NeurIPS 2024 Papers — Page 27
Conference on Neural Information Processing Systems · 4035 papers
OneActor: Consistent Subject Generation via Cluster-Conditioned Guidance
Jiahao Wang (Xi'an Jiaotong University), Hao Sun (China Telecom Artificial Intelligence Technology)
GenerationData SynthesisDiffusion modelImage
🎯 What it does: A one-shot OneActor method is proposed, achieving high-quality consistent image generation without additional data by learning clustering guidance in the semantic space.
OneBit: Towards Extremely Low-bit Large Language Models
Yuzhuang Xu (Harbin Institute of Technology), Wanxiang Che (Harbin Institute of Technology)
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a OneBit framework that compresses LLM weights to 1-bit and maintains accuracy through bidirectional vectors, achieving extremely low-bit quantization.
OneRef: Unified One-tower Expression Grounding and Segmentation with Mask Referring Modeling
Linhui Xiao (Institute of Automation Chinese Academy of Sciences), Changsheng Xu (Institute of Automation Chinese Academy of Sciences)
RecognitionObject DetectionSegmentationTransformerVision Language ModelImageMultimodality
🎯 What it does: Proposes the OneRef framework and Mask Referring Modeling (MRefM) technique, utilizing a single-tower shared Transformer for unified encoding of vision and text, achieving direct regression for expression localization and segmentation.
Online Adaptation of Language Models with a Memory of Amortized Contexts
Jihoon Tack (KAIST), Jonathan Richard Schwarz (Harvard University)
Domain AdaptationOptimizationMeta LearningTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes an online adaptation framework based on Memory-Augmented Context (MAC), which can quickly adapt large language models (LLMs) by compressing new documents into learnable modulations (PEFT) and storing them in a memory bank, allowing for the aggregation of a single modulation during query time without updating the large model parameters.
Online Bayesian Persuasion Without a Clue
Francesco Bacchiocchi (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
🎯 What it does: This paper proposes a learning algorithm for online Bayesian persuasion without any prior knowledge of the prior distribution and the receiver's utility, and provides its asymptotic no-regret property and sample complexity analysis.
Online Budgeted Matching with General Bids
Jianyi Yang (University of Houston), Shaolei Ren (University of California)
Recommendation SystemOptimizationTabular
🎯 What it does: A meta-algorithm named MetaAd is proposed, which achieves provable competitive ratios for the online budget matching problem without satisfying the small bid and FLM assumptions.
Online Classification with Predictions
Vinod Raman (University of Michigan), Ambuj Tewari (University of Michigan)
ClassificationOptimization
🎯 What it does: This paper proposes an online classification framework with a predictor that can access future samples, and designs a learning algorithm that maintains worst-case guarantees even when predictions are incorrect, while achieving lower errors when predictions are accurate.
Online Composite Optimization Between Stochastic and Adversarial Environments
Yibo Wang (Nanjing University), Lijun Zhang (Nanjing University)
OptimizationTabular
🎯 What it does: A new framework for online composite optimization called Composite SEA is proposed, which can adapt between random and adversarial environments;
Online Consistency of the Nearest Neighbor Rule
Geelon So (University of California San Diego), Sanjoy Dasgupta (University of California San Diego)
🎯 What it does: Proved the online consistency of the nearest neighbor rule in a more general online learning environment, covering measurable functions, metric spaces, and non-i.i.d. processes.
Online Control in Population Dynamics
Noah Golowich (Massachusetts Institute of Technology), Y. Jennifer Sun (Princeton University)
OptimizationReinforcement LearningTime Series
🎯 What it does: This paper proposes a general framework based on online control for real-time regulation of population dynamics under disturbances.
Online Control with Adversarial Disturbance for Continuous-time Linear Systems
Jingwei Li (Tsinghua University), Jingzhao Zhang (Tsinghua University)
OptimizationReinforcement LearningTime Series
🎯 What it does: This study investigates online control of continuous-time linear systems under limited sampling rates, aiming to design an online program that learns and performs comparably to a fixed optimal linear controller in the presence of non-random noise.
Online Convex Optimisation: The Optimal Switching Regret for all Segmentations Simultaneously
Stephen Pasteris (Alan Turing Institute), Mark Herbster (University College London)
Optimization
🎯 What it does: This paper proposes a new meta-algorithm RESET for online convex optimization problems, which can achieve optimal switching regret for all possible segments and adaptive dynamic regret control without prior knowledge of segmentations.
Online Estimation via Offline Estimation: An Information-Theoretic Framework
Dylan J Foster, Alexander Rakhlin (Massachusetts Institute of Technology)
🎯 What it does: This paper proposes a new information-theoretic framework called Oracle-Efficient Online Estimation (OEOE), aimed at studying how to convert offline estimation algorithms into online estimation algorithms and exploring their statistical and computational complexities.
Online Feature Updates Improve Online (Generalized) Label Shift Adaptation
Ruihan Wu (University of California San Diego), Kilian Q Weinberger
Domain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: An online label shift adaptation framework OLS-OFU is proposed, which enhances prediction performance in online label shift and generalized label shift scenarios by dynamically updating the feature extractor using self-supervised learning (SSL) during the testing phase.
Online Iterative Reinforcement Learning from Human Feedback with General Preference Model
Chenlu Ye (University of Illinois Urbana-Champaign), Tong Zhang (University of Illinois Urbana-Champaign)
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a RLHF framework based on general preferences or oracles (independent of traditional BT reward models), employing a minimax game with KL regularization, and presents two sample-efficient algorithms: offline (PELHF) and online (ELHF-IPO);
Online Learning of Delayed Choices
Recep Yusuf Bekci (University of Waterloo)
Recommendation SystemOptimizationTabular
🎯 What it does: An online learning algorithm for the MNL model is proposed to achieve adaptive combination recommendations in the presence of delayed feedback.
Online Learning with Sublinear Best-Action Queries
Matteo Russo (Sapienza University of Rome), Niek Tax (Meta)
🎯 What it does: This paper proposes the use of a limited number of best-action queries in online learning to significantly reduce loss, providing optimal upper and lower bounds under both full information and label efficiency feedback models.
Online Non-convex Learning in Dynamic Environments
Zhipan Xu (Nanjing University), Lijun Zhang (Nanjing University)
OptimizationMeta LearningReinforcement LearningSequential
🎯 What it does: This paper studies the problem of online non-convex learning in dynamic environments, proposing the FTPL-D and FTPL-D+ algorithms to minimize dynamic regret, and the FTPL-A algorithm to minimize adaptive regret.
Online Posterior Sampling with a Diffusion Prior
Branislav Kveton (Adobe Research), Rui Song (Amazon)
Recommendation SystemDiffusion modelTabular
🎯 What it does: A posterior sampling method using diffusion model priors in contextual multi-armed bandits is proposed, along with approximate implementations for linear models and generalized linear models.
Online Relational Inference for Evolving Multi-agent Interacting Systems
Beomseok Kang (Georgia Institute of Technology), Saibal Mukhopadhyay (Georgia Institute of Technology)
Graph Neural NetworkReinforcement LearningGraph
🎯 What it does: An online relational inference framework ORI is proposed, which utilizes a trainable adjacency matrix and online backpropagation to identify the hidden interaction graph of multi-agent systems in real-time.
Online Weighted Paging with Unknown Weights
Orin Levy (Tel Aviv University), Aviv Rosenberg (Google Research)
Optimization
🎯 What it does: An online weighted replacement algorithm is proposed, which can learn online and make caching decisions in the absence of known page weights.
OnlineTAS: An Online Baseline for Temporal Action Segmentation
Qing Zhong (University of Adelaide), Angela Yao (National University of Singapore)
SegmentationRecurrent Neural NetworkTransformerVideo
🎯 What it does: An online temporal action segmentation framework is proposed, which includes an adaptive memory bank, a context-aware feature enhancement module, and post-processing methods for action segmentation of untrimmed videos without a look-back window.
Only Strict Saddles in the Energy Landscape of Predictive Coding Networks?
Francesco Innocenti (University of Sussex), Christopher Buckley
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper studies the energy landscape of the Predictive Coding Network (PCN) at the inference equilibrium point, derives a closed-form expression for deep linear networks, and proves that many originally non-strict saddle points become strict saddle points under equilibrium energy. Experiments confirm the theoretical predictions and propose the conjecture that all saddle points are strict saddle points.
OPEL: Optimal Transport Guided ProcedurE Learning
Sayeed Shafayet Chowdhury (Purdue University), Kaushik Roy (Purdue University)
OptimizationContrastive LearningVideo
🎯 What it does: Unsupervised process learning for multi-video tasks, utilizing the optimal transport (OT) framework for alignment and localization of key steps.
Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives
Vincent Hanke (CISPA Helmholtz Center for Information Security), Adam Dziedzic (CISPA Helmholtz Center for Information Security)
GenerationSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper systematically evaluates the differences between open-source LLMs and closed-source LLMs in terms of privacy protection adaptation, analyzing privacy leakage, performance, and costs, and verifies that the open-source LLM with private gradient tuning methods outperforms the closed-source LLM's private prompt learning methods.
Open-Book Neural Algorithmic Reasoning
Hefei Li (East China Normal University), Zhengfeng Yang (East China Normal University)
Graph Neural NetworkBenchmark
🎯 What it does: An open-book neural algorithm inference framework is proposed, allowing the network to access and utilize instances from the training set during the inference process.
Open-Vocabulary Object Detection via Language Hierarchy
Jiaxing Huang (Nanyang Technological University), Shijian Lu (Nanyang Technological University)
Object DetectionConvolutional Neural NetworkPrompt EngineeringImage
🎯 What it does: The DetLH method is proposed, which utilizes the WordNet language hierarchy to expand weakly supervised image-level labels and co-regularizes with self-training, while introducing the language hierarchy in prompt generation to enhance general object detection performance across datasets.
OpenDlign: Open-World Point Cloud Understanding with Depth-Aligned Images
Ye Mao (Imperial College London), Krystian Mikolajczyk (Imperial College London)
ClassificationObject DetectionRetrievalTransformerDiffusion modelContrastive LearningImagePoint Cloud
🎯 What it does: The OpenDlign framework is proposed, which achieves open-world zero-shot and few-shot learning of 3D point clouds by projecting point clouds into multi-view depth maps and using diffusion models to generate texture-diverse depth-aligned images.
OpenGaussian: Towards Point-Level 3D Gaussian-based Open Vocabulary Understanding
Yanmin Wu (Peking University), Jian Zhang (Peking University)
Object DetectionSegmentationContrastive LearningGaussian SplattingPoint Cloud
🎯 What it does: Proposes a point-level open vocabulary understanding method based on 3D Gaussian Splatting called OpenGaussian, which trains consistent and distinguishable instance features and achieves a lossless association between points and language features;
OPERA: Automatic Offline Policy Evaluation with Re-weighted Aggregates of Multiple Estimators
Allen Nie (Stanford University), Emma Brunskill (Stanford University)
Reinforcement LearningTabularSequentialBiomedical Data
🎯 What it does: A new offline policy evaluation framework called OPERA is proposed, which can achieve lower mean square error evaluation results by weighted fusion of multiple OPE estimators.
Operator World Models for Reinforcement Learning
Pietro Novelli (Istituto Italiano di Tecnologia), Carlo Ciliberto (University College London)
Reinforcement LearningWorld ModelTabular
🎯 What it does: This paper proposes a new reinforcement learning algorithm called POWR, which combines Policy Mirror Descent (PMD) with world model learning based on Conditional Mean Embedding (CME). It can estimate the action value function of any policy through closed-form matrix operations in an infinite state space, thereby achieving globally optimal policy updates.
Opponent Modeling based on Subgoal Inference
XiaoPeng Yu, Zongqing Lu (Peking University)
Reinforcement LearningSequential
🎯 What it does: A subgoal reasoning-based opponent modeling method (OMG) is proposed, which infers the opponent's subgoals from historical trajectories and inputs them as state features into the policy network, enabling rapid adaptation to unknown opponents.
Opponent Modeling with In-context Search
Yuheng Jing (Institute of Automation Chinese Academy of Sciences), Jian Cheng (Institute of Automation Chinese Academy of Sciences)
TransformerReinforcement LearningAgentic AITabular
🎯 What it does: A context learning and decision-time search-based opponent modeling method called OMIS is proposed, which can adapt and perform stably in unknown non-stationary opponent environments.
OptEx: Expediting First-Order Optimization with Approximately Parallelized Iterations
Yao Shu (Guangdong Lab of AI and Digital Economy), Fei Yu
OptimizationTransformerReinforcement LearningImageText
🎯 What it does: The OptEx framework is proposed, which utilizes kernelized gradient estimation based on historical gradients to achieve approximate parallel iterations of first-order optimization (FOO) algorithms, significantly reducing the required number of sequential iterations.
Optical Diffusion Models for Image Generation
Ilker Oguz (École Polytechnique Fédérale de Lausanne), Demetri Psaltis (École Polytechnique Fédérale de Lausanne)
GenerationDiffusion modelImage
🎯 What it does: A full optical denoising diffusion model based on optical diffraction has been developed, using light wave propagation to achieve image generation.
Optimal ablation for interpretability
Maximilian Li (Harvard University), Lucas Janson (Harvard University)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: A new method called 'Optimal Ablation' (OA) is proposed to evaluate the importance of internal components of models, and it is applied to interpretive tasks such as circuit discovery, fact recall, and potential prediction.
Optimal Aggregation of Prediction Intervals under Unsupervised Domain Shift
Jiawei Ge (Princeton University), Jianqing Fan (Princeton University)
Domain AdaptationOptimizationTabular
🎯 What it does: In the unsupervised domain transfer scenario, we construct prediction intervals and obtain coverage-reliable, minimal-width intervals through model aggregation.
Optimal Algorithms for Augmented Testing of Discrete Distributions
Maryam Aliakbarpour (Rice University), Sandeep Silwal (University of Wisconsin Madison)
TabularTime Series
🎯 What it does: This paper studies the hypothesis testing problem of discrete distributions and proposes a new algorithmic framework that utilizes predictive distribution information to reduce sample complexity, applicable to uniformity testing, identity testing, and proximity testing.
Optimal Algorithms for Learning Partitions with Faulty Oracles
Adela Frances DePavia (University of Chicago), Erasmo Tani (University of Chicago)
Optimization
🎯 What it does: The study investigates the scenario of learning unknown partitions with at most ℓ errors under the same-cluster query model, designing an optimal query algorithm and providing upper and lower bounds for the minimum number of queries.
Optimal Algorithms for Online Convex Optimization with Adversarial Constraints
Abhishek Sinha (Tata Institute of Fundamental Research), Rahul Vaze (Tata Institute of Fundamental Research)
Anomaly DetectionOptimizationTabularFinance Related
🎯 What it does: This paper designs an efficient first-order algorithm for the constrained online convex optimization (COCO) problem, achieving optimal O(√T) regret and O(√T log T) cumulative constraint violation (CCV) without additional assumptions.
Optimal and Approximate Adaptive Stochastic Quantization
Ran Ben-Basat, shay vargaftik
OptimizationComputational EfficiencyData-Centric Learning
🎯 What it does: An efficient algorithm for the problem of Adaptive Stochastic Quantization (ASQ) is proposed, capable of achieving optimal or approximate quantization on large-scale vectors within acceptable time and memory limits.
Optimal Batched Best Arm Identification
Tianyuan Jin (National University of Singapore), Pan Xu (Duke University)
OptimizationBiomedical Data
🎯 What it does: This paper studies the problem of Batch Best Arm Identification (BBAI) and proposes the Tri-BBAI algorithm and the nearly optimal Batch Best Arm Identification (Opt-BBAI) algorithm, aiming to find the best arm with minimal sample complexity and batch complexity.
Optimal Classification under Performative Distribution Shift
Edwige Cyffers (University of Lille), Olivier Cappé
ClassificationOptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes modeling the performative effect as a push-forward operation and provides a corresponding gradient estimation method; it proves that linear shifts can lead to the convexity of performative risk in the binary classification case and relates it to robust classification; it also introduces the RPPerfGD algorithm based on gradient descent, which enhances performance by learning the shift matrix.
Optimal deep learning of holomorphic operators between Banach spaces
Ben Adcock (Simon Fraser University), Sebastian Moraga (Simon Fraser University)
Auto EncoderPhysics Related
🎯 What it does: This study addresses the challenge of learning holomorphic operators between Banach spaces and proposes a deep neural network (DNN) architecture that combines arbitrary approximation encoders and decoders, focusing on the learning of holomorphic operators.
Optimal Design for Human Preference Elicitation
Subhojyoti Mukherjee (University of Wisconsin Madison), Branislav Kveton (Adobe Research)
Recommendation SystemOptimizationText
🎯 What it does: An optimal design method is proposed for efficiently selecting a list of questions (including multiple answers) to obtain human preference feedback and learn the corresponding preference model under a given question budget.
Optimal Flow Matching: Learning Straight Trajectories in Just One Step
Nikita Maksimovich Kornilov (Skolkovo Institute of Science and Technology), Alexander Korotin (Skolkovo Institute of Science and Technology)
Image TranslationOptimizationFlow-based ModelRectified FlowImage
🎯 What it does: Developed Optimal Flow Matching (OFM), which learns the optimal transport flow of straight-line trajectories in single-step Flow Matching, allowing for direct acquisition of straight paths without solving ODEs;
Optimal Hypothesis Selection in (Almost) Linear Time
Maryam Aliakbarpour (Rice University), Adam Smith (Boston University)
OptimizationComputational Efficiency
🎯 What it does: This paper proposes two nearly linear time hypothesis selection algorithms, achieving the best error amplification factors α=3 (Algorithm 1) and α=4 (Algorithm 4) under the premise of sample optimality (O(log(n/δ)/ε²), and provides theoretical analysis of error and time complexity.
Optimal Multi-Fidelity Best-Arm Identification
Riccardo Poiani (Politecnico di Milano), Marcello Restelli (Politecnico di Milano)
OptimizationReinforcement Learning
🎯 What it does: This paper proposes a multi-fidelity optimal arm identification method, provides an instance-specific cost lower bound, and designs a gradient ascent algorithm MF-GRAD that can achieve this lower bound with high confidence.
Optimal Multiclass U-Calibration Error and Beyond
Haipeng Luo (University of Southern California), Vatsal Sharan (University of Southern California)
OptimizationReinforcement Learning
🎯 What it does: This paper addresses the online multi-class U-calibration problem and studies the optimal empirical return when facing all finite ranges of correct loss functions; it provides theoretical analysis of the lower and upper bounds of the pseudo U-calibration error, and gives a better upper bound for specific subclasses of loss.
Optimal Parallelization of Boosting
Arthur da Cunha (Aarhus University), Kasper Green Larsen (Aarhus University)
Optimization
🎯 What it does: A new parallel Boosting algorithm is proposed, and an improved lower bound proof is provided, which essentially brings the trade-off between the time complexity of parallel training and the number of training rounds closer to optimal.
Optimal Private and Communication Constraint Distributed Goodness-of-Fit Testing for Discrete Distributions in the Large Sample Regime
Lasse Vuursteen (University of Pennsylvania)
OptimizationSafty and Privacy
🎯 What it does: This paper studies the limit detection boundaries of the goodness-of-fit (uniformity) test for discrete multinomial distributions in a distributed environment under bandwidth constraints and differential privacy constraints in large samples.
Optimal Rates for Vector-Valued Spectral Regularization Learning Algorithms
Dimitri Meunier (Gatsby Computational Neuroscience Unit University College London), Zhu Li (Imperial College London)
🎯 What it does: This study investigates the theoretical properties of vector-valued spectral regularization learning algorithms, focusing on the case of infinite-dimensional output spaces.
Optimal Scalarizations for Sublinear Hypervolume Regret
Qiuyi Zhang (Google DeepMind)
Optimization
🎯 What it does: This paper studies the use of hypervolume scalarization methods in multi-objective optimization and proves that under random weights, this method can converge with a sublinear hypervolume regret rate of O(T^{-1/k}); it also proposes a non-Euclidean analysis algorithm for the multi-objective linear stochastic bandit problem, achieving a hypervolume regret rate of O(dT^{-1/2}+T^{-1/k}); finally, experiments were conducted to validate this method in three types of tasks: synthetic, linear bandit, and black-box optimization.
Optimal Top-Two Method for Best Arm Identification and Fluid Analysis
Agniv Bandyopadhyay (Tata Institute of Fundamental Research), Shubhada Agrawal (Georgia Institute of Technology)
OptimizationTabularOrdinary Differential Equation
🎯 What it does: This paper proposes a new algorithm based on Anchor Top-2 (AT2) for optimal arm identification under fixed confidence, and provides an asymptotic optimality proof regarding its sample complexity.
Optimal Transport-based Labor-free Text Prompt Modeling for Sketch Re-identification
Rui Li (Harbin Institute of Technology), Jinxing Li (Harbin Institute of Technology)
RecognitionRetrievalTransformerPrompt EngineeringImageText
🎯 What it does: This paper proposes an unsupervised optimal transport-based text prompt modeling network (OLTM), which automatically generates text attributes and combines prompt learning with optimal transport to achieve multi-level and multi-granularity alignment between sketches and images, thereby completing the Sketch Re-ID task.
Optimal-state Dynamics Estimation for Physics-based Human Motion Capture from Videos
Cuong Le (Linköping University), Bastian Wandt (Linköping University)
Pose EstimationRecurrent Neural NetworkReinforcement LearningVideoPhysics Related
🎯 What it does: This paper proposes an online human motion capture framework called OSDCap, which combines video pose estimation, differentiable physical simulation, and a learnable Kalman filter to achieve dynamic estimation of joint torques, external forces, and inertial biases.
Optimistic Critic Reconstruction and Constrained Fine-Tuning for General Offline-to-Online RL
Qin-Wen Luo (Nanjing University of Aeronautics and Astronautics), Sheng-Jun Huang (Nanjing University of Aeronautics and Astronautics)
Reinforcement LearningSequential
🎯 What it does: A general offline-to-online reinforcement learning framework is designed, which first reconstructs the critic through optimistic re-evaluation and aligns it with the offline policy, then incorporates constraint fine-tuning to alleviate distribution shift, achieving stable and efficient online fine-tuning from any offline method to SAC, TD3, and PPO.
Optimistic Verifiable Training by Controlling Hardware Nondeterminism
Megha Srivastava (Stanford University), Dan Boneh (Stanford University)
TransformerSupervised Fine-TuningText
🎯 What it does: This paper proposes a verifiable training method that eliminates non-determinism between different GPUs by training at higher precision and recording rounding decisions, allowing auditors to accurately reproduce the training process.
Optimization Algorithm Design via Electric Circuits
Stephen P. Boyd (Stanford University), Jaewook J. Suh (Rice University)
OptimizationTabular
🎯 What it does: This paper proposes a convex optimization algorithm design method based on circuit simulation, mapping the continuous-time dynamics of optimization problems to an inductor-capacitor-resistor (RLC) circuit, and utilizing automatic discretization techniques to generate convergent algorithms.
Optimization Can Learn Johnson Lindenstrauss Embeddings
Nikos Tsikouras (University of Athens), Christos Tzamos (University of Athens)
OptimizationDiffusion modelTabular
🎯 What it does: An optimized deterministic Johnson-Lindenstrauss embedding method is proposed, which directly learns the projection matrix that satisfies the JL guarantee through second-order optimization in the Gaussian sampler space.
Optimized Feature Generation for Tabular Data via LLMs with Decision Tree Reasoning
Jaehyun Nam (Korea Advanced Institute of Science and Technology), Jinwoo Shin (Korea Advanced Institute of Science and Technology)
OptimizationTransformerLarge Language ModelTabularFinance Related
🎯 What it does: A framework called OCTree is proposed for generating column features using LLM and decision tree inference.
Optimizing Automatic Differentiation with Deep Reinforcement Learning
Jamie Lohoff (Peter Grünberg Institute Forschungszentrum Jülich), Emre Neftci (Peter Grünberg Institute Forschungszentrum Jülich)
OptimizationTransformerReinforcement LearningGraphFinance Related
🎯 What it does: Utilizing deep reinforcement learning (a variant of AlphaZero) to find the optimal cross-country elimination order, significantly reducing the number of multiplications while preserving the exact Jacobian;
Optimizing over Multiple Distributions under Generalized Quasar-Convexity Condition
Shihong Ding (Peking University), Cong Fang (Peking University)
OptimizationReinforcement Learning
🎯 What it does: Two types of structured conditions for multi-distribution optimization are proposed—Generalized Quasi-Convexity (GQC) and Generalized Quasi-Convex-Concavity (GQCC). Based on these, two Optimistic Mirror Descent (OMD) algorithms are designed to find approximate global optima or approximate Nash equilibria without needing to know the convexity parameters of each distribution in advance.
Optimizing the coalition gain in Online Auctions with Greedy Structured Bandits
Dorian Baudry (University of Oxford), Clément Calauzènes (Inria)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the problem of how much advertising activity to pre-determine in multi-round second-price auctions for online display advertising using DSP, modeling it as a structured N-armed bandit problem.
Optimus-1: Hybrid Multimodal Memory Empowered Agents Excel in Long-Horizon Tasks
Zaijing Li (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
TransformerLarge Language ModelVision Language ModelVision-Language-Action ModelMultimodalitySequential
🎯 What it does: A composable multimodal agent, Optimust-1, is proposed in the Minecraft environment, utilizing a hybrid multimodal memory module (HDKG+AMEP) to achieve planning and reflection for long-term tasks.
OPUS: Occupancy Prediction Using a Sparse Set
JiaBao Wang, Ming-Ming Cheng (Nankai University)
SegmentationAutonomous DrivingTransformerSupervised Fine-TuningPoint Cloud
🎯 What it does: This paper proposes a method that treats the occupancy prediction task as a direct set prediction problem, and implements parallel regression of occupied locations and semantic categories in 3D space using a Transformer encoder-decoder architecture.
Oracle-Efficient Differentially Private Learning with Public Data
Adam Block (Massachusetts Institute of Technology), Steven Wu
OptimizationSafty and Privacy
🎯 What it does: This paper proposes a general algorithm for achieving differential privacy learning using publicly available unlabeled data, and presents an improved version for convex function classes and binary classification problems;
Oracle-Efficient Reinforcement Learning for Max Value Ensembles
Marcel Hussing (University of Pennsylvania), Jessica Sorrell (Johns Hopkins University)
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: An algorithm named MaxIteration is proposed, which utilizes an existing set of benchmark policies in reinforcement learning with large or infinite state spaces to approximate their value functions through a regression oracle, thereby learning a policy that competes closely with the maximum value-following policy.
Orchid: Flexible and Data-Dependent Convolution for Sequence Modeling
Mahdi Karami (Google Research), Ali Ghodsi (University of Waterloo)
ClassificationOptimizationComputational EfficiencyConvolutional Neural NetworkTransformerSupervised Fine-TuningImageSequential
🎯 What it does: Proposes the Orchid architecture, which utilizes data-adaptive global convolution to replace traditional attention, achieving scalability and efficiency in sequence mixing;
Order-Independence Without Fine Tuning
Reid McIlroy-Young (Harvard University), Cynthia Dwork (Harvard University)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the Set-Based Prompting method, which modifies the attention mask and position encoding during inference to make LLM insensitive to the order of specified subsequences, thereby eliminating the order dependency problem in scenarios such as multiple-choice questions.
Ordered Momentum for Asynchronous SGD
Chang-Wei Shi (Nanjing University), Wu-Jun Li (Nanjing University)
OptimizationReinforcement LearningImage
🎯 What it does: Proposed an Ordered Momentum (OrMo) mechanism for asynchronous SGD and implemented it under the parameter server framework;
Ordering-Based Causal Discovery for Linear and Nonlinear Relations
Zhuopeng Xu (Central South University), Ning Gui (Central South University)
Score-based ModelGraph
🎯 What it does: A unified causal discovery algorithm CaPS is proposed, capable of performing DAG structure learning on data with both linear and nonlinear causal relationships.
OSLO: One-Shot Label-Only Membership Inference Attacks
Yuefeng Peng (University of Massachusetts Amherst), Amir Houmansadr (University of Massachusetts Amherst)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a One-Shot Label-Only Membership Inference Attack (OSLO), which can determine whether a sample belongs to the training set of the target model with just one query, and the target model only returns hard labels.
OT4P: Unlocking Effective Orthogonal Group Path for Permutation Relaxation
Yaming Guo (Jilin University), Tieru Wu (Jilin University)
OptimizationImageGraph
🎯 What it does: We propose OT4P, a temperature-controlled differentiable transformation that maps unconstrained vectors to the orthogonal group, thereby relaxing permutation matrices and enabling gradient-based optimization and stochastic optimization.
OTTER: Effortless Label Distribution Adaptation of Zero-shot Models
Changho Shin (University of Wisconsin Madison), Frederic Sala (University of Wisconsin Madison)
ClassificationDomain AdaptationTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: This paper proposes a method that adjusts the prediction distribution of zero-shot models (such as CLIP and BERT) through Optimal Transport during the inference phase to address the issue of mismatched label distributions in pre-training, while also being compatible with few-shot scenarios and correcting the selection bias of large language models.
Out-of-Distribution Detection with a Single Unconditional Diffusion Model
Alvin Heng (National University of Singapore), Harold Soh (National University of Singapore)
Anomaly DetectionDiffusion modelImage
🎯 What it does: Using a single unconditional diffusion model, the judgment of whether a sample belongs to the distribution (OOD detection) is achieved by analyzing the rate of change and curvature of the diffusion path from the sample to the standard normal distribution.
Out-Of-Distribution Detection with Diversification (Provably)
Haiyun Yao (Tianjin University), Changqing Zhang (Tianjin University)
Anomaly DetectionConvolutional Neural NetworkMixture of ExpertsImage
🎯 What it does: This paper theoretically analyzes and experimentally verifies the impact of auxiliary outlier sample diversity on OOD detection performance, and proposes a diversity-inducing Mixup (diverseMix) method to enhance the model's generalization ability to unknown OOD.
Outlier-Robust Distributionally Robust Optimization via Unbalanced Optimal Transport
Zifan Wang (KTH Royal Institute of Technology), Karl Henrik Johansson
Anomaly DetectionOptimizationComputational EfficiencyTabular
🎯 What it does: This paper proposes a distributionally robust optimization framework based on Unbalanced Optimal Transport (UOT), utilizing a soft penalty UOT distance to construct a fuzzy discrimination set that is tolerant to outliers, and designs a Lagrangian penalty form that can be solved using stochastic subgradients.
Over-parameterized Student Model via Tensor Decomposition Boosted Knowledge Distillation
Yu-Liang Zhan (Renmin University of China), Ze-Feng Gao (Renmin University of China)
Knowledge DistillationTransformerImageText
🎯 What it does: This paper proposes a framework called OPDF for over-parameterization of the student model during the knowledge distillation training phase. It utilizes MPO (Matrix Product Operator) to decompose the weight matrix into high-order tensors and designs a tensor alignment loss based on this to improve the distillation effect.
Overcoming Brittleness in Pareto-Optimal Learning Augmented Algorithms
Alex Elenter (Sorbonne University), Yanni LEFKI
OptimizationReinforcement Learning from Human FeedbackTime SeriesFinance Related
🎯 What it does: A learning-enhanced online algorithm framework based on performance curves (profile) is proposed, and feasibility determination and algorithm construction are implemented for the one-way trading problem; at the same time, an adaptive Pareto-optimal algorithm ADA-PO is designed, which can achieve dominance relations across all sequences while maintaining robustness.
Overcoming Common Flaws in the Evaluation of Selective Classification Systems
Jeremias Traub (German Cancer Research Center), Paul F Jaeger
ClassificationSegmentationImageBenchmark
🎯 What it does: This paper proposes and validates a new multi-threshold evaluation metric AUGRC, aimed at comprehensively assessing the performance of selective classification systems, overcoming the shortcomings of the existing AURC metric.
Overcoming the Sim-to-Real Gap: Leveraging Simulation to Learn to Explore for Real-World RL
Andrew Wagenmaker (University of California), Abhishek Gupta (University of Washington)
Robotic IntelligenceReinforcement Learning
🎯 What it does: Learn a set of exploration strategies in a simulation environment, and then transfer these strategies to the real environment, achieving efficient RL learning through random exploration and least squares regression.
Overfitting Behaviour of Gaussian Kernel Ridgeless Regression: Varying Bandwidth or Dimensionality
Marko Medvedev (University of Chicago), Nathan Srebro (TTI Chicago)
🎯 What it does: This study investigates the performance of Gaussian kernel ridge regression in the context of ridgeless regression, providing a theoretical analysis of the test error of the minimum norm interpolation solution under two settings: fixed dimensions with adjustable kernel width, and dimensions growing with the sample size.
OW-VISCapTor: Abstractors for Open-World Video Instance Segmentation and Captioning
Anwesa Choudhuri (University of Illinois at Urbana-Champaign), Alex Schwing
Object TrackingSegmentationGenerationTransformerLarge Language ModelContrastive LearningVideoText
🎯 What it does: We propose OW-VISCapTor, which jointly accomplishes open-world video instance segmentation, tracking, and object-centered subtitle generation.
OwMatch: Conditional Self-Labeling with Consistency for Open-World Semi-Supervised Learning
Shengjie Niu (Hong Kong Polytechnic University), Chao Wang (Southern University of Science and Technology)
ClassificationContrastive LearningImage
🎯 What it does: An open-world semi-supervised learning framework called OwMatch is proposed, which utilizes conditional self-labeling and hierarchical thresholds to achieve simultaneous classification and clustering of known and unknown classes.
OxonFair: A Flexible Toolkit for Algorithmic Fairness
Eoin D. Delaney, Chris Russell (University of Oxford)
ClassificationOptimizationTextTabular
🎯 What it does: Developed OxonFair, a scalable fairness toolkit for binary classification tasks, supporting three types of data: tabular, NLP, and computer vision;
P$^2$C$^2$Net: PDE-Preserved Coarse Correction Network for efficient prediction of spatiotemporal dynamics
Qi Wang (Renmin University of China), Yang Liu (University of Chinese Academy of Sciences)
Time SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: Proposes the P2C Net model, which efficiently predicts spatiotemporal PDE dynamics on coarse grids and with large time steps, supporting training with very little data;
PAC-Bayes-Chernoff bounds for unbounded losses
Ioar Casado (Basque Center for Applied Mathematics), Andres R Masegosa
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a new PAC-Bayes-Chernoff prediction that provides oracle-form generalization bounds for unbounded loss functions and allows for precise optimization of the free parameter λ without the need for grid search.
PACE: Marrying generalization in PArameter-efficient fine-tuning with Consistency rEgularization
Yao Ni (Australian National University), Piotr Koniusz (Data61 CSIRO)
Domain AdaptationTransformerSupervised Fine-TuningImageText
🎯 What it does: The PACE method is proposed, combining Parameter-Efficient Fine-Tuning (PEFT) with consistency regularization. By adding multiplicative noise to the adapter features and requiring the outputs for the same input to remain consistent under different noise conditions, it reduces the gradient norm and aligns the fine-tuned model with the pre-trained model.
PACE: Pacing Operator Learning to Accurate Optical Field Simulation for Complicated Photonic Devices
Hanqing Zhu (University of Texas at Austin), David Z. Pan (University of Texas at Austin)
OptimizationComputational EfficiencyKnowledge DistillationTabularPhysics Related
🎯 What it does: This paper proposes a new cross-axis factorized Pace operator and a two-stage learning framework to accelerate and improve the simulation accuracy of optical fields in complex photonic devices.
PaCE: Parsimonious Concept Engineering for Large Language Models
Jinqi Luo (University of Pennsylvania), Rene Vidal
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A sparse coding-based activation engineering framework, PaCE, is proposed for aligning large language models without adjusting parameters.
PaDeLLM-NER: Parallel Decoding in Large Language Models for Named Entity Recognition
Jinghui Lu (ByteDance), Can Huang (ByteDance)
RecognitionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes and implements PaDeLLM-NER, which transforms the instruction tuning task to enable LLMs to generate all label-entity pairs in parallel during a single inference, significantly reducing generation latency.
PageRank Bandits for Link Prediction
Yikun Ban (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
Recommendation SystemGraph Neural NetworkGraph
🎯 What it does: A link prediction framework called PRB is proposed, which integrates the contextual bandit algorithm with PageRank, capable of balancing exploitation and exploration in both online and offline environments while utilizing graph structural information.
PaGoDA: Progressive Growing of a One-Step Generator from a Low-Resolution Diffusion Teacher
Dongjun Kim (Stanford University), Stefano Ermon (Stanford University)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationDiffusion modelGenerative Adversarial NetworkImageText
🎯 What it does: This paper proposes a three-stage training pipeline called PaGoDA, which first pre-trains a diffusion model at low resolution, then uses DDIM inversion for first-order generator distillation, and finally enhances the model to high resolution through a progressive upsampling network, significantly reducing training and inference costs.
Panacea: Pareto Alignment via Preference Adaptation for LLMs
Yifan Zhong (Peking University), Yaodong Yang (Peking University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A method called Panacea is proposed, utilizing Pareto set learning to achieve online alignment of large language models under multidimensional human preferences.
Pandora's Box: Towards Building Universal Attackers against Real-World Large Vision-Language Models
Daizong Liu (Peking University), Lichao Sun (Lehigh University)
OptimizationAdversarial AttackTransformerVision Language ModelImageTextMultimodality
🎯 What it does: In the context of large-scale visual language models (LVLM) in the real world, which can only be queried in a black-box scenario, a position-fixed, task-agnostic universal adversarial patch is constructed to induce the model to generate target text specified by the attacker under any input and task prompt.
PANORAMIA: Privacy Auditing of Machine Learning Models without Retraining
Mishaal Kazmi (University of British Columbia), Mathias Lécuyer (University of British Columbia)
GenerationSafty and PrivacyTransformerLarge Language ModelGenerative Adversarial NetworkImageText
🎯 What it does: A privacy auditing framework named PANORAMIA is proposed, which uses generated data to replace non-member samples for evaluating the privacy leakage of machine learning models;
Paralinguistics-Aware Speech-Empowered Large Language Models for Natural Conversation
Heeseung Kim (Seoul National University), Kang Min Yoo (NAVER)
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningMultimodalityAudio
🎯 What it does: A unified speech-text large language model (USDM) has been developed, achieving end-to-end natural speech dialogue generation, capable of producing speech responses with natural prosody without using explicit ASR/TTS.
Parallel Backpropagation for Shared-Feature Visualization
Alexander Lappe (Hertie Institute, University Clinics Tübingen), Rufin Vogels (KU Leuven)
Explainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a parallel backpropagation method based on deep learning to visualize the visual features shared by neurons activated by stimuli beyond categories in the higher visual cortex.
ParallelEdits: Efficient Multi-Aspect Text-Driven Image Editing with Attention Grouping
Mingzhen Huang (University at Buffalo), Siwei Lyu (University at Buffalo)
Image TranslationGenerationComputational EfficiencyTransformerDiffusion modelImageBenchmark
🎯 What it does: This paper proposes a multi-faceted text-driven image editing task and develops the ParallelEdits method to simultaneously edit multiple objects, attributes, or relationships in an image in one go.