NeurIPS 2024 Papers — Page 3
Conference on Neural Information Processing Systems · 4035 papers
Addressing Asynchronicity in Clinical Multimodal Fusion via Individualized Chest X-ray Generation
Wenfang Yao (Hong Kong Polytechnic University), Jing Qin (Hong Kong Baptist University)
GenerationData SynthesisAnomaly DetectionTransformerDiffusion modelAuto EncoderContrastive LearningMultimodalityBiomedical DataElectronic Health Records
🎯 What it does: This paper proposes a method called DDL-CXR based on the Latent Diffusion Model (LDM), which dynamically generates personalized chest X-ray (CXR) latent representations synchronized with the prediction time to address the asynchrony issue of multimodal clinical data (EHR and CXR) and integrates the generated latent CXR with historical data for clinical predictions.
Addressing Bias in Online Selection with Limited Budget of Comparisons
Ziyad Benomar (ENSAE), Vianney Perchet (CREST)
🎯 What it does: This study investigates the multi-color secretary problem under a limited comparison budget, proposing a dynamic threshold algorithm and providing optimal memoryless algorithms and their threshold calculation methods for two scenarios.
Addressing Hidden Confounding with Heterogeneous Observational Datasets for Recommendation
Yanghao Xiao (University of Chinese Academy of Sciences), Wensheng Zhang (Guangzhou University)
Recommendation SystemMeta LearningTabular
🎯 What it does: A meta-learning based debiasing recommendation method called MetaDebias is proposed, which uses heterogeneous observational data to address selection bias caused by hidden confounding.
Addressing Spatial-Temporal Heterogeneity: General Mixed Time Series Analysis via Latent Continuity Recovery and Alignment
Jiawei Chen (Zhejiang University), Chunhui Zhao (Zhejiang University)
Anomaly DetectionRepresentation LearningTransformerTime Series
🎯 What it does: This paper proposes the MiTSformer framework for unified modeling of mixed temporal data (including continuous and discrete variables) to achieve general representation learning for multiple tasks.
Addressing Spectral Bias of Deep Neural Networks by Multi-Grade Deep Learning
Ronglong Fang (Old Dominion University), Yuesheng Xu (Old Dominion University)
Spiking Neural NetworkImageTabular
🎯 What it does: This paper proposes a method to decompose high-frequency functions into a combination of multiple low-frequency sub-functions through Multi-Grade Deep Learning (MGDL), thereby alleviating the spectral bias problem of DNNs.
AdjointDEIS: Efficient Gradients for Diffusion Models
Zander W. Blasingame (Clarkson University), Chen Liu (Clarkson University)
GenerationData SynthesisDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: The AdjointDEIS method is proposed for efficiently solving gradients in diffusion models through continuous adjoint equations, enabling training-free guided generation.
Adjust Pearson's $r$ to Measure Arbitrary Monotone Dependence
Xinbo Ai (Beijing University of Posts and Telecommunications)
Tabular
🎯 What it does: A rearranged correlation coefficient (r#) is proposed, which quantifies any monotonic dependence by improving the Pearson correlation coefficient, and its theoretical foundation and calculation method are provided.
ADOPT: Modified Adam Can Converge with Any $\beta_2$ with the Optimal Rate
Shohei Taniguchi (University of Tokyo), Yutaka Matsuo (University of Tokyo)
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes a new adaptive gradient optimizer called ADOPT, which theoretically achieves a convergence rate of O(1/√T) without requiring a specific choice of β2 and without relying on the bounded gradient noise assumption. It outperforms Adam and its variants across various tasks.
AdvAD: Exploring Non-Parametric Diffusion for Imperceptible Adversarial Attacks
Jin Li (Sun Yat-Sen University), Xiangui Kang (Sun Yat-Sen University)
Adversarial AttackConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper proposes a non-parametric diffusion process-based adversarial attack framework, AdvAD, and its extreme version, AdvAD-X, to generate imperceptible adversarial samples.
Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models
Yicheng Xu (Institute of Science), Manabu Okumura
ClassificationDomain AdaptationTransformerVision Language ModelImage
🎯 What it does: A cross-domain task-agnostic incremental learning (X-TAIL) framework is proposed, along with a regression-based analytical incremental learning (RAIL) method, to achieve continuous learning across multiple domains while maintaining the zero-shot capability of VLM.
Advancing Fine-Grained Classification by Structure and Subject Preserving Augmentation
Eyal Michaeli (Reichman University), Ohad Fried (Reichman University)
ClassificationGenerationLarge Language ModelDiffusion modelImage
🎯 What it does: A generative data augmentation method SaSPA based on edge maps and thematic representations is proposed for fine-grained visual classification.
Advancing Open-Set Domain Generalization Using Evidential Bi-Level Hardest Domain Scheduler
Kunyu Peng (Karlsruhe Institute of Technology), Rainer Stiefelhagen (University of Stuttgart)
ClassificationDomain AdaptationMeta LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: An adaptive domain scheduling method for Open Set Domain Generalization (OSDG) is proposed—Evidential Bi-Level Hardest Domain Scheduler (EBiL-HaDS), which enhances the model's generalization and recognition capabilities for unknown domains and unknown categories by dynamically selecting the hardest domains.
Advancing Spiking Neural Networks for Sequential Modeling with Central Pattern Generators
Changze Lv (Fudan University), Dongsheng Li (Microsoft Research Asia)
ClassificationRecognitionOptimizationSpiking Neural NetworkTransformerImageTextTime Series
🎯 What it does: A location information encoding method based on Central Pattern Generators (CPG) is proposed in spiking neural networks (SNN), and its effectiveness is validated in tasks such as time series prediction, text classification, and image classification.
Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees
Sijia Chen (Nanjing University), Lijun Zhang (Nanjing University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By constructing a stepwise preference dataset based on failure exploration and applying Direct Preference Optimization (DPO) after Supervised Fine-Tuning (SFT), the multi-step reasoning ability of tool-enhanced large language models has been improved.
Advancing Training Efficiency of Deep Spiking Neural Networks through Rate-based Backpropagation
Chengting Yu (Zhejiang University), Aili Wang (Zhejiang University)
Spiking Neural NetworkImage
🎯 What it does: A rate coding-based backpropagation method is proposed to simplify the time dependency of SNNs and reduce the computational and memory costs of BPTT.
Advection Augmented Convolutional Neural Networks
Niloufar Zakariaei (University of British Columbia), Moshe Eliasof (University of Cambridge)
RestorationGenerationOptimizationConvolutional Neural NetworkVideoTime SeriesPhysics Related
🎯 What it does: This paper proposes an architecture that integrates the advection-diffusion-reaction (ADR) process into convolutional neural networks for efficient prediction of spatio-temporal sequences.
Adversarial Environment Design via Regret-Guided Diffusion Models
Hojun Chung (Seoul National University), Songhwai Oh (Seoul National University)
Reinforcement LearningDiffusion model
🎯 What it does: This paper proposes a penalty-guided diffusion model (ADD) to generate adversarial training environments, thereby enhancing the generalization performance of deep reinforcement learning.
Adversarial Moment-Matching Distillation of Large Language Models
Chen Jia (SI-TECH Information Technology)
Knowledge DistillationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a knowledge distillation method based on adversarial action value moment matching, utilizing a reinforcement learning framework to view distillation as imitation learning. The goal is to minimize the difference in action value moments between the teacher and the student, combining both on-policy and off-policy perspectives.
Adversarial Representation Engineering: A General Model Editing Framework for Large Language Models
Yihao Zhang (Peking University), Meng Sun (Peking University)
Representation LearningAdversarial AttackTransformerLarge Language ModelGenerative Adversarial NetworkText
🎯 What it does: A framework based on Adversarial Representation Engineering (ARE) is proposed for precise editing of concepts in large language models without affecting the original model's performance.
Adversarial Schrödinger Bridge Matching
Nikita Gushchin (Skolkovo Institute of Science and Technology), Alexander Korotin (Skolkovo Institute of Science and Technology)
Image TranslationData SynthesisDomain AdaptationGenerative Adversarial NetworkImageStochastic Differential Equation
🎯 What it does: A discrete-time iterative Markov fitting (D-IMF) method is proposed to efficiently solve the Schrödinger bridge problem, achieving unsupervised domain transformation from distribution p0 to p1.
Adversarially Robust Decision Transformer
Xiaohang Tang (University College London), Ilija Bogunovic (University College London)
TransformerReinforcement LearningSequential
🎯 What it does: A method called Adversarially Robust Decision Transformer (ARDT) is designed and implemented, which enhances the robustness of agents in adversarial environments by using Expectile Regression to label worst-case returns of trajectories in offline reinforcement learning and utilizing these labeled returns during the training phase of the Decision Transformer.
Adversarially Robust Dense-Sparse Tradeoffs via Heavy-Hitters
David Woodruff, Samson Zhou (Texas A&M University)
OptimizationSafty and PrivacyAdversarial AttackTime Series
🎯 What it does: An improved algorithm for L_p estimation and heavy-hitter detection is proposed in adversarial stream models, implemented using sparse-dense trade-off techniques.
Adversarially Robust Multi-task Representation Learning
Austin Watkins (Johns Hopkins University), Raman Arora (Johns Hopkins University)
Representation LearningAdversarial Attack
🎯 What it does: This paper theoretically studies adversarially robust transfer learning under the Multi-Task Representation Learning (MTRL) framework, providing an upper bound on adversarial excess transfer risk under Lipschitz loss and smooth non-negative loss, and proving that learning robust representations for multiple tasks can reduce the impact of adversarial attacks.
Adversarially Trained Weighted Actor-Critic for Safe Offline Reinforcement Learning
Honghao Wei (Washington State University), Xin Liu (ShanghaiTech University)
Safty and PrivacyReinforcement LearningTabular
🎯 What it does: A weighted safe actor-critic algorithm (WSAC) for offline safe reinforcement learning is proposed, which achieves safe robust policy improvement using Stackelberg games and relative inertia penalties.
AED: Adaptable Error Detection for Few-shot Imitation Policy
Jia-Fong Yeh (National Taiwan University), Winston H. Hsu (National Taiwan University)
Anomaly DetectionRobotic IntelligenceRecurrent Neural NetworkReinforcement LearningVideoTime SeriesBenchmark
🎯 What it does: An Adaptive Error Detection (AED) task is proposed, along with a corresponding cross-domain benchmark, and an algorithm named Pattern Observer (PrObe) is developed to monitor behavioral errors in Few-Shot Imitation (FSI) strategies online.
Agent Planning with World Knowledge Model
Shuofei Qiao (Zhejiang University), Huajun Chen (National University of Singapore)
TransformerLarge Language ModelAgentic AIContrastive LearningText
🎯 What it does: A parameterizable World Knowledge Model (WKM) is proposed, which synthesizes task knowledge and state knowledge from expert trajectories and experiential exploration trajectories to assist large language models in interactive planning.
AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases
Zhaorun Chen (University of Chicago), Bo Li (University of Chicago)
Autonomous DrivingOptimizationAdversarial AttackTransformerLarge Language ModelAgentic AITextSequentialBiomedical DataElectronic Health RecordsRetrieval-Augmented Generation
🎯 What it does: This paper proposes a backdoor attack method called AGENTPOISON for large language model (LLM) agents using retrieval-augmented generation (RAG). By injecting a small number of malicious examples into the agent's long-term memory or knowledge base, the attack can be triggered when specific keywords are retrieved, inducing the agent to perform malicious actions.
Aggregate-and-Adapt Natural Language Prompts for Downstream Generalization of CLIP
Chen Huang (Apple), Joshua M. Susskind (Apple)
ClassificationRetrievalKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageText
🎯 What it does: A CLIP prompt learning method that aggregates and adapts to natural language prompts is proposed, enhancing CLIP's multi-task downstream generalization ability through text knowledge distillation.
Aggregating Quantitative Relative Judgments: From Social Choice to Ranking Prediction
Yixuan Even Xu (Carnegie Mellon University), Vincent Conitzer (Carnegie Mellon University)
Tabular
🎯 What it does: Proposes and studies the Quantitative Relative Judgment Aggregation (QRJA) model, which serves as a bridge between social choice and ranking prediction;
AGILE: A Novel Reinforcement Learning Framework of LLM Agents
Peiyuan Feng (ByteDance Research), Hang Li (ByteDance Research)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: A unified reinforcement learning framework called AGILE is proposed, allowing large language model agents to complete complex dialogue tasks through memory, tool invocation, reflection, and actively seeking human advice.
AHA: Human-Assisted Out-of-Distribution Generalization and Detection
Haoyue Bai (University of Wisconsin Madison), Robert D Nowak
Domain AdaptationAnomaly DetectionImage
🎯 What it does: Proposes the AHA framework, which enhances the model's OOD generalization and detection capabilities using a small amount of manual labeling in the maximum discernment region (where the densities of covariate OOD and semantic OOD are roughly equal).
AID: Attention Interpolation of Text-to-Image Diffusion
Qiyuan He (National University of Singapore), Angela Yao (National University of Singapore)
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelImageText
🎯 What it does: A training-independent attention interpolation method AID/PAID is proposed for smooth and consistent interpolation of different text conditions in text-to-image diffusion models.
AirSketch: Generative Motion to Sketch
Hui Xian Grace Lim (University of Central Florida), Ser-Nam Lim (University of Central Florida)
RestorationGenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: This paper presents AirSketch, which utilizes a controllable diffusion model to recover clean, coherent, and user-intent-compliant hand-drawn sketches from extremely noisy trajectory images generated by hand tracking, without the need for markers or expensive hardware.
AlchemistCoder: Harmonizing and Eliciting Code Capability by Hindsight Tuning on Multi-source Data
Zifan Song (Tongji University), Cairong Zhao (Tongji University)
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Developed the AlchemistCoder series of code LLMs, significantly enhancing code generation and reasoning capabilities through multi-source data, AlchemistPrompt, and fine-tuning on code understanding tasks.
Algebraic Positional Encodings
Konstantinos Kogkalidis (Aalto University), Vikas Garg (YaiYai Ltd)
ImageSequential
🎯 What it does: A general positional encoding framework based on group theory, Algebraic Positional Encodings (APE), is proposed, which maps structures such as sequences, trees, and two-dimensional grids to orthogonal matrix subgroups, preserving and utilizing the algebraic properties of the original structures.
Algorithmic Capabilities of Random Transformers
Ziqian Zhong (Massachusetts Institute of Technology), Jacob Andreas (Massachusetts Institute of Technology)
GenerationRetrievalTransformerText
🎯 What it does: This paper studies whether a randomly initialized Transformer that only trains the embedding layer (i.e., freezing the internal parameters) can perform algorithmic tasks such as arithmetic, associative retrieval, and bracket matching, and explores its potential in memory and natural language generation.
Algorithmic Collective Action in Recommender Systems: Promoting Songs by Reordering Playlists
Joachim Baumann (University of Zurich), Celestine Mendler-Dünner (Max-Planck Institute for Intelligent Systems)
Recommendation SystemTransformerSequentialAudio
🎯 What it does: This paper studies the feasibility of enhancing the visibility of minority artists in a Transformer-based music recommendation system through algorithmic collective action. It proposes two lightweight, authenticity-constrained song insertion strategies (InClust, DirLoF) and experimentally validates their effectiveness using large industry-level models.
Algorithmic progress in language models
Anson Ho (Epoch), Jaime Sevilla (Epoch)
TransformerLarge Language ModelText
🎯 What it does: Constructed and analyzed evaluation data for over 200 language models on WikiText and Penn Treebank, quantifying the relative contributions of advancements in pre-training algorithms and computational scale to performance improvements from 2012 to 2023;
ALI-Agent: Assessing LLMs' Alignment with Human Values via Agent-based Evaluation
Jingnan Zheng (National University of Singapore), Tat-Seng Chua (Singapore Management University)
TransformerLarge Language ModelSupervised Fine-TuningAgentic AITextChain-of-Thought
🎯 What it does: The ALI-Agent framework is designed to automatically generate, evaluate, and iteratively improve alignment tests for human values through LLM agents.
Alias-Free Mamba Neural Operator
Jianwei Zheng (Zhejiang University of Technology), Xiaoqin Zhang (Zhejiang University of Technology)
BenchmarkPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper proposes a neural operator MambaNO that integrates the Mamba state space model with convolutional integration for efficiently approximating the analytical operator of PDEs while maintaining alias-free properties, with a time complexity of O(N).
Aligner-Encoders: Self-Attention Transformers Can Be Self-Transducers
Adam Stooke (Google), Pedro J Moreno Mengibar
RecognitionOptimizationComputational EfficiencyTransformerAudio
🎯 What it does: Designed the Aligner-Encoder model, allowing the Transformer encoder to achieve audio-to-text alignment during the forward pass, using only frame-level cross-entropy during training, without the need for dynamic programming or blank tokens;
Aligner: Efficient Alignment by Learning to Correct
Jiaming Ji (Peking University), Yaodong Yang (Peking University)
Computational EfficiencyReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper presents Aligner, a lightweight and pluggable alignment module that corrects the outputs of any large language model by learning the correction residuals between good and bad answers in preference data, thereby achieving model alignment.
Aligning Audio-Visual Joint Representations with an Agentic Workflow
Shentong Mo (Carnegie Mellon University), Yibing Song (Alibaba Group)
ClassificationSegmentationData-Centric LearningTransformerLarge Language ModelAgentic AIVision Language ModelVideoMultimodalityAudio
🎯 What it does: This paper proposes an LLM-based Agent workflow (AVAgent) that converts audio and video into text descriptions separately, utilizes LLM for decision-making, and executes a series of audio editing actions (such as denoising, temporal correction, pitch/tempo/volume adjustment, etc.). It then uses VLM to evaluate the consistency between audio and video, iteratively achieving precise alignment of audio and video, thereby enhancing the quality of audio-visual joint representation.
Aligning Diffusion Behaviors with Q-functions for Efficient Continuous Control
Huayu Chen (Tsinghua University), Jun Zhu (Tsinghua University)
OptimizationRobotic IntelligenceReinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: This paper proposes a two-stage method called Efficient Diffusion Alignment (EDA), which breaks down offline reinforcement learning tasks into behavior pre-training and policy alignment. By pre-training a diffusion behavior model and fine-tuning it with a Q-function, efficient continuous control is achieved.
Aligning Diffusion Models by Optimizing Human Utility
Shufan Li (University of California, Los Angeles), Kazuki Kozuka (Panasonic AI Research)
GenerationOptimizationReinforcement Learning from Human FeedbackDiffusion modelImageText
🎯 What it does: Proposes the Diffusion-KTO method, which aligns text-to-image diffusion models by optimizing human utility maximization;
Aligning Embeddings and Geometric Random Graphs: Informational Results and Computational Approaches for the Procrustes-Wasserstein Problem
Mathieu Even (Paris Sciences et Lettres University), Laurent Massoulié (Paris Sciences et Lettres University)
OptimizationGaussian SplattingPoint Cloud
🎯 What it does: This paper studies the Procrustes-Wasserstein alignment problem, providing information-theoretic limits in both high-dimensional and low-dimensional settings, and proposes a Ping-Pong alternating algorithm based on Frank-Wolfe convex relaxation to simultaneously estimate orthogonal transformations and permutations.
Aligning Individual and Collective Objectives in Multi-Agent Cooperation
Yang Li (University of Manchester), Wei Pan (University of Manchester)
OptimizationReinforcement Learning
🎯 What it does: By modeling the mixed-motivation cooperation problem as a differentiable game, the Altruistic Gradient Adjustment (AgA) algorithm is proposed to achieve gradient alignment between individual and collective goals in multi-agent systems, thereby promoting cooperative behavior.
Aligning Large Language Models with Representation Editing: A Control Perspective
Lingkai Kong (Georgia Tech), Chao Zhang (Georgia Tech)
OptimizationRepresentation LearningReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes a framework for aligning large language models through dynamic representation editing during the inference phase;
Aligning LLM Agents by Learning Latent Preference from User Edits
Ge Gao (Cornell University), Dipendra Misra (Microsoft Research)
Recommendation SystemTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This paper proposes an interactive learning framework called PRELUDE, based on user edits, to learn and infer users' implicit preference descriptions, thereby guiding LLM agents to generate text that better meets user needs.
Aligning Model Properties via Conformal Risk Control
William Overman (Stanford University), Mohsen Bayati (Stanford University)
Tabular
🎯 What it does: Post-process the pre-trained model by converting property testing queries into loss functions, and construct prediction intervals using conformal risk control, thereby ensuring that there is at least one function approximately satisfying the specified properties (such as univariate monotonicity and concavity) within the interval at a given confidence level.
Aligning Target-Aware Molecule Diffusion Models with Exact Energy Optimization
Siyi Gu (Stanford University), Stefano Ermon (Stanford University)
OptimizationDrug DiscoveryDiffusion modelContrastive LearningGraph
🎯 What it does: The ALIDIFF framework is proposed to align pre-trained target-aware molecular diffusion models to enhance the binding energy and structural rationality between ligands and targets.
Aligning to Thousands of Preferences via System Message Generalization
Seongyun Lee (Korea Advanced Institute of Science and Technology), Minjoon Seo (Korea Advanced Institute of Science and Technology)
Recommendation SystemOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Trained and evaluated a 7B language model (JANUS) for personalized alignment based on system messages, and constructed a training set called MULTIFACETED COLLECTION containing 197k multidimensional system messages.
Aligning Vision Models with Human Aesthetics in Retrieval: Benchmarks and Algorithms
Miaosen Zhang (Southeast University), Baining Guo (Southeast University)
RetrievalRecommendation SystemReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringImage
🎯 What it does: By aligning the visual model in the retrieval system, it enhances its satisfaction of human aesthetic preferences.
Alignment at Pre-training! Towards Native Alignment for Arabic LLMs
Juhao Liang (Shenzhen Research Institute of Big Data), Jinchao Xu (King Abdullah University of Science and Technology)
TransformerLarge Language ModelText
🎯 What it does: Proposed and implemented a 'Native Alignment' method during the pre-training phase, focusing on Arabic LLM, and released two open-source models: LLaMA3-Tamed-8B and LLaMA3-Tamed-70B.
Alignment for Honesty
Yuqing Yang (Fudan University), Pengfei Liu (Shanghai Jiao Tong University)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A framework for aligning the honesty of large language models is proposed, encouraging models to proactively refuse to answer when they do not know the answer.
All-in-One Image Coding for Joint Human-Machine Vision with Multi-Path Aggregation
Xu Zhang (Nanjing University), Zhan Ma (Nanjing University)
SegmentationCompressionGenerative Adversarial NetworkImage
🎯 What it does: A multi-path aggregation (MPA) model is proposed to achieve an all-in-one unified architecture for human-machine vision joint image coding.
Alleviate Anchor-Shift: Explore Blind Spots with Cross-View Reconstruction for Incomplete Multi-View Clustering
Suyuan Liu (National University of Defense Technology), Kunlun He (Chinese PLA General Hospital)
OptimizationImage
🎯 What it does: This study addresses the anchor point shift problem in incomplete multi-view clustering and proposes the AIMC-CVR method for cross-view reconstruction to enhance clustering performance.
Alleviating Distortion in Image Generation via Multi-Resolution Diffusion Models and Time-Dependent Layer Normalization
Qihao Liu (ByteDance), Liang-Chieh Chen (ByteDance)
GenerationData SynthesisTransformerDiffusion modelImage
🎯 What it does: This paper proposes the DiMR model, which utilizes a multi-resolution network for feature cascading in diffusion models and introduces time-dependent layer normalization (TD-LN) to improve the injection of temporal information.
Alleviating Hallucinations in Large Vision-Language Models through Hallucination-Induced Optimization
Xinyu Lyu (Southwestern University of Finance and Economics), Jingkuan Song (Shenzhen Institute for Advanced Study, University of Electronic Science and Technology of China)
RetrievalOptimizationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: A strategy called Hallucination-Induced Optimization (HIO) is proposed to enhance contrastive decoding by inducing more hallucinations, significantly reducing the hallucination output of large visual language models.
Almost Free: Self-concordance in Natural Exponential Families and an Application to Bandits
Shuai Liu (University of Alberta), Csaba Szepesvari (University of Alberta)
OptimizationReinforcement Learning
🎯 What it does: This study investigates the self-conformality of one-parameter natural exponential families and applies this property to generalized linear multi-armed bandits with sub-exponential/sub-Gaussian distributions, providing a second-order regret bound without exponential upper bounds.
Almost Minimax Optimal Best Arm Identification in Piecewise Stationary Linear Bandits
Yunlong Hou (National University of Singapore), Zixin Zhong (Hong Kong University of Science and Technology)
Reinforcement LearningTime Series
🎯 What it does: A piecewise stationary linear bandit (PSLB) model is proposed, and the PS ε BAI+ algorithm is designed to identify the ε-optimal arm under a given confidence level δ.
Almost Surely Asymptotically Constant Graph Neural Networks
Sam Adam-Day (University of Oxford), Ben Finkelshtein (University of Oxford)
ClassificationGraph Neural NetworkGraph
🎯 What it does: The convergence of the output of real-valued GNN classifiers with respect to graph size is studied under random graph models, proving that most GNNs ultimately output a constant.
Almost-Linear RNNs Yield Highly Interpretable Symbolic Codes in Dynamical Systems Reconstruction
Manuel Brenner (Heidelberg University), Daniel Durstewitz (Heidelberg University)
Explainability and InterpretabilityRecurrent Neural NetworkTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: Proposes and trains an Almost Linear Recursive Neural Network (AL-RNN) that automatically learns the simplest piecewise linear representation from time series data and generates interpretable symbolic encodings.
AlphaMath Almost Zero: Process Supervision without Process
Guoxin Chen (Tongyi Lab), Kai Fan (Tongyi Lab)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: AlphaMath automatically generates high-quality mathematical reasoning processes by combining a pre-trained LLM with Monte Carlo Tree Search (MCTS) and a lightweight value model, without the need for manual or GPT-4 process annotations.
AlphaPruning: Using Heavy-Tailed Self Regularization Theory for Improved Layer-wise Pruning of Large Language Models
Haiquan Lu (Nankai University), Yaoqing Yang (University of California)
TransformerLarge Language ModelText
🎯 What it does: This paper proposes a hierarchical sparsification method called AlphaPruning based on the theory of reparameterization self-regularization (HT-SR), using the ESD shape index (PL Alpha Hill) to allocate the hierarchical sparsity ratio of LLM layers, further enhancing the sparsity rate while maintaining performance.
AlphaTablets: A Generic Plane Representation for 3D Planar Reconstruction from Monocular Videos
Yuze He (Tsinghua University), Yong-jin Liu
SegmentationDepth EstimationNeural Radiance FieldVideo
🎯 What it does: This paper proposes AlphaTablets, which expresses three-dimensional planes using rectangular shapes with alpha channels, and achieves end-to-end 3D plane reconstruction from monocular video based on differentiable rasterization.
ALPINE: Unveiling The Planning Capability of Autoregressive Learning in Language Models
Siwei Wang (Microsoft Research), Wei Chen (Microsoft Research)
TransformerGraph
🎯 What it does: This study investigates the performance of Transformer in path planning tasks and explains its expressive power and learning dynamics from both theoretical and experimental perspectives.
ALPS: Improved Optimization for Highly Sparse One-Shot Pruning for Large Language Models
Xiang Meng (Massachusetts Institute of Technology), Rahul Mazumder (Massachusetts Institute of Technology)
OptimizationLarge Language ModelText
🎯 What it does: An optimized single-shot sparse pruning framework called ALPS is proposed, which uses ADMM to solve non-convex ℓ0 constraints, combined with dynamic penalty parameter updates and PCG post-processing, achieving high-quality one-click pruning for large-scale LLMs.
AlterMOMA: Fusion Redundancy Pruning for Camera-LiDAR Fusion Models with Alternative Modality Masking
Shiqi Sun (Northwestern Polytechnical University), Ying Zhang (Didi Chuxing)
Object DetectionSegmentationAutonomous DrivingMultimodalityPoint Cloud
🎯 What it does: A pruning framework for camera-lidar fusion models is proposed—AlterMOMA, which identifies and removes redundant parameters using an alternating masking approach.
AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers
Jake Grigsby (University of Texas at Austin), Yuke Zhu (University of Texas at Austin)
Meta LearningTransformerReinforcement LearningSequential
🎯 What it does: Using the Transformer sequence model, the updates of the actor and critic are transformed from reward-based regression to a scale-invariant classification loss, thereby addressing the issue of imbalanced reward scales across different tasks in multi-task reinforcement learning, enhancing performance in multi-task, meta-learning, and long-term memory environments.
Amnesia as a Catalyst for Enhancing Black Box Pixel Attacks in Image Classification and Object Detection
Dongsu Song (Korea Aerospace University), Jay Hoon Jung (Korea Aerospace University)
Object DetectionAdversarial AttackReinforcement LearningImage
🎯 What it does: A reinforcement learning-based pixel attack method called RFPAR, which utilizes memory and forgetting mechanisms, is proposed to achieve black-box attacks with minimal pixel perturbation in image classification and object detection.
AmoebaLLM: Constructing Any-Shape Large Language Models for Efficient and Instant Deployment
Yonggan Fu (Georgia Institute of Technology), Yingyan Celine Lin (Georgia Institute of Technology)
CompressionComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The AmoebaLLM framework is proposed, which can instantly extract sub-networks of any depth/width combination after a one-time fine-tuning, allowing large language models to be flexibly and efficiently deployed across various platforms and application scenarios.
AMOR: A Recipe for Building Adaptable Modular Knowledge Agents Through Process Feedback
Jian Guan (Tsinghua University), Minlie Huang (Ant Group)
Domain AdaptationExplainability and InterpretabilityComputational EfficiencyTransformerLarge Language ModelReinforcement LearningMixture of ExpertsTextBiomedical Data
🎯 What it does: This paper proposes an adaptable modular knowledge agent (AMOR) based on finite state machines (FSM), capable of reasoning, retrieving, and answering from knowledge bases in different domains through process feedback.
Amortized Active Causal Induction with Deep Reinforcement Learning
Yashas Annadani (Technical University of Munich), Adam Foster
OptimizationTransformerReinforcement LearningGraph
🎯 What it does: This paper proposes a transformer-based amortized active intervention design method called CAASL, which can generate optimal intervention sequences in real-time without intermediate causal graph inference.
Amortized Bayesian Experimental Design for Decision-Making
Daolang Huang (Aalto University), Samuel Kaski (Aalto University)
OptimizationTransformerReinforcement LearningTabularBiomedical Data
🎯 What it does: Design a Bayesian experimental design framework that enhances decision awareness, achieving joint inference of experimental design and decision prediction through a Transformer neural decision process.
Amortized Eigendecomposition for Neural Networks
Tianbo Li (SEA AI Lab), Min Lin (SEA AI Lab)
OptimizationComputational EfficiencyGraph Neural NetworkAuto EncoderImageGraph
🎯 What it does: This paper proposes an 'Amortized Eigendecomposition' method that replaces traditional eigendecomposition with QR decomposition by introducing eigen loss, thereby efficiently implementing eigenvalue decomposition in neural network training.
Amortized Fourier Neural Operators
Zipeng Xiao (Shanghai Jiao Tong University), Zhijie Deng (Tsinghua University)
TabularPhysics Related
🎯 What it does: AM-FNO is proposed, which amplifies the parameterization of the Fourier space kernel function through neural networks, supporting arbitrary frequency patterns while significantly reducing the number of parameters.
Amortized Planning with Large-Scale Transformers: A Case Study on Chess
Anian Ruoss (Google DeepMind), Tim Genewein (Google DeepMind)
Reinforcement Learning from Human FeedbackTransformerSupervised Fine-TuningTabularBenchmark
🎯 What it does: Using a large-scale Transformer to predict action values on a chessboard, and based on this, constructing a search-free game strategy, evaluating its playing strength against traditional search engines.
Amortizing intractable inference in diffusion models for vision, language, and control
Siddarth Venkatraman (Mila, Université de Montréal), Nikolay Malkin (University of Edinburgh)
GenerationReinforcement LearningDiffusion modelImageTextMultimodality
🎯 What it does: This study investigates how to use diffusion model priors and arbitrary constraint functions to train posterior samplers, and proposes the Relative Trajectory Balancing (RTB) objective.
An Accelerated Algorithm for Stochastic Bilevel Optimization under Unbounded Smoothness
Xiaochuan Gong (George Mason University), Mingrui Liu (George Mason University)
OptimizationTextStochastic Differential Equation
🎯 What it does: This paper proposes a new accelerated bi-level optimization algorithm, AccBO, for solving upper-level non-convex, unbounded smooth functions and lower-level strongly convex problems, which can find ε-critical points with an oracle complexity of ˜O(ε⁻³) under random gradient noise.
An Accelerated Gradient Method for Convex Smooth Simple Bilevel Optimization
Jincheng Cao (University of Texas at Austin), Aryan Mokhtari (University of Texas at Austin)
OptimizationTabular
🎯 What it does: This paper proposes an Accelerated Gradient Method (AGM-BiO) for solving convex smooth simple bilevel optimization problems, which involves minimizing the upper-level objective over the optimal solution set of the lower-level problem;
An Adaptive Approach for Infinitely Many-armed Bandits under Generalized Rotting Constraints
Jung-hun Kim (Seoul National University), Se-Young Yun (KAIST)
OptimizationReinforcement Learning
🎯 What it does: An adaptive UCB algorithm for the infinite multi-armed bandit problem is proposed, along with theoretical upper and lower bounds on returns under slow and burst corrosion constraints.
An Analysis of Elo Rating Systems via Markov Chains
Sam Olesker-Taylor (University of Warwick), Luca Zanetti (University of Bath)
Graph
🎯 What it does: This paper models the online learning process of the Elo rating system under the Bradley-Terry-Luce model as a fixed-step Markov chain and provides a theoretical analysis of its convergence rate related to the spectral gap of the graph.
An Analysis of Tokenization: Transformers under Markov Data
Nived Rajaraman (University of California Berkeley), Kannan Ramchandran (University of California Berkeley)
TransformerText
🎯 What it does: This paper theoretically studies the impact of tokenization on the learning effectiveness of Transformer in the k-order Markov data generation process. It proves that without tokenization, the Transformer often remains at the character-level model; after introducing tokenization, the Transformer can achieve near-optimal cross-entropy. Theoretical and experimental analyses of tokenizers such as LZW and BPE in this context are also provided.
An Analytical Study of Utility Functions in Multi-Objective Reinforcement Learning
Manel Rodriguez-Soto (Artificial Intelligence Research Institute), Maite López-Sánchez
OptimizationReinforcement Learning
🎯 What it does: Analyzed the theoretical properties of utility functions in multi-objective reinforcement learning (MORL), provided sufficient conditions for the existence of optimal policies, and studied whether preference relations can be expressed as utility functions.
An Autoencoder-Like Nonnegative Matrix Co-Factorization for Improved Student Cognitive Modeling
Shenbao Yu (Fujian Normal University), Mingwei Lin (Fujian Normal University)
OptimizationExplainability and InterpretabilityAuto EncoderTabular
🎯 What it does: A non-negative matrix co-factorization model based on autoencoders (AE-NMCF) is proposed to simultaneously predict students' homework scores and estimate their mastery of knowledge concepts.
An effective framework for estimating individualized treatment rules
Joowon Lee (University of Wisconsin-Madison), Guanhua Chen (University of Minnesota)
OptimizationTabular
🎯 What it does: A unified Individualized Treatment Rule (ITR) estimation framework is proposed, transforming ITR learning into a convex optimization problem with weighted, hard L1 ball constraints and soft L2 regularization, solved using Projected Gradient Descent (PGD), while incorporating techniques such as distributed covariate balancing weights, variable selection, outcome enhancement, and inverse variance weighting.
An Efficient High-dimensional Gradient Estimator for Stochastic Differential Equations
Shengbo Wang (Stanford University), Peter Glynn (Stanford University)
Stochastic Differential Equation
🎯 What it does: A new unbiased gradient estimator, called the Generator Gradient Estimator, is proposed, which maintains computational complexity independent of the parameter dimension even when the SDE parameter dimension is extremely high.
An Efficient Memory Module for Graph Few-Shot Class-Incremental Learning
Dong Li (Tsinghua University), Biqing Qi (Shanghai Artificial Intelligence Laboratory)
ClassificationKnowledge DistillationGraph Neural NetworkGraph
🎯 What it does: The Mecoin framework is proposed, which implements graph few-shot incremental learning (GFSCIL) using Structured Memory Units (SMU) and Memory Representation Adaptive Modules (MRaM) to address the problem of catastrophic forgetting.
An Efficient Recipe for Long Context Extension via Middle-Focused Positional Encoding
Tong Wu (BIGAI), Zilong Zheng (BIGAI)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A long context extension method based on position encoding interpolation, called CREAM, is proposed;
An End-To-End Graph Attention Network Hashing for Cross-Modal Retrieval
Huilong Jin (Hebei Normal University), Jia Luo (Communication University of China)
RetrievalGraph Neural NetworkTransformerContrastive LearningMultimodality
🎯 What it does: An end-to-end Graph Attention Network Hashing (EGATH) framework is proposed for cross-modal retrieval.
An engine not a camera: Measuring performative power of online search
Celestine Mendler-Dünner (ELLIS Institute), Moritz Hardt (Max Planck Institute for Intelligent Systems)
🎯 What it does: Designed and implemented an online experiment (Powermeter) based on a browser plugin, which randomly altered the arrangement and hiding of Google and Bing search results without the users' knowledge, collecting click data to measure the search engines' 'performative power' (i.e., the ability to causally influence traffic through result ranking).
An Equivalence Between Static and Dynamic Regret Minimization
Andrew Jacobsen (Università degli Studi di Milano Politecnico di Milano), Francesco Orabona (KAUST)
Optimization
🎯 What it does: This paper studies the dynamic regret minimization problem in online convex optimization, demonstrating the equivalence between dynamic regret minimization under linear loss and static regret minimization in an extended decision space.
An exactly solvable model for emergence and scaling laws in the multitask sparse parity problem
Yoonsoo Nam (University of Oxford), Ard A. Louis (University of Oxford)
TransformerTabularPhysics Related
🎯 What it does: A parsable multi-task sparse parity problem model is constructed to study its emergence and scaling laws in deep networks.
An Expectation-Maximization Algorithm for Training Clean Diffusion Models from Corrupted Observations
WeiminBai, He Sun (Peking University)
RestorationGenerationDiffusion modelImage
🎯 What it does: Using the Expectation-Maximization algorithm to jointly train diffusion models from only damaged observations (without clear samples or with only a few clear samples) and solve imaging inverse problems;
An eye for an ear: zero-shot audio description leveraging an image captioner with audio-visual token distribution matching
Hugo Malard (Telecom Paris), Slim Essid (Telecom Paris)
GenerationData SynthesisRepresentation LearningTransformerPrompt EngineeringVision Language ModelVideoMultimodalityAudio
🎯 What it does: This paper proposes an unsupervised zero-shot audio description method that utilizes an image description model to map audio features into a visual encoding space through distribution alignment, thereby achieving natural language descriptions of audio.
An Image is Worth 32 Tokens for Reconstruction and Generation
Qihang Yu (ByteDance), Liang-Chieh Chen (ByteDance)
GenerationCompressionTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a Transformer-based 1D image tokenizer called TiTok, which can compress images into only 32 discrete tokens, enabling image reconstruction and generation.
An Improved Empirical Fisher Approximation for Natural Gradient Descent
Xiaodong Wu (University of Cambridge), Phil Woodland
OptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: An improved empirical Fisher (iEF) approximation method is proposed to enhance the approximation quality of natural gradient descent (NGD), and an efficient evaluation framework is designed to quantify the performance of different approximate NGD methods.
An In-depth Investigation of Sparse Rate Reduction in Transformer-like Models
Yunzhe Hu (University of Hong Kong), Dong Xu (University of Hong Kong)
CompressionOptimizationTransformerImage
🎯 What it does: This paper studies the implementation details of the Transformer-like model CRATE and its optimization objective Sparse Rate Reduction (SRR), and experimentally verifies the hierarchical optimization and generalization performance of different implementation variants.
An Information Theoretic Perspective on Conformal Prediction
Alvaro Correia, Arash Behboodi (Qualcomm AI Research)
Federated LearningImage
🎯 What it does: This paper establishes the connection between Segmented Conformal Prediction (SCP) and concepts such as conditional entropy and list decoding from an information-theoretic perspective, and derives three upper bounds; subsequently, these upper bounds are used as training objectives to achieve end-to-end conformal training, and a method to enhance the efficiency of the prediction set using side information is proposed.