ICML 2024 Papers — Page 11
International Conference on Machine Learning · 2610 papers
Graphon Mean Field Games with a Representative Player: Analysis and Learning Algorithm
Fuzhong Zhou (Columbia University), Xuan Di (Columbia University)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes a discrete-time graphon equilibrium game framework based on representative players, studying graphon equilibria in continuous state and action spaces, and provides theoretical proofs for existence, uniqueness, and approximate equilibria; it also designs the first completely oracle-free online learning algorithm and presents sample complexity; numerical experiments validate the performance of the algorithm.
GRATH: Gradual Self-Truthifying for Large Language Models
Weixin Chen (University of Illinois at Urbana-Champaign), Bo Li (University of Chicago)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: The GRATH method is proposed, which enhances the model's truthfulness under OOD queries by allowing the LLM to generate pairs of true and false answers itself and using DPO for iterative fine-tuning.
Grokking Group Multiplication with Cosets
Dashiell Stander (EleutherAI), Stella Biderman (EleutherAI)
🎯 What it does: The researchers conducted reverse engineering on a single-layer fully connected network, which implements multiplication on the full permutations of the symmetric groups S5 and S6, revealing that its internal computation is completed using coset circuits of the covariant subgroup.
GroupCover: A Secure, Efficient and Scalable Inference Framework for On-device Model Protection based on TEEs
Zheng Zhang (Beihang University), Ye Wu (Douyin Vision Co., Ltd.)
Safty and PrivacyComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: A secure inference framework called GROUPCOVER is proposed, which is based on Trusted Execution Environment (TEE) and GPU acceleration to protect the intellectual property of deep learning models deployed on edge devices.
Guarantees for Nonlinear Representation Learning: Non-identical Covariates, Dependent Data, Fewer Samples
Thomas TCK Zhang, Nikolai Matni (University of Pennsylvania)
Representation LearningSequential
🎯 What it does: Under the framework of multi-task representation learning, the statistical generalization guarantees in the case of non-identically distributed covariates and dependent samples are studied, and new task coverage and diversity metrics are proposed.
Guidance with Spherical Gaussian Constraint for Conditional Diffusion
Lingxiao Yang (ShanghaiTech University), Ye Shi (ShanghaiTech University)
RestorationGenerationSuper ResolutionDiffusion modelImage
🎯 What it does: This paper proposes a training-free conditional diffusion model plugin—Spherical Gaussian Constraint (DSG), which addresses the manifold deviation problem encountered in traditional methods during the sampling process by constraining the guiding steps on the surface of a high-dimensional Gaussian distribution, significantly improving sample quality and sampling speed.
Guiding LLMs The Right Way: Fast, Non-Invasive Constrained Generation
Luca Beurer-Kellner (ETH Zurich), Martin Vechev (ETH Zurich)
GenerationComputational EfficiencyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: The DOMINO algorithm is proposed, achieving minimally intrusive constraint generation for large language models.
HALC: Object Hallucination Reduction via Adaptive Focal-Contrast Decoding
Zhaorun Chen (University of Chicago), Jiawei Zhou (Toyota Technological Institute at Chicago)
Object DetectionTransformerVision Language ModelImageText
🎯 What it does: A new decoding algorithm HALC is proposed, which utilizes adaptive focal contrast localization and global visual matching bundle search to reduce object hallucination in large visual language models in real-time.
HAMLET: Graph Transformer Neural Operator for Partial Differential Equations
Andrey Bryutkin (Massachusetts Institute of Technology), Angelica I Aviles-Rivero
Graph Neural NetworkTransformerGraphBenchmarkPhysics Related
🎯 What it does: A neural operator framework called HAMLET based on graph Transformer is proposed for solving partial differential equations with arbitrary geometries and input formats.
Handling Heterogeneous Curvatures in Bandit LQR Control
Yu-Hu Yan (Nanjing University), Peng Zhao (Nanjing University)
OptimizationReinforcement Learning from Human FeedbackReinforcement Learning
🎯 What it does: This paper studies the Bandit LQR control problem with heterogeneous curvature, proposing a lossless with-history reduction and utilizing Newton decrement for adaptive tuning of heterogeneous curvature.
Hard Tasks First: Multi-Task Reinforcement Learning Through Task Scheduling
Myungsik Cho (Korea Advanced Institute of Science and Technology), Youngchul Sung (Korea Advanced Institute of Science and Technology)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: A multi-task reinforcement learning algorithm based on task scheduling (SMT) is proposed, which alleviates negative transfer by prioritizing the training of more difficult tasks.
HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal
Mantas Mazeika (University of Illinois Urbana-Champaign), Dan Hendrycks (Center for AI Safety)
OptimizationAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningTextMultimodalityBenchmark
🎯 What it does: This paper proposes HarmBench, a standardized evaluation framework for automated red teaming and robust rejection, and utilizes this framework for a large-scale comparison of 18 red teaming methods against 33 large language models (including defenses). It further introduces an efficient adversarial training method called R2D2.
HarmoDT: Harmony Multi-Task Decision Transformer for Offline Reinforcement Learning
Shengchao Hu (Shanghai Jiao Tong University), Dacheng Tao (Nanyang Technological University)
Meta LearningTransformerReinforcement LearningSequential
🎯 What it does: Proposes HarmoDT, a Harmony Multi-Task Decision Transformer for offline multi-task reinforcement learning;
Harmonic Self-Conditioned Flow Matching for joint Multi-Ligand Docking and Binding Site Design
Hannes Stark (Massachusetts Institute of Technology), Tommi Jaakkola (Massachusetts Institute of Technology)
Drug DiscoveryGraph Neural NetworkFlow-based ModelGraph
🎯 What it does: This paper proposes two models, HARMONICFLOW and FLOWSITE, which can jointly generate the 3D poses of ligands and the residue types of protein binding sites based solely on the protein backbone and the chemical graph of the ligand, achieving automatic design of small molecule binding sites.
Harmonizing Generalization and Personalization in Federated Prompt Learning
Tianyu Cui (ShanghaiTech University), Ye Shi (ShanghaiTech University)
Federated LearningVision Language ModelContrastive LearningImage
🎯 What it does: A prompt learning framework called FedPGP is proposed, which simultaneously considers personalization and generalization in federated learning.
Harmony in Diversity: Merging Neural Networks with Canonical Correlation Analysis
Stefan Horoi (Universite Montreal), Guy Wolf (Universite Montreal)
OptimizationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A model merging method based on regularized CCA, called CCA Merge, is proposed. It aligns features of multiple neural networks trained with different initializations, data splits, or training strategies, and averages the weights through linear transformation to obtain a single model.
HarmonyDream: Task Harmonization Inside World Models
Haoyu Ma (Tsinghua University), Mingsheng Long (Tsinghua University)
Robotic IntelligenceReinforcement LearningAuto EncoderWorld ModelImageVideo
🎯 What it does: A framework named HarmonyDream is proposed, which enhances the sample efficiency of model-based RL by dynamically balancing observation modeling and reward modeling in the world model.
Harnessing Hierarchical Label Distribution Variations in Test Agnostic Long-tail Recognition
Zhiyong Yang (University of Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
RecognitionMixture of ExpertsImage
🎯 What it does: This paper proposes a model based on Dirichlet Mixture Experts (DirMixE) to address the long-tail recognition problem where the label distribution is unknown and extremely imbalanced during testing.
Harnessing Neural Unit Dynamics for Effective and Scalable Class-Incremental Learning
Depeng Li (Huazhong University of Science and Technology), Zhigang Zeng (Huazhong University of Science and Technology)
ClassificationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: The AutoActivator model is proposed, which achieves task-free incremental learning through dynamic expansion and reactivation of neural units, avoiding catastrophic forgetting.
Harnessing the Power of Neural Operators with Automatically Encoded Conservation Laws
Ning Liu (Global Engineering and Materials, Inc.), Yue Yu (Lehigh University)
Time SeriesPhysics Related
🎯 What it does: A conservative law encoding neural operator (clawNO) is proposed, which ensures the conservation of quality (and other) by embedding divergence-free outputs in the network structure;
HelmFluid: Learning Helmholtz Dynamics for Interpretable Fluid Prediction
Lanxiang Xing (Tsinghua University), Mingsheng Long (Tsinghua University)
Explainability and InterpretabilityComputational EfficiencyConvolutional Neural NetworkOptical FlowTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A fluid prediction framework called HelmFluid is proposed based on the Helmholtz theorem, which models fluid dynamics by learning potential and stream functions, and utilizes a multi-scale multi-head integration architecture for future fluid field predictions.
Helpful or Harmful Data? Fine-tuning-free Shapley Attribution for Explaining Language Model Predictions
Jingtan Wang (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
ClassificationExplainability and InterpretabilityLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A Shapley value-based instance attribution method is proposed, introducing the concept of β-robustness to evaluate the sign stability of attribution scores under different data resampling.
HexGen: Generative Inference of Large Language Model over Heterogeneous Environment
YOUHE JIANG, Binhang Yuan (Hong Kong University of Science and Technology)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper proposes HEXGEN, a distributed large language model inference engine that supports heterogeneous multi-GPU environments, capable of balancing tensor model parallelism and pipeline parallelism.
HGAP: Boosting Permutation Invariant and Permutation Equivariant in Multi-Agent Reinforcement Learning via Graph Attention Network
Bor-Jiun Lin, Chun-Yi Lee (National Tsing Hua University)
Recurrent Neural NetworkGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A multi-agent reinforcement learning framework called HGAP based on graph attention networks is proposed to address the PI/PE and POMDP problems.
HGCN2SP: Hierarchical Graph Convolutional Network for Two-Stage Stochastic Programming
Yang Wu (Institute of Automation), Jian Cheng (University of Chinese Academy of Sciences)
OptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: In response to the scenario selection and ranking problem in two-stage stochastic programming (2SP), the HGCN2SP model is proposed, which encodes scenarios using a hierarchical graph structure and trains within a reinforcement learning framework.
Hidden Traveling Waves bind Working Memory Variables in Recurrent Neural Networks
Arjun Karuvally (University of Massachusetts Amherst), Hava T Siegelmann
Recurrent Neural NetworkTransformerSequential
🎯 What it does: This paper proposes a wave-based working memory model (TWM) and connects its theory with recurrent neural networks (RNN) and self-attention Transformers, demonstrating that RNNs naturally form hidden propagating waves when learning linear history-dependent systems (HDS), thereby improving gradient propagation.
Hierarchical Integral Probability Metrics: A distance on random probability measures with low sample complexity
Marta Catalano (Luiss University), Hugo Lavenant (Bocconi University)
🎯 What it does: A new distance (Hierarchical Integral Probability Metric, HIPM) is proposed to measure the similarity of random probability distributions.
Hierarchical Neural Operator Transformer with Learnable Frequency-aware Loss Prior for Arbitrary-scale Super-resolution
Xihaier Luo (Brookhaven National Laboratory), Byung-Jun Yoon (Texas A&M University)
RestorationSuper ResolutionTransformerImage
🎯 What it does: A method for arbitrary scale super-resolution based on Hierarchical Neural Operator Transformer (HiNOTE) is proposed, utilizing Fourier cascade upsampling and Galerkin self-attention to achieve high-quality detail recovery on scientific data.
Hierarchical Novelty Detection via Fine-Grained Evidence Allocation
Spandan Pyakurel (Rochester Institute of Technology), Qi Yu (Rochester Institute of Technology)
Anomaly DetectionImage
🎯 What it does: A novel hierarchical novelty detection method based on evidence allocation, E-HND, is proposed, which can simultaneously detect novel samples and locate their nearest parent class within an existing hierarchical structure.
Hierarchical State Space Models for Continuous Sequence-to-Sequence Modeling
Raunaq Bhirangi (Carnegie Mellon University), Lerrel Pinto (New York University)
Recurrent Neural NetworkTransformerTime SeriesSequentialBenchmark
🎯 What it does: The paper proposes a benchmark for continuous sequence-to-sequence prediction called CSP-Bench and designs a hierarchical state space model (HiSS) for predicting high-frequency sensor data.
Hieros: Hierarchical Imagination on Structured State Space Sequence World Models
Paul Mattes (Hasso Plattner Institute), Ralf Herbrich (Hasso Plattner Institute)
Reinforcement LearningAuto EncoderWorld ModelSequentialBenchmark
🎯 What it does: HIEROS is proposed, a multi-level hierarchical reinforcement learning framework that combines a hierarchical world model with a subgoal autoencoder, utilizing the S5 structured state space layer to achieve efficient temporal abstraction simulation.
High-Dimensional Bayesian Optimization via Semi-Supervised Learning with Optimized Unlabeled Data Sampling
Yuxuan Yin (University of California), Peng Li (University of California)
OptimizationDrug DiscoveryAuto EncoderTabular
🎯 What it does: A semi-supervised learning framework called Teacher-Student Bayesian Optimization (TSBO) is proposed, which predicts pseudo-labels for unlabeled points through a teacher model. The student model learns these labels and provides feedback to the teacher, significantly reducing the number of expensive label queries in high-dimensional Bayesian optimization.
High-Dimensional Geometric Streaming for Nearly Low Rank Data
Hossein Esfandiari (Google Research), Peilin Zhong (Google Research)
OptimizationComputational EfficiencyImage
🎯 What it does: A single-pass linear time and linear space streaming algorithm is proposed to construct a strong coreset, enabling efficient solutions to the p-subspace approximation problem in high-dimensional data (especially for p=∞).
High-Dimensional Kernel Methods under Covariate Shift: Data-Dependent Implicit Regularization
Yihang Chen (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
🎯 What it does: This study investigates the bias-variance trade-off of high-dimensional kernel ridge regression under covariate shift and provides an analysis of importance-weighted adaptive implicit regularization.
High-dimensional Linear Bandits with Knapsacks
Wanteng Ma (Hong Kong University of Science and Technology), Jiashuo Jiang (Hong Kong University of Science and Technology)
OptimizationTabular
🎯 What it does: This paper proposes an online hard-threshold-based sparse estimator within a primal-dual framework to address the high-dimensional context bandit with constraints (CBwK) problem.
High-Order Contrastive Learning with Fine-grained Comparative Levels for Sparse Ordinal Tensor Completion
Yu Dai (East China Normal University), Kai Zhang (East China Normal University)
Recommendation SystemGraph Neural NetworkContrastive LearningTime SeriesSequential
🎯 What it does: A high-order contrastive learning framework HOCTC is proposed for the completion of sparse ordinal tensors.
High-Performance Temporal Reversible Spiking Neural Networks with $\mathcal{O}(L)$ Training Memory and $\mathcal{O}(1)$ Inference Cost
JiaKui Hu, Guoqi Li (Institute of Automation, Chinese Academy of Sciences)
ClassificationComputational EfficiencySpiking Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A time-reversible spiking neural network (T-RevSNN) is proposed, which achieves low training memory and low inference energy consumption by shutting down most temporal dynamics and retaining temporal information only in key layers.
High-Probability Bound for Non-Smooth Non-Convex Stochastic Optimization with Heavy Tails
Langqi Liu (Nanjing University), Lijun Zhang (Nanjing University)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a high-probability convergence algorithm for non-smooth non-convex stochastic optimization problems with heavy-tailed noise.
High-Probability Convergence for Composite and Distributed Stochastic Minimization and Variational Inequalities with Heavy-Tailed Noise
Eduard Gorbunov (Mohamed bin Zayed University of Artificial Intelligence), Peter Richtárik (King Abdullah University of Science and Technology)
Optimization
🎯 What it does: This paper proposes a gradient differential clipping method for composite and distributed stochastic optimization and variational inequality problems under heavy-tailed noise, and provides a high-probability convergence analysis.
Highway Value Iteration Networks
Yuhui Wang (King Abdullah University of Science and Technology), Jürgen Schmidhuber
Convolutional Neural NetworkReinforcement LearningSequential
🎯 What it does: This paper proposes the Highway Value Iteration Network (Highway VIN), which integrates highway value iteration into the planning module of VIN and adds an aggregate gate, exploration module, and filter gate to enhance long-term planning capabilities.
Homomorphism Counts for Graph Neural Networks: All About That Basis
Emily Jin (University of Oxford), Matthias Lanzinger (Institute for Logic and Computation)
Graph Neural NetworkGraph
🎯 What it does: This paper proposes a new feature injection framework that injects homomorphic counts from graph homomorphism bases into Graph Neural Networks (GNNs), significantly enhancing the model's expressiveness while leveraging the advantages of both subgraph counting and homomorphic counting.
How Deep Do We Need: Accelerating Training and Inference of Neural ODEs via Control Perspective
Keyan Miao (University of Oxford), Konstantinos Gatsis (University of Southampton)
OptimizationComputational EfficiencyImageOrdinary Differential Equation
🎯 What it does: This study investigates how to accelerate the training and inference of Neural ODEs from the perspective of control theory through minimum time control and Lyapunov methods.
How Deep Networks Learn Sparse and Hierarchical Data: the Sparse Random Hierarchy Model
Umberto Maria Tomasini (Institute of Physics EPFL), Matthieu Wyart (Institute of Physics EPFL)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: A Sparse Random Hierarchical Model (SRHM) is proposed, demonstrating that the sparse property makes classification tasks insensitive to discrete translations (similar to differential transformations), and learns corresponding hierarchical invariant representations through deep networks.
How do Large Language Models Navigate Conflicts between Honesty and Helpfulness?
Ryan Liu (Princeton University), Thomas L. Griffiths (Princeton University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextChain-of-Thought
🎯 What it does: Using a psychological experimental paradigm (signal gambling game) to assess the trade-off between honesty and helpfulness in large language models, and exploring the impact of RLHF, chain reasoning, and prompting strategies on this trade-off through multiple experiments.
How Do Nonlinear Transformers Learn and Generalize in In-Context Learning?
Hongkang Li (Rensselaer Polytechnic Institute), Pin-Yu Chen (IBM)
TransformerLarge Language ModelText
🎯 What it does: This paper conducts a theoretical analysis to study the training dynamics of nonlinear Transformers under self-attention and ReLU MLP structures, as well as their generalization performance in in-context learning (ICL);
How do Transformers Perform In-Context Autoregressive Learning ?
Michael Eli Sander, Gabriel Peyré (Ecole Normale Superieure)
TransformerTextSequential
🎯 What it does: This paper studies the context learning mechanism implemented by linear Transformers in the autoregressive (AR) process;
How Does Goal Relabeling Improve Sample Efficiency?
Sirui Zheng (Northwestern University), Zhaoran Wang (Princeton University)
Reinforcement Learning
🎯 What it does: This paper proposes a theoretical and algorithmic framework that utilizes goal relabeling to significantly enhance the sample efficiency of sparse reward reinforcement learning.
How Far Can Fairness Constraints Help Recover From Biased Data?
mohit sharma, Amit Deshpande
ClassificationOptimization
🎯 What it does: The study investigates whether fairness constraints (mainly equal opportunity) can recover the optimal fair classifier under the original data distribution in the presence of systematic bias in the data distribution, and provides recovery conditions from extreme bias to arbitrary distributions, arbitrary hypothesis classes, Massart noise, rejection options, and time-varying scenarios.
How Flawed Is ECE? An Analysis via Logit Smoothing
Muthu Chidambaram (Duke University), Semon Rezchikov (Princeton University)
ClassificationImage
🎯 What it does: This paper systematically analyzes the discontinuity of Expected Calibration Error (ECE), provides a complete characterization of the breakpoints under any probability measure, and based on this, proposes a continuous Logit-Smoothed ECE (LS-ECE). It also offers a consistent estimation method and experimentally verifies the high consistency of LS-ECE with traditional ECE in image classification tasks.
How Free is Parameter-Free Stochastic Optimization?
Amit Attia (Tel Aviv University), Tomer Koren (Google Research)
OptimizationHyperparameter Search
🎯 What it does: The study investigates the problem of parameter-free stochastic optimization, exploring the conditions under which completely parameter-free methods exist, which have convergence rates comparable to the best tuning methods without requiring significant knowledge of the true problem parameters.
How Graph Neural Networks Learn: Lessons from Training Dynamics
Chenxiao Yang (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Graph Neural NetworkGraph
🎯 What it does: This study investigates the training dynamics of graph neural networks (GNNs) during the gradient descent process, revealing the phenomenon of 'kernel-graph alignment' in function space aligned with graph structure, and proposes a non-parametric residual propagation algorithm based on sparse adjacency matrices.
How Interpretable Are Interpretable Graph Neural Networks?
Yongqiang Chen (Chinese University of Hong Kong), James Cheng (Chinese University of Hong Kong)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: This paper proposes a theoretical framework to analyze the expressive power of explainable graph neural networks (XGNNs) and finds that existing attention-based XGNNs exhibit significant errors when approximating subgraph multilinear extensions (SubMT), leading to compromised interpretability and out-of-distribution (OOD) generalization. Based on this, the authors design a new XGNN architecture called Graph Multilinear Net (GMT), which more accurately approximates SubMT through linearization or random subgraph sampling, and further retrains the classifier using a frozen subgraph extractor after training to enhance performance.
How Language Model Hallucinations Can Snowball
Muru Zhang (University of Washington), Noah A. Smith (University of Washington)
GenerationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: This study investigates the phenomenon of 'avalanche hallucination' in language models, where they generate false positives and self-identify them when answering questions.
How Learning by Reconstruction Produces Uninformative Features For Perception
Randall Balestriero (Brown University), Yann LeCun (New York University)
ClassificationRepresentation LearningConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: This paper discusses the significant mismatch problem between the features learned through reconstruction (especially autoencoders) for representation learning and downstream perceptual tasks (such as classification), and investigates how to improve this alignment through noise design (such as occlusion).
How Private are DP-SGD Implementations?
Lynn Chua (Google Research), Chiyuan Zhang
Safty and Privacy
🎯 What it does: This paper analyzes the privacy loss curves of DP-SGD under different batch sampling methods (deterministic, Poisson, and randomized) and reveals significant privacy gaps.
How Smooth Is Attention?
Valérie Castin (École Normale Supérieure PSL), Gabriel Peyré (CNRS)
TransformerLarge Language ModelText
🎯 What it does: This paper studies the Lipschitz constant of the self-attention mechanism, analyzes the impact of sequence length and layer normalization on the local Lipschitz constant of self-attention, and proposes new theoretical bounds.
How Spurious Features are Memorized: Precise Analysis for Random and NTK Features
Simone Bombari (Institute of Science and Technology), Marco Mondelli
Gaussian SplattingImage
🎯 What it does: This study investigates the memory mechanism of deep learning models regarding irrelevant noisy features in the training set and provides an accurate theoretical description.
How to Escape Sharp Minima with Random Perturbations
Kwangjun Ahn (Massachusetts Institute of Technology), Suvrit Sra (Technische Universität München)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A gradient-based algorithm is proposed that uses random perturbations to approach and escape sharp local minima, seeking flat minima.
How to Explore with Belief: State Entropy Maximization in POMDPs
Riccardo Zamboni (Politecnico di Milano), Mirco Mutti (Technion - Israel Institute of Technology)
OptimizationRobotic IntelligenceReinforcement LearningAgentic AI
🎯 What it does: This paper presents research on maximizing the entropy of the true state distribution (State Entropy Maximization, SEM) within the framework of partially observable Markov decision processes (POMDP), aiming to enable agents to achieve high-entropy coverage of the true environmental state even when only partial information is observable.
How to Leverage Diverse Demonstrations in Offline Imitation Learning
Sheng Yue (Tsinghua University), Yaoxue Zhang (Tsinghua University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper proposes an offline imitation learning method called ILID, which utilizes diverse behaviors from non-expert demonstrations to select beneficial actions from noisy data and learn a policy.
How to Make the Gradients Small Privately: Improved Rates for Differentially Private Non-Convex Optimization
Andrew Lowy (University of Wisconsin Madison), Stephen Wright
OptimizationSafty and Privacy
🎯 What it does: A simple and flexible framework is proposed for designing differential privacy algorithms to find approximate stationary points of non-convex loss functions.
How to Trace Latent Generative Model Generated Images without Artificial Watermark?
Zhenting Wang (Rutgers University), Shiqing Ma (University of Massachusetts at Amherst)
GenerationData SynthesisOptimizationDiffusion modelAuto EncoderImage
🎯 What it does: This paper proposes LATENTTRACER, which utilizes a gradient reverse method initialized by an encoder to trace the source of images generated by latent generative models without adding artificial watermarks.
How Transformers Learn Causal Structure with Gradient Descent
Eshaan Nichani (Princeton University), Jason D. Lee (Princeton University)
TransformerSequential
🎯 What it does: This paper proposes a new autoregressive two-layer Transformer training task—Random Sequences and Causal Structures (Task 2.4), and demonstrates that gradient descent can learn the underlying causal graph in the first layer of attention, thereby achieving in-context estimation of transition probabilities in the sequence.
How Uniform Random Weights Induce Non-uniform Bias: Typical Interpolating Neural Networks Generalize with Narrow Teachers
Gon Buzaglo (Technion Institute of Technology), Daniel Soudry (Toyota Technological Institute at Chicago)
Image
🎯 What it does: This paper studies the generalization performance of over-parameterized neural networks obtained through 'guess and check' (G&C) sampling of random weights under zero loss.
How Universal Polynomial Bases Enhance Spectral Graph Neural Networks: Heterophily, Over-smoothing, and Over-squashing
Keke Huang (National University of Singapore), Pietro Lio
Graph Neural NetworkGraph
🎯 What it does: A unified polynomial basis (UniBasis) and a graph convolutional network based on it, called UniFilter, are proposed to handle different heterogeneous graphs and suppress over-smoothing and over-compression.
How Well Can LLMs Negotiate? NegotiationArena Platform and Analysis
Federico Bianchi (Stanford University), James Zou (Stanford University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper develops the open-source platform NEGOTIATIONARENA to evaluate the behavior of large language models (LLMs) in three multi-round negotiation scenarios: resource allocation, the Ultimatum game, and buy-sell transactions. It conducts systematic experiments comparing models such as GPT-4, GPT-3.5, and Claude-2/2.1.
Human Alignment of Large Language Models through Online Preference Optimisation
Daniele Calandriello (Google DeepMind), Bilal Piot (Google DeepMind)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningContrastive LearningText
🎯 What it does: This paper studies how to improve the alignment effect of large language models through online preference optimization, proposing two new algorithms: Online IPO and IPO-MD, and proving that Online IPO is equivalent to Nash-MD-PG at the expected update level.
Human vs. Generative AI in Content Creation Competition: Symbiosis or Conflict?
Fan Yao (University of Virginia), Haifeng Xu (University of Chicago)
🎯 What it does: A generalized model based on the Tullock competition is constructed to analyze the competition and collaboration between human content creators and generative AI (GenAI) in the content creation market. It discusses two competitive scenarios (exclusive and inclusive) and provides the existence, uniqueness, and micro/macro characteristics of pure Nash equilibrium (PNE).
Human-like Category Learning by Injecting Ecological Priors from Large Language Models into Neural Networks
Akshay Kumar Jagadish, Marcel Binz (Max Planck Institute for Biological Cybernetics)
ClassificationMeta LearningTransformerLarge Language ModelSupervised Fine-TuningTabular
🎯 What it does: Utilizing large language models (LLM) to generate category learning tasks that conform to the true ecological distribution, and then training the model through meta-learning (ERMI) to achieve human-like performance in learning difficulty, strategy transfer, and generalization, achieving state-of-the-art performance on the OpenML-CC18 classification benchmark.
HumanTOMATO: Text-aligned Whole-body Motion Generation
Shunlin Lu (Tsinghua University), Heung-Yeung Shum
GenerationTransformerContrastive LearningVideoText
🎯 What it does: We propose HumanTOMATO, a text-based framework for generating full-body motion that can produce high-quality, semantically consistent 3D motion sequences containing hand, facial, and body movements based on natural language descriptions.
Hybrid Inverse Reinforcement Learning
Juntao Ren (Cornell University), Sanjiban Choudhury (Cornell University)
Reinforcement Learning
🎯 What it does: A novel inverse reinforcement learning framework is proposed—Hybrid Inverse Reinforcement Learning (Hybrid IRL), which significantly reduces exploration and interaction requirements by simultaneously utilizing online data and expert demonstrations within the policy search.
Hybrid Neural Representations for Spherical Data
Hyomin Kim (Pohang University of Science and Technology), Sungsoo Ahn (Pohang University of Science and Technology)
Pose EstimationSuper ResolutionCompressionTime Series
🎯 What it does: This study investigates a hybrid neural representation method suitable for spherical data, combining feature grids with MLP to generate pose features in spherical coordinates, achieving regression, super-resolution, temporal interpolation, and compression of climate and cosmic microwave background data.
Hybrid Reinforcement Learning from Offline Observation Alone
Yuda Song (Carnegie Mellon University), Aarti Singh (Carnegie Mellon University)
Reinforcement Learning
🎯 What it does: This paper studies a hybrid reinforcement learning framework (HYRLO) that combines offline data containing only state information with online interaction, proposes the FOOBAR algorithm, and provides theoretical sample complexity guarantees along with experimental validation.
Hybrid$^2$ Neural ODE Causal Modeling and an Application to Glycemic Response
Bob Junyi Zou (Stanford University), Emily Fox
Biomedical DataOrdinary Differential Equation
🎯 What it does: A hybrid model combining ODE mechanisms and neural networks is proposed, optimizing both prediction and causal effects through a mixed loss function.
Hyperbolic Active Learning for Semantic Segmentation under Domain Shift
Luca Franco (ITALAI S.R.L.), Fabio Galasso (Sapienza University of Rome)
SegmentationDomain AdaptationImage
🎯 What it does: In the active learning for domain adaptation semantic segmentation, the HALO method is proposed, which utilizes the radius and predictive entropy in hyperbolic space to assess the uncertainty of pixels and select pixels that need labeling.
Hyperbolic Geometric Latent Diffusion Model for Graph Generation
Xingcheng Fu (Guangxi Normal University), Xianxian LI
GenerationGraph Neural NetworkDiffusion modelGraph
🎯 What it does: A new implicit diffusion model, HypDiff, is proposed for generating graph structures in hyperbolic space.
Hyperbolic Optimizer as a Dynamical System
Nico Alvarado (Pontificia Universidad Catolica de Chile), Hans Lobel (Pontificia Universidad Catolica de Chile)
OptimizationOrdinary Differential Equation
🎯 What it does: A hyperbolic curvature optimizer based on the Poincaré ball model is proposed, treating it as a continuous dynamical system of nonlinear differential equations, and its convergence properties are analyzed using Lyapunov stability theory.
HyperFields: Towards Zero-Shot Generation of NeRFs from Text
Sudarshan Babu (Toyota Technological Institute at Chicago), Rana Hanocka (University of Chicago)
GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelScore-based ModelNeural Radiance FieldImageText
🎯 What it does: HyperFields maps text descriptions to NeRF weights through a dynamic hypernetwork in a single forward pass, achieving zero-shot 3D scene generation.
Hypergraph-enhanced Dual Semi-supervised Graph Classification
Wei Ju (Peking University), Ming Zhang (Peking University)
ClassificationGraph Neural NetworkContrastive LearningGraph
🎯 What it does: A HEAL framework is proposed, utilizing both hypergraphs and line graphs for semi-supervised graph classification.
I/O Complexity of Attention, or How Optimal is FlashAttention?
Barna Saha (University of California San Diego), Christopher Ye (University of California San Diego)
Transformer
🎯 What it does: Analyze the I/O complexity of the attention mechanism under two levels of memory hierarchy, and provide optimal algorithms and lower bounds.
IBD-PSC: Input-level Backdoor Detection via Parameter-oriented Scaling Consistency
Linshan Hou (Harbin Institute of Technology), Yiming Li (Nanyang Technological University)
Anomaly DetectionImage
🎯 What it does: A method for detecting input-level backdoors based on parameter amplification consistency, IBD-PSC, has been designed to identify induced samples in a white-box scenario by amplifying batch normalization (BN) parameters and evaluating confidence consistency.
Identifiability Matters: Revealing the Hidden Recoverable Condition in Unbiased Learning to Rank
Mouxiang Chen (Zhejiang University), Jianling Sun (Zhejiang University)
Recommendation SystemOptimizationGraph Neural NetworkGraph
🎯 What it does: This study investigates the issue of identifiability of relevance in unbiased click learning, proposing identifiability conditions based on connectivity tests, and presents two model-free data processing methods (node intervention and node merging) to recover relevance.
Identification and Estimation for Nonignorable Missing Data: A Data Fusion Approach
Zixiao Wang (Johns Hopkins University), Ilya Shpitser (Johns Hopkins University)
TabularBiomedical Data
🎯 What it does: This paper proposes a data fusion framework that utilizes the MAR characteristics of auxiliary datasets in conjunction with the MNAR characteristics of the main dataset to achieve the identifiability and estimation of missing data parameters.
IIANet: An Intra- and Inter-Modality Attention Network for Audio-Visual Speech Separation
Kai Li (Tsinghua University), Xiaolin Hu (Tsinghua University)
RecognitionData SynthesisConvolutional Neural NetworkVideoMultimodalityAudio
🎯 What it does: This paper proposes a hierarchical attention network for audio-visual speech separation, IIANet, which achieves higher quality separation by cross-utilizing intra-modal and inter-modal attention in the audio and visual pathways.
ILILT: Implicit Learning of Inverse Lithography Technologies
Haoyu Yang (NVIDIA), Haoxing Ren (NVIDIA)
GenerationOptimizationImageBenchmarkOrdinary Differential Equation
🎯 What it does: This paper proposes an ILILT framework that models inverse lithography technology (ILT) as a fixed-point iteration using implicit layer learning, achieving direct generation of high-quality masks without the need for traditional numerical solvers' iterations.
IM-3D: Iterative Multiview Diffusion and Reconstruction for High-Quality 3D Generation
Luke Melas-Kyriazi (Meta), Filippos Kokkinos (Meta)
GenerationData SynthesisOptimizationDiffusion modelGaussian SplattingImageVideoText
🎯 What it does: An end-to-end text/image to 3D model generation framework IM-3D is proposed, which first uses a fine-tuned text-to-video generator to produce multi-view images, and then directly optimizes these views using Gaussian Splatting and image-level losses to obtain high-quality 3D models, further enhancing consistency through an iterative generation-reconstruction loop.
IM-Unpack: Training and Inference with Arbitrarily Low Precision Integers
Zhanpeng Zeng (University of Wisconsin-Madison), Vikas Singh (University of Wisconsin-Madison)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper verifies that the Transformer model can directly use integer GEMM in both training and inference, and proposes the IM-Unpack algorithm to decompose high-bit integers into low-bit integers, thereby achieving full low-precision integer computation.
Image Clustering with External Guidance
Yunfan Li (Sichuan University), Xi Peng (Lenovo Research)
Knowledge DistillationContrastive LearningImage
🎯 What it does: Proposes an image clustering method TAC based on external textual knowledge.
Image Fusion via Vision-Language Model
Zixiang Zhao (Xi'an Jiaotong University), Luc Van Gool (KULeuven)
TransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A visual language model is used to guide image fusion through text descriptions, proposing the FILM framework.
Image Hijacks: Adversarial Images can Control Generative Models at Runtime
Luke Bailey (Harvard University), Scott Emmons (University of California, Berkeley)
GenerationAdversarial AttackTransformerPrompt EngineeringVision Language ModelImageTextMultimodality
🎯 What it does: The concept of 'Image Hijacks' is proposed, and based on two methods: Behaviour Matching and Prompt Matching, it can automatically generate slightly perturbed images to control the output of multimodal visual language models during inference.
Image Restoration Through Generalized Ornstein-Uhlenbeck Bridge
Conghan Yue (Sun Yat-sen University), Dongyu Zhang (Sun Yat-sen University)
RestorationSuper ResolutionDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A framework for image restoration based on the Generalized Ornstein-Uhlenbeck Bridge (GOUB) is proposed, utilizing Doob's h-transformation to eliminate endpoint variance, achieving a point-to-point mapping from low-quality images to high-quality images, and providing a Mean-ODE variant to enhance pixel details and structural perception.
Imitation Learning from Purified Demonstrations
Yunke Wang (Wuhan University), Chang Xu (University of Sydney)
Robotic IntelligenceReinforcement Learning from Human FeedbackDiffusion modelGenerative Adversarial NetworkSequential
🎯 What it does: To address the presence of noisy (suboptimal) example data, the authors propose first using a diffusion model to denoise the examples (forward diffusion + reverse reconstruction) to obtain 'pure' examples, which are then used in traditional imitation learning frameworks such as behavior cloning or GAIL.
Imitation Learning in Discounted Linear MDPs without exploration assumptions
Luca Viano (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: This paper proposes a new imitation learning algorithm ILARL, specifically designed for infinite-horizon linear MDPs, which can learn a policy close to that of an expert without the need for exploration assumptions. The algorithm transforms imitation learning into an adversarial loss problem in online learning for full-information linear MDPs and derives the final policy through an online-to-batch conversion.
Impact of Decentralized Learning on Player Utilities in Stackelberg Games
Kate Donahue (Cornell University), Aleksandrs Slivkins (Microsoft Research)
Recommendation SystemOptimizationReinforcement Learning
🎯 What it does: This study investigates the long-term interaction between two learners (such as recommendation systems and users, chatbots and humans) in a decentralized, error-aligned Stackelberg game, and its impact on their respective utilities, providing corresponding learning algorithms.
Implicit Bias of AdamW: $\ell_\infty$-Norm Constrained Optimization
Shuo Xie (Toyota Technological Institute at Chicago), Zhiyuan Li (Toyota Technological Institute at Chicago)
OptimizationText
🎯 What it does: In deterministic optimization, this paper proves that AdamW is guaranteed to be a KKT point of the objective function satisfying the ℓ∞ constraint (∥x∥∞ ≤ 1/λ) at convergence, thereby revealing the implicit regularization effect of AdamW.
Implicit Bias of Policy Gradient in Linear Quadratic Control: Extrapolation to Unseen Initial States
Noam Razin (Tel Aviv University), Nadav Cohen (Tel Aviv University)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper theoretically studies how the implicit bias of policy gradient learning in the underdetermined linear quadratic regulation (LQR) problem affects the controller's extrapolation ability for unseen initial states during training, and experimentally verifies the theoretical conclusions, further extending to nonlinear systems and neural network controllers.
Implicit Compressibility of Overparametrized Neural Networks Trained with Heavy-Tailed SGD
Yijun Wan (Paris Research Center, Huawei Technologies France), Umut Simsekli (Inria)
CompressionOptimizationConvolutional Neural NetworkImageTime SeriesElectrocardiogramStochastic Differential Equation
🎯 What it does: This paper proposes injecting heavy-tailed noise with infinite variance into SGD and proves that the weight matrix obtained after training in a hyperparameterized single hidden layer network can achieve compressibility with high probability at any compression ratio.
Implicit meta-learning may lead language models to trust more reliable sources
Dmitrii Krasheninnikov (University of Cambridge), David Krueger
Meta LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The study investigates whether language models implicitly learn to recognize and trust more reliable sources during the fine-tuning process, and subsequently internalize information from reliable sources more thoroughly in later training.
Implicit Regularization in Feedback Alignment Learning Mechanisms for Neural Networks
Zachary Robertson (Stanford University), Sanmi Koyejo
ClassificationOptimizationConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: This study proposes a theoretical framework and implicit regularization mechanism for the Feedback Alignment learning mechanism, proving its properties in maintaining gradient alignment and convergence, and conducting experimental validation on multi-class tasks.
Implicit Representations for Constrained Image Segmentation
Jan Philipp Schneider (University of Siegen), Michael Moeller (University of Siegen)
SegmentationConvolutional Neural NetworkImageVideo
🎯 What it does: Utilizing implicit function representation for image segmentation, explicitly enforcing geometric constraints (convexity, star shape, symmetry, periodicity, path connectivity, etc.) in the implicit domain through a specific network structure, and embedding it into variational and deep learning segmentation frameworks.