper studies the accelerated convergence of the stochastic heavy ball method under anisotropic gradient noise conditions, filling a gap in theoretical analysis.
Accelerated Sampling with Stacked Restricted Boltzmann Machines
Jorge Fernandez-de-Cossio-Diaz (Ecole Normale Supérieure), Remi Monasson
OptimizationComputational EfficiencyRepresentation LearningProtein Structure PredictionImageSequential
🎯 What it does: This paper proposes an accelerated sampling method based on stacked restricted Boltzmann machines—Stacked Tempering—by exchanging hidden/visible layer states between different layers of RBMs to enhance sampling efficiency.
Accelerating Data Generation for Neural Operators via Krylov Subspace Recycling
Hong Wang (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
OptimizationComputational EfficiencyTabularPhysics Related
🎯 What it does: The SKR algorithm is proposed, utilizing sorting and Krylov subspace reuse to accelerate the solution of linear equations in the neural operator training data generation process.
Accelerating Distributed Stochastic Optimization via Self-Repellent Random Walks
Jie Hu (North Carolina State University), Do Young Eun (North Carolina State University)
OptimizationGraph
🎯 What it does: This paper studies and proposes a distributed stochastic optimization algorithm SA-SRRW based on self-regulating random walks (SRRW), and provides its almost convergence and central limit theorem.
Accelerating Sinkhorn algorithm with sparse Newton iterations
Xun Tang (Stanford University), Lexing Ying (Stanford University)
OptimizationComputational EfficiencyImage
🎯 What it does: The Sinkhorn-Newton-Sparse (SNS) algorithm is proposed, which introduces early stopping and a sparse Newton subroutine based on the Sinkhorn algorithm to accelerate the solution of entropy-regularized optimal transport;
Accurate and Scalable Estimation of Epistemic Uncertainty for Graph Neural Networks
Puja Trivedi (University of Michigan), Jayaraman J. Thiagarajan
ClassificationDomain AdaptationGraph Neural NetworkGraph
🎯 What it does: A G-Δ UQ training framework is proposed, which enhances the intrinsic uncertainty estimation of Graph Neural Networks (GNNs) through graph-specific anchoring strategies (node feature anchoring, hidden layer anchoring, output layer anchoring), achieving better calibration and OOD detection under distribution shifts.
Accurate Forgetting for Heterogeneous Federated Continual Learning
Abudukelimu Wuerkaixi (Tsinghua University), Masashi Sugiyama (University of Tokyo)
Federated LearningKnowledge DistillationFlow-based ModelImage
🎯 What it does: In a federated continual learning environment with distribution heterogeneity and potentially unrelated tasks, an adaptive knowledge filtering mechanism using 'accurate forgetting' is proposed to maintain performance on previous tasks.
Accurate Retraining-free Pruning for Pretrained Encoder-based Language Models
Seungcheol Park (Seoul National University), U Kang (Seoul National University)
CompressionTransformerLarge Language ModelText
🎯 What it does: This paper proposes K-prune, a structured pruning method that does not require retraining, for compressing pre-trained encoder-based language models.
Achieving Fairness in Multi-Agent MDP Using Reinforcement Learning
Peizhong Ju (Ohio State University), Ness Shroff
OptimizationReinforcement Learning
🎯 What it does: A fair policy based on reinforcement learning is proposed in multi-agent MDPs, aiming to maximize the α-fairness function rather than simply the total reward.
Achieving Human Parity in Content-Grounded Datasets Generation
Asaf Yehudai (IBM Israel Research Lab), Leshem Choshen (IBM Israel Research Lab)
GenerationData SynthesisTransformerLarge Language ModelReinforcement LearningTextBiomedical Data
🎯 What it does: Proposes the Genie framework, which utilizes LLM to automatically generate and filter high-quality content-related data (such as long-form Q&A, summaries, information extraction).
Achieving Sample and Computational Efficient Reinforcement Learning by Action Space Reduction via Grouping
Yining Li (Ohio State University), Ness Shroff (Ohio State University)
Computational EfficiencyReinforcement Learning
🎯 What it does: The paper divides the action space into several groups based on action transition probabilities and reward similarities, and utilizes a linear decomposition model for model estimation and planning at the grouping level, proposing a reinforcement learning framework that reduces sample and computational complexity while maintaining approximately optimal performance.
ACRF: Compressing Explicit Neural Radiance Fields via Attribute Compression
Guangchi Fang (Sun Yat-sen University), Yulan Guo (National University of Defense Technology)
CompressionNeural Radiance FieldPoint Cloud
🎯 What it does: This study investigates how to compress explicit NeRF models, treating it as a 3D data compression task, and proposes the ACRF framework.
Active Retrosynthetic Planning Aware of Route Quality
Luotian Yuan (Zhejiang University), Fei Wu (Zhejiang University)
Drug DiscoveryReinforcement LearningAgentic AITabularBenchmark
🎯 What it does: A proactive backtracking synthesis planning framework, ARP, is proposed, which utilizes an actor-critic model to actively decide whether to query reaction quality, thereby evaluating and planning high-quality synthesis routes.
Active Test-Time Adaptation: Theoretical Analyses and An Algorithm
Shurui Gui (Texas A and M University), Shuiwang Ji (Texas A and M University)
Domain AdaptationImage
🎯 What it does: A new Active Testing Time Adaptation (ATTA) framework is proposed, which can adapt in real-time to the continuously changing test distribution under the condition of obtaining only a limited number of labeled samples, avoiding catastrophic forgetting.
AdaMerging: Adaptive Model Merging for Multi-Task Learning
Enneng Yang (Northeastern University), Dacheng Tao (Nanyang Technological University)
ClassificationTransformerImage
🎯 What it does: This paper studies the model fusion problem in multi-task learning and proposes an adaptive model fusion method called AdaMerging.
Adapting Large Language Models via Reading Comprehension
Daixuan Cheng (Beijing Institute for General Artificial Intelligence), Furu Wei (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBiomedical DataFinance Related
🎯 What it does: Investigate the impact of continuous pre-training on domain-specific corpora for large language models and propose an adaptation method to convert raw corpora into reading comprehension texts.
Adapting to Distribution Shift by Visual Domain Prompt Generation
Zhixiang Chi (University of Toronto), Yang Wang (Concordia University)
Domain AdaptationMeta LearningTransformerPrompt EngineeringContrastive LearningImage
🎯 What it does: Proposes a Visual Domain Prompt Generator (VDPG) method for domain adaptation of large pre-trained models (such as CLIP) using a small number of unlabeled samples during testing;
Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts
Jian Xie (Fudan University), Yu Su (Ohio State University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The system evaluates the behavior of large language models in knowledge conflict scenarios and proposes a complete framework for constructing counter-memory.
Adaptive deep spiking neural network with global-local learning via balanced excitatory and inhibitory mechanism
Tingting Jiang (Dalian University of Technology), Gang Pan (Zhejiang University)
Spiking Neural NetworkImage
🎯 What it does: This paper proposes an adaptive training method for deep spiking neural networks (EIHL) that combines local (STDP) and global (STBP) learning, automatically adjusting network sparsity and switching learning modes through an excitation-inhibition mechanism.
Adaptive Federated Learning with Auto-Tuned Clients
Junhyung Lyle Kim (Rice University), Anastasios Kyrillidis (Rice University)
Federated LearningImageText
🎯 What it does: This paper proposes Δ-SGD, a distributed SGD algorithm that adaptively adjusts the learning rate for each client in a federated learning environment;
Adaptive Instrument Design for Indirect Experiments
Yash Chandak (Stanford University), Emma Brunskill (Stanford University)
Reinforcement Learning
🎯 What it does: This paper proposes an adaptive instrument design framework based on influence functions and multiple rejection sampling, which can significantly improve sample efficiency in indirect experiments and validate its effectiveness on various linear and nonlinear IV estimators.
Adaptive Rational Activations to Boost Deep Reinforcement Learning
Quentin Delfosse (TU Darmstadt), Kristian Kersting (TU Darmstadt)
Reinforcement LearningImage
🎯 What it does: Proposes and evaluates learnable rational activation functions to enhance neural plasticity in deep reinforcement learning.
Adaptive Regret for Bandits Made Possible: Two Queries Suffice
Zhou Lu (Princeton University), Elad Hazan (Google Deepmind)
OptimizationReinforcement LearningTabularBenchmark
🎯 What it does: Proposed a multi-armed bandit and convex optimization algorithm that achieves optimal strongly adaptive regret with only two queries.
Adaptive Regularization of Representation Rank as an Implicit Constraint of Bellman Equation
Qiang He (Ruhr University Bochum), Setareh Maghsudi (Ruhr University Bochum)
Reinforcement LearningSequential
🎯 What it does: This paper proposes BEER (Bellman Equation-based automatic rank Regularizer) to automatically adjust the rank of representation layers in deep reinforcement learning, preventing the model from becoming overly complex or underfitting.
Adaptive Retrieval and Scalable Indexing for k-NN Search with Cross-Encoders
Nishant Yadav (University of Massachusetts Amherst), Andrew McCallum (Google DeepMind)
RetrievalOptimizationTextBenchmark
🎯 What it does: This paper proposes an offline indexing method based on sparse matrix decomposition and an online adaptive retrieval method (AXN), achieving efficient approximation and retrieval for cross-encoders (CE) in k-NN search.
Adaptive Self-training Framework for Fine-grained Scene Graph Generation
Kibum Kim (Korea Advanced Institute of Science and Technology), Chanyoung Park (Korea Advanced Institute of Science and Technology)
Object DetectionGenerationGraph Neural NetworkSupervised Fine-TuningGraph
🎯 What it does: This paper proposes a self-training framework (ST-SGG) that generates pseudo-labels for unlabeled relational triples, utilizing large-scale unlabeled data to enhance the performance of scene graph generation models, particularly for fine-grained predicates.