ICML 2024 Papers with AI Summaries
International Conference on Machine Learning · 2610 papers
→ ICML 2024 papers with code (550)
Each paper below shows an AI-generated one-line summary. Get the full 6-part summary (innovation, method, data, results, limitations) and search all 2610 ICML 2024 papers by keyword, author or institution —
free trial on arXivSub.
${\rm E}(3)$-Equivariant Actor-Critic Methods for Cooperative Multi-Agent Reinforcement Learning
Dingyang Chen (University of South Carolina), Qi Zhang (University of South Carolina)
Graph Neural NetworkReinforcement LearningPoint Cloud
🎯 What it does: A multi-agent reinforcement learning framework utilizing E(3) Euclidean symmetry is proposed, designing an actor-critic architecture based on E(3) equivariant message passing networks (SEGNN);
$\bf{\Phi}_\textrm{Flow}$: Differentiable Simulations for PyTorch, TensorFlow and Jax
Philipp Holl (Technical University of Munich), Nils Thuerey (Technical University of Munich)
OptimizationComputational EfficiencyPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper presents PhiFlow, a differentiable simulation toolkit that spans PyTorch, TensorFlow, Jax, and NumPy, providing a complete set of functionalities from differential operators and boundary handling to full fluid solvers, greatly simplifying the writing of scientific simulation code such as PDEs/ODEs.
$\mathtt{VITS}$ : Variational Inference Thompson Sampling for contextual bandits
Pierre Clavier (Ecole Polytechnique), Alain Oliviero Durmus
Recommendation SystemReinforcement LearningTabular
🎯 What it does: A Thompson Sampling algorithm based on Gaussian variational inference (VITS) is proposed, addressing the difficulties of posterior sampling and high computational cost in traditional TS, suitable for both linear and nonlinear (e.g., quadratic, logistic) scenarios;
$\texttt{MoE-RBench}$: Towards Building Reliable Language Models with Sparse Mixture-of-Experts
Guanjie Chen (Shanghai Artificial Intelligence Laboratory), Yu Cheng (The Chinese University of Hong Kong)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelMixture of ExpertsTextBenchmark
🎯 What it does: This paper proposes MoE-RBench, a comprehensive benchmark for evaluating the multi-dimensional reliability metrics of sparse expert models (MoE) in terms of safety, hallucination, adversarial robustness, and out-of-distribution (OOD) robustness.
$f$-Divergence Based Classification: Beyond the Use of Cross-Entropy
Nicola Novello (University of Klagenfurt), Andrea M Tonello
ClassificationOptimizationImage
🎯 What it does: A posterior probability learning framework is derived from the variational representation of f-divergence, and a new Shifted Log divergence is proposed for classification.
$H$-Consistency Guarantees for Regression
Anqi Mao (New York University), Yutao Zhong (Google Research)
Tabular
🎯 What it does: This paper proposes and studies the H-consistency boundary for regression problems, providing consistency theorems for various alternative loss functions (Huber, p-norm, epsilon-insensitive, etc.) for squared loss.
$S^2$IP-LLM: Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting
Zijie Pan (University of Connecticut), Dongjin Song (University of Connecticut)
TransformerLarge Language ModelPrompt EngineeringTime Series
🎯 What it does: The S2IP-LLM framework is proposed, which aligns the semantic space of the pre-trained LLM with the temporal embedding space and utilizes prompt learning to achieve temporal prediction.
3D Geometric Shape Assembly via Efficient Point Cloud Matching
Nahyuk Lee (POSTECH), Minsu Cho (POSTECH)
OptimizationRobotic IntelligencePoint Cloud
🎯 What it does: Proposes the Proxy Match Transform (PMT) and the PMTR framework to achieve efficient high-order matching of point cloud features, thereby completing geometric shape assembly tasks.
3D-VLA: A 3D Vision-Language-Action Generative World Model
Haoyu Zhen (Shanghai Jiao Tong University), Chuang Gan (Massachusetts Institute of Technology)
GenerationRobotic IntelligenceTransformerLarge Language ModelVision-Language-Action ModelDiffusion modelWorld ModelMultimodalityPoint Cloud
🎯 What it does: Proposes 3D-VLA, a generative world model that integrates 3D perception, reasoning, and action planning;
A Bayesian Approach to Online Planning
Nir Greshler (General Motors), Aviv Tamar (Technion - Israel Institute of Technology)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabularBenchmark
🎯 What it does: A Bayesian inference-based online planning method is proposed, utilizing the uncertainty estimation of neural networks to improve Monte Carlo Tree Search (MCTS), and providing a finite-time Bayesian return upper bound.
A Bias-Variance-Covariance Decomposition of Kernel Scores for Generative Models
Sebastian Gregor Gruber, Florian Buettner (German Cancer Consortium)
GenerationData SynthesisImageTextAudio
🎯 What it does: A bias-variance-covariance decomposition for kernel scoring is proposed, along with an unbiased consistent estimator based solely on generated samples.
A Circuit Domain Generalization Framework for Efficient Logic Synthesis in Chip Design
Zhihai Wang (University of Science and Technology of China), Feng Wu (University of Science and Technology of China)
OptimizationComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A data-driven heuristic method for logic synthesis called PruneX is proposed, which uses domain-independent representations to predict and prune most invalid node-level transformations, significantly accelerating the logic synthesis process.
A Closer Look at the Limitations of Instruction Tuning
Sreyan Ghosh (University of Maryland), Dinesh Manocha (University of Maryland)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Investigate the limitations of Instruction Tuning and evaluate its impact on the knowledge and capabilities of LLMs through experiments.
A Computational Framework for Solving Wasserstein Lagrangian Flows
Kirill Neklyudov (University of Montreal), Alireza Makhzani (Vector Institute)
OptimizationFlow-based ModelBiomedical Data
🎯 What it does: A unified modeling of various variants (such as OT, Schrödinger bridge, unbalanced OT, etc.) is proposed through the minimization of the Lagrangian action in probability density space, along with a deep learning solving framework;
A connection between Tempering and Entropic Mirror Descent
Nicolas Chopin (ENSAE), Anna Korba (ENSAE)
Optimization
🎯 What it does: This paper establishes the convergence rate of tempering iterations by equating tempering with entropic mirror descent, and derives a set of adaptive tempering rules from an optimization perspective, which are subsequently validated in experiments for their effectiveness under high-dimensional Gaussian targets.
A Contextual Combinatorial Bandit Approach to Negotiation
Yexin Li (State Key Laboratory of General Artificial Intelligence), Siyuan Qi (State Key Laboratory of General Artificial Intelligence)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a negotiation learning framework based on Contextual Combinatorial Multi-Armed Bandit (CCMAB) and designs the NegUCB algorithm to balance exploration and exploitation in negotiation tasks with partial observations, complex acceptance functions, and large action spaces.
A decoder-only foundation model for time-series forecasting
Abhimanyu Das (Google Research), Yichen Zhou (Google Research)
TransformerTime Series
🎯 What it does: A time series foundational model called TimesFM based on a decoder Transformer is proposed, capable of predicting multi-domain and multi-granularity time series under zero-shot conditions.
A Dense Reward View on Aligning Text-to-Image Diffusion with Preference
Shentao Yang (University of Texas at Austin), Mingyuan Zhou (University of Texas at Austin)
GenerationOptimizationReinforcement LearningDiffusion modelImageText
🎯 What it does: This paper proposes an optimization of preference alignment using a dense reward perspective in text-to-image diffusion models.
A Differentiable Partially Observable Generalized Linear Model with Forward-Backward Message Passing
Chengrui Li (Georgia Institute of Technology), Anqi Wu (Georgia Institute of Technology)
OptimizationExplainability and InterpretabilityTabularBiomedical Data
🎯 What it does: A differentiable partially observable generalized linear model (POGLM) is proposed, and a forward-backward message passing sampling scheme is introduced to improve neural connectivity inference.
A Diffusion Model Framework for Unsupervised Neural Combinatorial Optimization
Sebastian Sanokowski (Johannes Kepler University), Sebastian Lehner (Johannes Kepler University)
OptimizationGraph Neural NetworkDiffusion modelGraph
🎯 What it does: This paper proposes an unsupervised neural combinatorial optimization framework based on diffusion models (DiffUCO), achieving approximation and sampling of discrete Boltzmann distributions under data-free conditions.
A Distributional Analogue to the Successor Representation
Harley Wiltzer (McGill University), Mark Rowland (Google DeepMind)
Reinforcement LearningSequential
🎯 What it does: This paper proposes a Distributional Successor Measure (DSM), which decomposes the return distribution into a reward function and DSM, enabling zero-shot distributed policy evaluation for any reward function.
A Doubly Recursive Stochastic Compositional Gradient Descent Method for Federated Multi-Level Compositional Optimization
Hongchang Gao (Temple University)
OptimizationFederated LearningComputational EfficiencyTabularFinance Related
🎯 What it does: A federated learning framework Fed-DR-SCGD is proposed to solve multi-layer stochastic combinatorial optimization problems and achieve linear acceleration in heterogeneous data environments.
A Dual-module Framework for Counterfactual Estimation over Time
Xin Wang (University of Science and Technology of China), Huanhuan Chen (University of Science and Technology of China)
Data SynthesisOptimizationRecurrent Neural NetworkGenerative Adversarial NetworkTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a dual-module framework ACTIN for counterfactual estimation of potential outcomes in time series, addressing the issues of time-varying confounding bias and long-range dependence.
A Dynamic Algorithm for Weighted Submodular Cover Problem
Kiarash Banihashem (University of Maryland), Morteza Monemizadeh (TU Eindhoven)
Optimization
🎯 What it does: This paper proposes a randomized algorithm for solving the weighted submodular covering problem in dynamic environments (where elements can be inserted and deleted), which can maintain an approximately optimal solution with expected polynomial logarithmic query complexity after each update.
A Dynamical Model of Neural Scaling Laws
Blake Bordelon (Harvard University), Cengiz Pehlevan (Harvard University)
TransformerTextStochastic Differential Equation
🎯 What it does: This paper analyzes the training and generalization dynamics of neural networks by constructing a solvable random feature model and using dynamical mean field theory, thereby explaining the emergence and characteristics of neural scaling laws.
A fast algorithm to simulate nonlinear resistive networks
Benjamin Scellier (Rain AI)
OptimizationComputational EfficiencyImage
🎯 What it does: A fast algorithm is proposed to solve the steady state of ideal nonlinear resistor networks using coordinate descent, and a parallel block coordinate descent implementation is designed for deep resistor networks (DRN).
A Federated Stochastic Multi-level Compositional Minimax Algorithm for Deep AUC Maximization
Xinwen Zhang (Temple University), Hongchang Gao (Temple University)
OptimizationFederated LearningImage
🎯 What it does: A federated multi-layer composite minimax algorithm, LocalSMCGDAM, is proposed for maximizing deep AUC.
A Field Guide for Pacing Budget and ROS Constraints
Santiago R. Balseiro (Google Research), Di Wang (Google Research)
OptimizationTabularOrdinary Differential Equation
🎯 What it does: This paper studies and compares three real-time bidding acceleration strategies under budget and ROS constraints: the fully coupled dual optimal method, the sequential decoupling method, and the minimization method that only takes the minimum bid of the two. It proves that the minimization method is theoretically as good as the dual optimal method.
A Fine-grained Analysis of Fitted Q-evaluation: Beyond Parametric Models
Jiayi Wang (University of Texas at Dallas), Raymond K. W. Wong (Texas A&M University)
Reinforcement LearningTabular
🎯 What it does: This paper conducts a fine-grained analysis of the statistical properties of the Fitted Q-Evaluation (FQE) method under both parametric and non-parametric Q-function models, deriving error upper bounds and addressing three key questions: optimal sample complexity, time dependence, and the role of the probability ratio function.
A Fixed-Point Approach for Causal Generative Modeling
Meyer Scetbon (Microsoft Research), Chao Ma (Microsoft Research)
GenerationData SynthesisTransformerGraph
🎯 What it does: This paper proposes a new framework that views Structural Causal Models (SCM) as a fixed point problem under a known topological order, and constructs a two-stage causal generative model.
A Fresh Take on Stale Embeddings: Improving Dense Retriever Training with Corrector Networks
Nicholas Monath (Google DeepMind), Manzil Zaheer (Google DeepMind)
RetrievalTextRetrieval-Augmented Generation
🎯 What it does: This work trains a small target correction network (corrector) to dynamically correct the staleness of cached target embeddings during the training process of dense retrieval and retrieval-augmented language models, thereby avoiding frequent target re-encoding.
A General Framework for Learning from Weak Supervision
Hao Chen (Carnegie Mellon University), Bhiksha Raj (Mohamed bin Zayed University of AI)
ClassificationRecognitionImage
🎯 What it does: A unified weakly supervised learning framework (GLWS) is proposed, which achieves learning for any form of weak supervision through the EM algorithm and non-deterministic finite automata (NFA);
A General Framework for Sequential Decision-Making under Adaptivity Constraints
Nuoya Xiong (Tsinghua University), Zhuoran Yang (Yale University)
Reinforcement LearningSequential
🎯 What it does: This paper proposes a general class of Eluder Condition (EC) functions and provides algorithm implementations under this framework for two adaptive constraints: sparse policy switching and batch learning, proving the corresponding convergence and switching cost upper bounds.
A General Online Algorithm for Optimizing Complex Performance Metrics
Wojciech Kotlowski, Krzysztof Dembczynski (Poznan University of Technology)
ClassificationOptimizationTabular
🎯 What it does: A general online algorithm OMMA is proposed to directly maximize complex non-decomposable performance metrics (such as F-measure, G-mean, etc.) in binary classification, multi-class classification, and multi-label problems, along with a theoretical regret upper bound.
A General Theory for Softmax Gating Multinomial Logistic Mixture of Experts
Huy Nguyen (University of Texas at Austin), Nhat Ho (University of Texas at Austin)
ClassificationOptimizationMixture of Experts
🎯 What it does: This paper presents a general theory of softmax gated mixture of experts models for multi-class tasks and provides the convergence rate of parameter estimation.
A Generative Approach for Treatment Effect Estimation under Collider Bias: From an Out-of-Distribution Perspective
Baohong Li (Zhejiang University), Kun Kuang (Zhejiang University)
GenerationData SynthesisOptimizationGenerative Adversarial NetworkTabular
🎯 What it does: This paper proposes a novel generative method called the Coupled Counterfactual Generative Adversarial Model (C2 GAM), which reinterprets confounding bias as an out-of-distribution (OOD) problem in discrete environments. It utilizes a small amount of unbiased representative data along with a large amount of biased observational data to jointly generate missing S=0 samples and missing S labels, thereby eliminating confounding bias and improving the accuracy of causal effect estimation.
A Geometric Decomposition of Finite Games: Convergence vs. Recurrence under Exponential Weights
Davide Legacci (University Grenoble Alpes), Bary Pradelski (CNRS)
Optimization
🎯 What it does: A geometric decomposition of finite games based on the Shahshahani metric is proposed, revealing the equivalence between potential games and incompressible (i.e., harmonic) games, and proving that under incompressible games, the exponential/multiplicative weight learning (EW) dynamics exhibit conservation, preservation, and Poincaré recurrence;
A Geometric Explanation of the Likelihood OOD Detection Paradox
Hamidreza Kamkari (Layer 6 AI), Gabriel Loaiza-Ganem (Layer 6 AI)
GenerationAnomaly DetectionDiffusion modelFlow-based ModelImage
🎯 What it does: This study investigates the paradox of likelihood-based deep generative models (such as normalizing flows and diffusion models) in OOD detection and proposes a dual-threshold detection method based on geometric interpretation, utilizing the local intrinsic dimension (LID) estimation from pre-trained models to identify OOV samples that have high likelihood but low probability quality.
A Global Geometric Analysis of Maximal Coding Rate Reduction
Peng Wang (University of Michigan), Yi Ma (University of Hong Kong)
OptimizationImage
🎯 What it does: The global geometric properties of the maximum coding rate reduction (MCR-2) objective function were studied, proving that all critical points are either local maxima or strict saddle points, and providing closed-form expressions for local and global optimal solutions.
A Graph is Worth $K$ Words: Euclideanizing Graph using Pure Transformer
Zhangyang Gao (Westlake University), Stan Z. Li (Westlake University)
GenerationData SynthesisDrug DiscoveryGraph Neural NetworkTransformerMultimodalityGraph
🎯 What it does: Transform non-Euclidean graphs into learnable Euclidean vectors (Graph Words) and achieve graph representation and autoregressive generation through pure Transformer.
A Hierarchical Adaptive Multi-Task Reinforcement Learning Framework for Multiplier Circuit Design
Zhihai Wang (University of Science and Technology of China), Feng Wu (Huawei Technologies)
OptimizationReinforcement Learning
🎯 What it does: A hierarchical adaptive multi-task reinforcement learning framework (HAVE) is designed to efficiently explore the design space of multipliers under given multi-objective (area, delay) constraints, ultimately obtaining Pareto optimal solutions that cover the entire weight space.
A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts
Kuang-Huei Lee (Google DeepMind), Ian Fischer (Google DeepMind)
RetrievalCompressionTransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: A system named ReadAgent is proposed, which is an LLM agent system that significantly extends the effective context length and improves long text reading comprehension performance through pagination, gist memory compression, and interactive retrieval.
A Language Model’s Guide Through Latent Space
Dimitri von Rütte (ETH Zurich), Thomas Hofmann (ETH Zurich)
TransformerLarge Language ModelText
🎯 What it does: This paper explores the activation space within large language models to probe linear concept vectors, further enabling reasoning and guidance on various concepts such as appropriateness, humor, creativity, quality, and sincerity. It also proposes a new evaluation metric to measure the trade-off between concept activation effectiveness and fluency loss. Systematic experiments were conducted across multiple models and probes to validate the relationship between concept guidability and detection accuracy.
A Linear Time and Space Local Point Cloud Geometry Encoder via Vectorized Kernel Mixture (VecKM)
Dehao Yuan (University of Maryland), Yiannis Aloimonos (University of Maryland)
ClassificationSegmentationPoint Cloud
🎯 What it does: A local geometric encoder VecKM is proposed, which encodes the local geometry of point clouds into fixed-length complex vectors using vectorized Gaussian kernel mixture.
A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity
Andrew Lee (University of Michigan), Rada Mihalcea (University of Michigan)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the mechanism of the alignment algorithm (DPO) in reducing toxic behavior in large language models, finding that toxic information is represented through specific MLP value vectors. It demonstrates that DPO avoids these toxic trigger areas by slight weight shifts (GPT2) or gating mechanisms (Llama2); it also shows that toxicity can be easily restored by amplifying key vectors or opening gates.
A Minimaximalist Approach to Reinforcement Learning from Human Feedback
Gokul Swamy (Carnegie Mellon University), Alekh Agarwal (Google Research)
Reinforcement Learning from Human FeedbackReinforcement LearningSequential
🎯 What it does: A reinforcement learning from human feedback (RLHF) algorithm called Self-Play Preference Optimization (SPO) is proposed, which directly utilizes human preferences (or preference models) for self-play training without the need for a reward model or adversarial training.
A Multimodal Automated Interpretability Agent
Tamar Rott Shaham (Massachusetts Institute of Technology), Antonio Torralba (Massachusetts Institute of Technology)
Explainability and InterpretabilityTransformerLarge Language ModelVision Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: MAIA is proposed - an automated interpretation agent that uses a multimodal language model (GPT-4V) and a set of programmable interpretation tools, capable of conducting interpretability analysis of neural networks by constructing experimental programs;
A Near-Linear Time Approximation Algorithm for Beyond-Worst-Case Graph Clustering
Vincent Cohen-Addad (Google Research), Aida Mousavifar (Google Research)
OptimizationGraph Neural NetworkGraphStochastic Differential Equation
🎯 What it does: An approximate algorithm with near-linear time complexity is presented under the semi-random graph model (Model 1.1) to solve the Balanced Cut problem, achieving solutions of comparable quality to the theoretically optimal approximation proposed by Makarychev et al.
A Nearly Optimal Single Loop Algorithm for Stochastic Bilevel Optimization under Unbounded Smoothness
Xiaochuan Gong (George Mason University), Mingrui Liu (George Mason University)
OptimizationRepresentation LearningText
🎯 What it does: A single-loop double-layer optimization algorithm SLIP is proposed for stochastic double-layer optimization problems where the upper-level function is unboundedly smooth and the lower-level function is strongly convex.
A Neural-Guided Dynamic Symbolic Network for Exploring Mathematical Expressions from Data
Wenqiang Li (Institute of Semiconductors, Chinese Academy of Sciences), Meilan Hao (Institute of Semiconductors, Chinese Academy of Sciences)
Recurrent Neural NetworkReinforcement LearningTabularTime SeriesSequentialPhysics Related
🎯 What it does: A neural-guided dynamic symbolic network DYSYMNET is proposed for automatically discovering interpretable mathematical expressions from data.
A Neural-Preconditioned Poisson Solver for Mixed Dirichlet and Neumann Boundary Conditions
Kai Weixian Lan (University of California), Joseph Teran (University of California)
OptimizationConvolutional Neural NetworkTabular
🎯 What it does: An iterative solver based on a neural network preprocessor is proposed to solve the Poisson equation with mixed Dirichlet/Neumann boundary conditions.
A New Branch-and-Bound Pruning Framework for $\ell_0$-Regularized Problems
Theo Guyard, Ayse-Nur Arslan
OptimizationTabular
🎯 What it does: A new Branch-and-Bound (BnB) pruning framework is proposed for solving optimization problems with ∏_0 regularization terms, which can efficiently evaluate multiple subregions for the presence of optimal solutions without solving convex relaxations.
A New Computationally Efficient Algorithm to solve Feature Selection for Functional Data Classification in High-dimensional Spaces
Tobia Boschi (IBM Research Europe), Jonathan P Epperlein
ClassificationComputational EfficiencyTabularTime SeriesElectronic Health Records
🎯 What it does: An efficient algorithm for simultaneous feature selection and functional classification (FSFC) is proposed and applied to multivariate longitudinal data.
A New Linear Scaling Rule for Private Adaptive Hyperparameter Optimization
Ashwinee Panda (Princeton University), Prateek Mittal (Princeton University)
OptimizationSafty and PrivacyHyperparameter SearchImageText
🎯 What it does: A private adaptive hyperparameter optimization method based on linear scaling rules is proposed, which first estimates hyperparameters through small-scale experiments under low privacy budgets and then scales up to larger budgets in a linear manner, significantly reducing the computational and privacy costs of hyperparameter search.
A New Robust Partial p-Wasserstein-Based Metric for Comparing Distributions
Sharath Raghvendra (North Carolina State University), Kaiyi Zhang (Virginia Tech)
RetrievalImage
🎯 What it does: A new distance based on partial p-Wasserstein, denoted as (p, k)-RPW, is proposed, which combines the sensitivity of 2-Wasserstein with robustness against outliers and sampling errors.
A New Theoretical Perspective on Data Heterogeneity in Federated Optimization
Jiayi Wang (University of Utah), Mingyue Ji (University of Utah)
OptimizationFederated LearningConvolutional Neural NetworkRecurrent Neural NetworkImage
🎯 What it does: This paper recharacterizes data heterogeneity in federated learning by introducing the 'heterogeneity-driven pseudo-Lipschitz constant' (L_h), and based on this, provides convergence upper bounds for FedAvg and its extended algorithms, proving that increasing the number of local updates can still improve convergence speed even in cases of highly non-homogeneous data.
A Persuasive Approach to Combating Misinformation
Safwan Hossain (Harvard University), Gauthier Gidel (Mila, Université de Montréal)
🎯 What it does: This paper proposes using a Bayesian persuasion (information design) framework to enable social media platforms to strategically signal based on predictions of user content, thereby influencing whether users share content and reducing the spread of misleading information.
A Primal-Dual Algorithm for Offline Constrained Reinforcement Learning with Linear MDPs
Kihyuk Hong (University of Michigan), Ambuj Tewari (University of Michigan)
Computational EfficiencyReinforcement Learning
🎯 What it does: A highly efficient primal-dual algorithm is proposed for offline constrained linear MDP learning, achieving a sample complexity of O(ε⁻²) under the assumption of partial data coverage.
A Probabilistic Approach to Learning the Degree of Equivariance in Steerable CNNs
Lars Veefkind (University of Amsterdam), Gabriele Cesa (Qualcomm AI Research)
ClassificationRecognitionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a probabilistic method to learn and explain the adjustable degree of equivariance in differentiable Steerable CNNs.
A Provable Decision Rule for Out-of-Distribution Detection
Xinsong Ma (Wuhan University), Weiwei Liu (Wuhan University)
Anomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an out-of-distribution (OOD) detection decision rule based on the generalized Benjamini–Hochberg (g-BH) procedure, and provides theoretical control of false positives and an upper bound on the expected false positive rate.
A Provably Effective Method for Pruning Experts in Fine-tuned Sparse Mixture-of-Experts
Mohammed Nowaz Rabbani Chowdhury (Rensselaer Polytechnic Institute), Christopher Carothers (Rensselaer Polytechnic Institute)
OptimizationComputational EfficiencySupervised Fine-TuningMixture of ExpertsImage
🎯 What it does: This paper studies an expert pruning method for fine-tuned sparse mixture of experts (MoE) models, providing both theoretical and experimental proofs.
A Rate-Distortion View of Uncertainty Quantification
Ifigeneia Apostolopoulou (Carnegie Mellon University), Artur Dubrawski (Carnegie Mellon University)
ClassificationAnomaly DetectionAuto EncoderImage
🎯 What it does: A distance-aware bottleneck (DAB) model based on information bottleneck and rate-distortion theory is proposed, which estimates uncertainty by learning a codebook to compress training samples and measuring the statistical distance between the input and the codebook.
A Resilient and Accessible Distribution-Preserving Watermark for Large Language Models
Yihan Wu (University of Maryland), Heng Huang (University of Maryland)
GenerationTransformerLarge Language ModelText
🎯 What it does: A distribution-preserving, easily retrievable, and robust LLM watermarking framework called DiPmark is proposed, which can mark generated content without affecting text quality.
A sampling theory perspective on activations for implicit neural representations
Hemanth Saratchandran (University of Adelaide), Simon Lucey (University of Adelaide)
RestorationData SynthesisNeural Radiance FieldImageTime SeriesOrdinary Differential Equation
🎯 What it does: This paper provides a unified analysis of activation functions in implicit neural representations (INR) from the perspective of sampling theory and proves that the sinc activation can theoretically optimally reconstruct signals.
A Simple Early Exiting Framework for Accelerated Sampling in Diffusion Models
Taehong Moon (KRAFTON), Juho Lee (Graduate School of AI, KAIST)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: An adaptive early exit framework ASE is proposed for diffusion models, which accelerates sampling by dynamically dropping network blocks based on time steps.
A Single-Loop Robust Policy Gradient Method for Robust Markov Decision Processes
Zhenwei Lin (Shanghai University of Finance and Economics), Yinyu Ye (Stanford University)
OptimizationReinforcement Learning
🎯 What it does: Proposes a Single-Cycle Robust Policy Gradient (SRPG) method for solving Robust Markov Decision Processes (RMDP) with s-rectangular uncertainty sets;
A Sober Look at LLMs for Material Discovery: Are They Actually Good for Bayesian Optimization Over Molecules?
Agustinus Kristiadi (Vector Institute), Geoff Pleiss (University of British Columbia)
OptimizationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper explores the use of large language models (LLMs) for Bayesian optimization (BO) in molecular space and evaluates their effectiveness in practical chemistry tasks.
A Space Group Symmetry Informed Network for O(3) Equivariant Crystal Tensor Prediction
Keqiang Yan (Texas A&M University), Shuiwang Ji (Texas A&M University)
Graph Neural NetworkTabularPhysics Related
🎯 What it does: GMTNet is proposed, a graph neural network based on crystal space group symmetry and O(3) equivariance, for predicting the dielectric, piezoelectric, and elastic tensors of crystals.
A Sparsity Principle for Partially Observable Causal Representation Learning
Danru Xu (University of Amsterdam), Sara Magliacane
Representation LearningTabular
🎯 What it does: This paper studies how to recover latent causal variables from high-dimensional observations using unpaired, instance-dependent biased observational data in the context of partially observable causal representation learning (CRL) problems.
A Statistical Framework for Data-dependent Retrieval-Augmented Models
Soumya Basu (Google), Manzil Zaheer (Google DeepMind)
RetrievalOptimizationTextBenchmarkRetrieval-Augmented Generation
🎯 What it does: This paper proposes a statistical framework that unifies the end-to-end training method of Retrieval-Augmented Models (RAM) and provides its risk upper bound.
A Statistical Theory of Regularization-Based Continual Learning
Xuyang Zhao (Peking University), Wei Lin (Peking University)
Supervised Fine-TuningTabular
🎯 What it does: In the framework of continuous learning linear regression, the statistical properties of the general ℓ2 regularization algorithm are systematically derived and analyzed. Optimal regularization weights that achieve a balance between forward and backward knowledge transfer are proposed, and equivalences with methods such as early stopping are provided.
A Study of First-Order Methods with a Deterministic Relative-Error Gradient Oracle
Nadav Hallak (Technion), Kfir Yehuda Levy
Optimization
🎯 What it does: This paper studies the theoretical convergence properties of classical first-order methods—Projection Gradient (PG) and Conditional Gradient (CG)—when using a deterministic relative error gradient operator (error gradient Oracle) in constrained optimization problems.
A Subquadratic Time Algorithm for Robust Sparse Mean Estimation
Ankit Pensia (IBM Research)
Anomaly DetectionOptimization
🎯 What it does: This paper proposes an algorithm for estimating sparse high-dimensional Gaussian means under the presence of robust attacks (strong contamination); the algorithm completes the estimation in sub-quadratic time and guarantees an error of O(ε√log(1/ε)) with a sample size of polynomial (k, log d, 1/ε).
A Tale of Tails: Model Collapse as a Change of Scaling Laws
Elvis Dohmatob (Meta), Julia Kempe (New York University)
GenerationData SynthesisLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies how the scaling laws of neural networks change when artificial synthetic data is added to the training corpus, and provides a theoretical explanation and experimental evidence for the phenomenon of model collapse.
A Tensor Decomposition Perspective on Second-order RNNs
Maude Lizaire (University of Montreal), Guillaume Rabusseau (University of Montreal)
Recurrent Neural NetworkText
🎯 What it does: Theoretical and experimental research on second-order recurrent neural networks (2RNN) is conducted, proposing the use of CP decomposition to compress its third-order weight tensor, forming the CPRNN model, and exploring the impact of rank and hidden size on expressive power.
A Theoretical Analysis of Backdoor Poisoning Attacks in Convolutional Neural Networks
Boqi Li (Wuhan University), Weiwei Liu (Wuhan University)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A dual analysis of backdoor injection attacks in convolutional neural networks is conducted, clarifying the mechanisms behind successful backdoors.
A Theory of Fault-Tolerant Learning
Changlong Wu (Purdue University), Ananth Grama (Purdue University)
🎯 What it does: A fault-tolerant PAC learning framework based on PAC learning theory is proposed, providing upper and lower bounds on sample complexity under random and adversarial fault scenarios, particularly offering a sample complexity that is almost independent of the size of the fault set for deletion faults in neural networks.
A Theory of Non-Linear Feature Learning with One Gradient Step in Two-Layer Neural Networks
Behrad Moniri (University of Pennsylvania), Edgar Dobriban (University of Pennsylvania)
OptimizationRepresentation LearningTabularStochastic Differential Equation
🎯 What it does: This paper studies how a two-layer neural network learns nonlinear features after one gradient descent step (for the first layer weights) and provides a high-dimensional limit analysis of training and testing errors.
A Touch, Vision, and Language Dataset for Multimodal Alignment
Letian Fu (University of California Berkeley), Ken Goldberg (University of California Berkeley)
ClassificationGenerationRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper presents the TVL dataset and trains an aligned tactile encoder and a TVL-LLaMA model capable of generating tactile descriptions through contrastive learning across the three modalities of touch, vision, and language.
A Unified Adaptive Testing System Enabled by Hierarchical Structure Search
Junhao Yu (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Reinforcement LearningTabular
🎯 What it does: A unified adaptive testing system framework is proposed, treating CAT and MST as hierarchical search problems, and achieving automatic generation of optimal problem sequences, eliminating the need for manual design.
A Unified Framework for Learning with Nonlinear Model Classes from Arbitrary Linear Samples
Ben Adcock (Simon Fraser University), Nick Dexter (Florida State University)
CompressionOptimizationMultimodalityMagnetic Resonance Imaging
🎯 What it does: A unified framework is proposed for learning unknown objects in nonlinear model classes under arbitrary linear sampling (which can be multimodal and vectorized), along with a corresponding upper bound on the learning error.
A Unified Linear Programming Framework for Offline Reward Learning from Human Demonstrations and Feedback
Kihyun Kim (Massachusetts Institute of Technology), Pablo Parrilo (Massachusetts Institute of Technology)
OptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper proposes a unified linear programming (LP) framework for offline reward learning, which can infer reward functions from both expert demonstration data (IRL) and human comparative feedback on trajectories (RLHF).
A Unified View of FANOVA: A Comprehensive Bayesian Framework for Component Selection and Estimation
yosra marnissi, Maxime Leiber (INRIA)
Tabular
🎯 What it does: A unified Bayesian FANOVA framework is proposed, capable of simultaneously performing component selection, covariate selection, and prediction, and providing scalable implementation details.
A Universal Class of Sharpness-Aware Minimization Algorithms
Behrooz Tahmasebi (Massachusetts Institute of Technology), Patrick Jaillet (Massachusetts Institute of Technology)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: A class of general sharpness measures and corresponding sharpness-aware optimization algorithms (Sharpness-Aware Minimization) is proposed, and it is proven to be expressive for any Hessian function; specific instances (Frob-SAM, Det-SAM) are validated for their generalization advantages in over-parameterized models.
A Universal Transfer Theorem for Convex Optimization Algorithms Using Inexact First-order Oracles
Phillip Kerger (Johns Hopkins University), Amitabh Basu (Johns Hopkins University)
Optimization
🎯 What it does: A general 'transfer theorem' is proposed, which can rewrite any convex optimization algorithm that uses precise first-order information into a version that accepts imprecise first-order information at a black-box level.
A2Q+: Improving Accumulator-Aware Weight Quantization
Ian Colbert (AMD), Yaman Umuroglu (AMD)
Super ResolutionOptimizationConvolutional Neural NetworkImage
🎯 What it does: A2Q+ is proposed, which improves the accumulator-aware quantization algorithm while avoiding accumulator overflow, further enhancing model accuracy under low-bit-width accumulators.
A3S: A General Active Clustering Method with Pairwise Constraints
Xun Deng (University of Science and Technology of China), Zheng Wang (Alibaba Group)
OptimizationImage
🎯 What it does: An adaptive active clustering framework A3S is proposed, which utilizes human pairwise constraints to aggregate and split the initial clustering results, significantly improving clustering quality.
Absolute Policy Optimization: Enhancing Lower Probability Bound of Performance with High Confidence
Weiye Zhao (Carnegie Mellon University), Changliu Liu (Carnegie Mellon University)
OptimizationReinforcement LearningSequential
🎯 What it does: This paper proposes Absolute Policy Optimization (APO) and its efficient variant PAPO, aiming to ensure a monotonic improvement of the lower probability bound of performance distribution (i.e., worst-case performance) with high confidence, and further enhance expected performance based on this.
Accelerated Algorithms for Constrained Nonconvex-Nonconcave Min-Max Optimization and Comonotone Inclusion
Yang Cai (Yale University), Weiqiang Zheng (Yale University)
Optimization
🎯 What it does: Two single-loop accelerated algorithms for constrained non-convex-concave min-max optimization and its generalized co-monotonic inclusion problem are proposed: composite-EAG and composite-FEG.
Accelerated Policy Gradient for s-rectangular Robust MDPs with Large State Spaces
Ziyi Chen (University of Maryland), Heng Huang (University of Maryland)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes an accelerated policy gradient algorithm for solving robust Markov decision processes (Robust MDP) with s-rectangular uncertainty sets and extends it to stochastic environments and large state spaces.
Accelerated Policy Gradient: On the Convergence Rates of the Nesterov Momentum for Reinforcement Learning
Yen-Ju Chen (National Yang Ming Chiao Tung University), Ping-Chun Hsieh (National Yang Ming Chiao Tung University)
Reinforcement LearningSequential
🎯 What it does: Proposes Accelerated Policy Gradient (APG), using Nesterov momentum to accelerate policy gradients in reinforcement learning;
Accelerated Speculative Sampling Based on Tree Monte Carlo
Zhengmian Hu (University of Maryland), Heng Huang (University of Maryland)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A sampling framework based on tree space, Tree Monte Carlo (TMC), and its special case, Accelerated Speculative Sampling (ASpS), are proposed to accelerate the inference of large language models (LLMs) by improving the maximum coupling strategy.
Accelerating Convergence in Bayesian Few-Shot Classification
Tianjun Ke (Renmin University of China), Feng Zhou (Renmin University of China)
ClassificationOptimizationMeta LearningImage
🎯 What it does: A variational inference method based on mirror descent is proposed for Bayesian few-shot classification under Gaussian processes (GP);
Accelerating Convergence of Score-Based Diffusion Models, Provably
Gen Li (Chinese University of Hong Kong), Yuxin Chen (University of Pennsylvania)
GenerationData SynthesisComputational EfficiencyDiffusion modelScore-based ModelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A training-independent accelerated sampling algorithm is proposed, achieving faster sampling speeds for both deterministic (DDIM) and stochastic (DDPM) diffusion models.
Accelerating Federated Learning with Quick Distributed Mean Estimation
Ran Ben-Basat (University College London), Michael Mitzenmacher (Harvard University)
OptimizationFederated LearningComputational EfficiencyRecurrent Neural NetworkReinforcement LearningImageText
🎯 What it does: A new Distributed Mean Estimation (DME) method called QUIC-FL is proposed, which significantly reduces client encoding time and server decoding time while ensuring optimal O(1/n) NMSE.
Accelerating Heterogeneous Federated Learning with Closed-form Classifiers
Eros Fanì (Polytechnic University of Turin), Marco Ciccone (Polytechnic University of Turin)
Federated LearningComputational EfficiencySupervised Fine-TuningImage
🎯 What it does: A FED3R method is proposed for training classifiers in federated learning using pre-trained features and closed-form ridge regression, along with its random feature version and fine-tuning extension.
Accelerating Iterative Retrieval-augmented Language Model Serving with Speculation
Zhihao Zhang (Carnegie Mellon University), Zhihao Jia (Carnegie Mellon University)
RetrievalOptimizationComputational EfficiencyTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes the RaLMSpec framework, which accelerates the iterative retrieval-enhanced language model service using techniques such as speculative retrieval with batch validation, cache prefetching, optimal speculation step scheduling, and asynchronous validation.
Accelerating Legacy Numerical Solvers by Non-intrusive Gradient-based Meta-solving
Sohei Arisaka (National University of Singapore), Qianxiao Li (National University of Singapore)
OptimizationComputational Efficiency
🎯 What it does: A non-invasive gradient-based Meta-solving method is proposed, which accelerates traditional non-automatic differentiation numerical solvers using control variate forward gradients.
Accelerating Look-ahead in Bayesian Optimization: Multilevel Monte Carlo is All you Need
Shangda Yang (University of Manchester), Kody J. H. Law (Meta Platforms, Inc.)
OptimizationTabular
🎯 What it does: A framework is proposed that applies the Multi-Layer Monte Carlo (MLMC) method to multi-step Bayesian optimization, significantly reducing the computational complexity of estimating the acquisition function in a forward-looking manner.
Accelerating Parallel Sampling of Diffusion Models
Zhiwei Tang (Chinese University of Hong Kong), Tsung-Hui Chang (Chinese University of Hong Kong)
GenerationComputational EfficiencyDiffusion modelImage
🎯 What it does: This paper proposes a parallel sampling framework called ParaTAA, which can significantly accelerate the sampling process of diffusion models while maintaining image quality.