ICML 2024 Papers — Page 14
International Conference on Machine Learning · 2610 papers
Learning to Compile Programs to Neural Networks
Logan Weber (Massachusetts Institute of Technology), Michael Carbin (Massachusetts Institute of Technology)
OptimizationMeta LearningAI Code AssistantLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes and implements a 'Neural Surrogate Compiler' that can directly compile program text into initial neural network parameters, which are then fine-tuned to obtain a neural network surrogate that approximates the program.
Learning to Continually Learn with the Bayesian Principle
Soochan Lee (Seoul National University), Gunhee Kim (Seoul National University)
ClassificationGenerationMeta LearningDiffusion modelAuto EncoderImage
🎯 What it does: A meta continual learning framework based on Bayesian principles (SB-MCL) is proposed, which learns a neural network as a data mapper during the meta-learning phase, allowing for the use of exponential family Bayesian sequential updates during the continual learning phase, thereby eliminating catastrophic forgetting.
Learning to Explore for Stochastic Gradient MCMC
SeungHyun Kim (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
Meta LearningImage
🎯 What it does: An adaptive Stochastic Gradient Markov Chain Monte Carlo (SGMCMC) sampler is constructed through meta-learning to enhance the exploration efficiency and mixing speed of Bayesian Neural Networks (BNN) posterior.
Learning to Explore in POMDPs with Informational Rewards
Annie Xie (Stanford University), Chelsea Finn (Stanford University)
OptimizationReinforcement LearningSequential
🎯 What it does: A POMDP exploration algorithm named PROBE is designed, which utilizes privileged information or future trajectories to construct information gain rewards and learn task-relevant information collection strategies.
Learning to Infer Generative Template Programs for Visual Concepts
R. Kenny Jones (Brown University), Daniel Ritchie (Adobe Research)
SegmentationGenerationTransformerImage
🎯 What it does: A neural symbolic framework called Template Programs is proposed to infer generalizable generative programs from a small number of visual samples.
Learning to Intervene on Concept Bottlenecks
David Steinmann (TU Darmstadt), Kristian Kersting (TU Darmstadt)
ClassificationRecognitionImage
🎯 What it does: This paper proposes the Concept Bottleneck Memory Model (CB2M), which adds two layers of memory to the traditional concept bottleneck model to continuously utilize and reuse human interventions.
Learning to Model the World With Language
Jessy Lin (University of California Berkeley), Anca Dragan (University of California Berkeley)
Recurrent Neural NetworkReinforcement LearningWorld ModelTextMultimodality
🎯 What it does: A multimodal world model named Dynalang is proposed, which learns the association between language and vision through future prediction, supporting the understanding and application of various forms of language (instructions, prompts, manuals, etc.) in multiple environments.
Learning to Play Atari in a World of Tokens
Pranav Agarwal (College of Technology Superior), Samira Ebrahimi Kahou (University of Calgary)
TransformerReinforcement LearningWorld ModelVideoBenchmark
🎯 What it does: This paper presents DART, which uses discrete VQ-VAE representations and a Transformer structure (decoder modeling the world, encoder learning the policy) to achieve sample-efficient learning in Atari games without lookahead search.
Learning to Predict Mutational Effects of Protein-Protein Interactions by Microenvironment-aware Hierarchical Prompt Learning
Lirong Wu (Westlake University), Stan Z. Li (Westlake University)
Protein Structure PredictionGraph Neural NetworkPrompt EngineeringBiomedical Data
🎯 What it does: A deep model called Prompt-DDG based on microenvironment hierarchical prompt learning is proposed for predicting the mutation effects (ΔΔG) of protein-protein interactions.
Learning to Reach Goals via Diffusion
Vineet Jain (McGill University), Siamak Ravanbakhsh (Mila Quebec Artificial Intelligence Institute)
Computational EfficiencyRobotic IntelligenceReinforcement LearningDiffusion modelContrastive LearningSequential
🎯 What it does: A target-conditioned reinforcement learning method called Merlin is proposed from the perspective of diffusion models. By constructing a forward diffusion process in the state space and learning a reverse 'denoising' policy, it achieves the transition from any initial state to a specified target.
Learning to Remove Cuts in Integer Linear Programming
Pol Puigdemont (École Polytechnique Fédérale de Lausanne), Volkan Cevher (École Polytechnique Fédérale de Lausanne)
OptimizationReinforcement LearningTabular
🎯 What it does: In the cutting plane method for integer linear programming, a strategy is proposed to simultaneously add multiple cuts and determine which old cuts to remove through a learning model;
Learning to Route Among Specialized Experts for Zero-Shot Generalization
Mohammed Muqeeth (Massachusetts Institute of Technology International Business Machines), Colin Raffel (University of Toronto)
TransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsTextBenchmark
🎯 What it does: This paper proposes a post-hoc adaptive token-level routing method called PHATGOOSE, which enables dynamic routing among a large number of parameter-efficient fine-tuned expert models (such as LoRA modules) to enhance the zero-shot generalization of pre-trained language models on unseen tasks.
Learning to Scale Logits for Temperature-Conditional GFlowNets
Minsu Kim (Korea Advanced Institute of Science and Technology), Yoshua Bengio (Universite de Montreal)
GenerationOptimizationGraph Neural NetworkReinforcement LearningGraphSequential
🎯 What it does: Logit-GFN is proposed—a framework that significantly enhances the training stability, convergence speed, and generalization ability of temperature-conditioned GFlowNet by directly applying temperature scaling to the logits of GFlowNet.
Learning to Stabilize Online Reinforcement Learning in Unbounded State Spaces
Brahma S Pavse, Josiah P. Hanna (University of Wisconsin Madison)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the stability problem of online reinforcement learning in unbounded state spaces and proposes a STOP method that combines Lyapunov cost shaping with state transformation, which can ensure long-term stability of the system while achieving optimal performance.
Learning Universal Predictors
Jordi Grau-Moya (Google DeepMind), Joel Veness (Google DeepMind)
Meta LearningRecurrent Neural NetworkTransformerSequential
🎯 What it does: Using meta-learning to train neural networks to approximate the universal predictive power of Solomonoff Induction (SI).
Learning Useful Representations of Recurrent Neural Network Weight Matrices
Vincent Herrmann (Swiss AI Lab IDSIA), Jürgen Schmidhuber (Swiss AI Lab IDSIA)
Representation LearningRecurrent Neural NetworkSequential
🎯 What it does: Proposes and compares various methods for encoding the representation of weights in recurrent neural networks, particularly developing an interactive probing architecture.
Learning with 3D rotations, a hitchhiker's guide to SO(3)
Andreas René Geist, Georg Martius (Max Planck Institute for Intelligent Systems)
Pose EstimationRepresentation LearningImagePoint CloudReview/Survey Paper
🎯 What it does: This paper reviews and evaluates the learnability of three-dimensional rotation representations in deep learning, providing usage recommendations in the contexts of rotation estimation and feature prediction.
Learning with Adaptive Resource Allocation
Jing Wang (Nanjing University), Zhi-Hua Zhou (Nanjing University)
ImageText
🎯 What it does: This paper proposes an adaptive resource allocation method called LARA for parallel training of multiple time-constrained machine learning tasks under limited computational resources.
Learning with Complementary Labels Revisited: The Selected-Completely-at-Random Setting Is More Practical
Wei Wang (University of Tokyo), Masashi Sugiyama (University of Tokyo)
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A consistency method for supplementary label learning called SCARCE is proposed, which does not rely on the assumption of uniform distribution or additional ordinary label data.
Learning with Partial-Label and Unlabeled Data: A Uniform Treatment for Supervision Redundancy and Insufficiency
Yangfan Liu (Southeast University), Ning Xu (Southeast University)
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes a unified framework SPMI for simultaneously addressing two weak supervision scenarios: label redundancy (partial labels) and label insufficiency (no labels). It gradually identifies the true labels by dynamically exchanging labels in the candidate label set of all samples through a label channel.
Learning-Efficient Yet Generalizable Collaborative Filtering for Item Recommendation
Yuanhao Pu (University of Science and Technology of China), Enhong Chen (University of Science and Technology of China)
Recommendation SystemGraph Neural NetworkTabular
🎯 What it does: This paper proposes a new squared loss function RG² for item recommendation in collaborative filtering.
Learning-Rate-Free Stochastic Optimization over Riemannian Manifolds
Daniel Dodd (Lancaster University), Christopher Nemeth (Lancaster University)
OptimizationTabular
🎯 What it does: A random Riemannian optimization algorithm without learning rates (RDoG, RDoWG, NRDoG) is proposed, and a high-probability convergence proof is provided.
Less is More: on the Over-Globalizing Problem in Graph Transformers
Yujie Xing (Beijing University of Posts and Telecommunications), Chuan Shi (Beijing University of Posts and Telecommunications)
Graph Neural NetworkTransformerGraph
🎯 What it does: This paper reveals the problem of excessive globalization in graph Transformers and proposes the CoBFormer architecture to alleviate this issue.
LESS: Selecting Influential Data for Targeted Instruction Tuning
Mengzhou Xia (Princeton University), Danqi Chen (Princeton University)
Supervised Fine-TuningText
🎯 What it does: A data selection method named LESS is designed to select instruction tuning data for specific tasks by estimating the influence of training samples on validation samples.
Lessons from Generalization Error Analysis of Federated Learning: You May Communicate Less Often!
Milad Sefidgaran (Huawei), Yijun Wan (Huawei)
Federated LearningConvolutional Neural NetworkImage
🎯 What it does: The theoretical analysis of the evolution of the generalization error of models in federated learning (FL) with respect to the communication rounds R is conducted, providing two types of upper bounds: PAC-Bayes and information theory (rate-distortion); subsequently, it is applied to federated SVM (FSVM) and experimental validation based on ResNet-56.
Let Go of Your Labels with Unsupervised Transfer
Artyom Gadetsky (EPFL), Maria Brbic
ClassificationDomain AdaptationRepresentation LearningTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A completely unsupervised transfer framework named TURTLE is proposed, which can automatically discover the true labels of a dataset from a pre-trained base model without using any labels.
Leverage Class-Specific Accuracy to Guide Data Generation for Improving Image Classification
Jay Gala (University of California San Diego), Pengtao Xie (Mohamed bin Zayed University of Artificial Intelligence)
ClassificationGenerationData SynthesisOptimizationGenerative Adversarial NetworkImage
🎯 What it does: A multi-layer optimization framework is proposed, which dynamically allocates generation budgets based on the accuracy of each category on the validation set, generating synthetic images that match category demands, thereby improving image classification performance.
Leveraging (Biased) Information: Multi-armed Bandits with Offline Data
Wang Chi Cheung (National University of Singapore), Lixing Lyu (National University of Singapore)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies how to utilize potentially biased offline data for online learning in multi-armed bandits.
Leveraging Attractor Dynamics in Spatial Navigation for Better Language Parsing
Xiaolong Zou (Qiyuan Lab), Bo Hong (Qiyuan Lab)
Recurrent Neural NetworkText
🎯 What it does: A PHE-trinity model is proposed, utilizing a hippocampal-entorhinal cortex continuous attractor network for the representation of grammatical structures, and achieving language instruction parsing through structure-content separation.
Leveraging Self-Consistency for Data-Efficient Amortized Bayesian Inference
Marvin Schmitt (University of Stuttgart), Stefan T. Radev (Rensselaer Polytechnic Institute)
OptimizationComputational EfficiencyFlow-based ModelTabularTime Series
🎯 What it does: This paper proposes a loss function based on the symmetry of joint probability models—self-consistency—to enhance the efficiency and accuracy of amortized Bayesian inference (ABI), especially in low data budget scenarios.
Leveraging VLM-Based Pipelines to Annotate 3D Objects
Rishabh Kabra (Google DeepMind), Niloy Mitra
ClassificationObject DetectionTransformerVision Language ModelPoint Cloud
🎯 What it does: A novel unsupervised multi-view aggregation algorithm called ScoreAgg based on visual language models is proposed to automatically generate reliable type and material annotations for a large number of 3D objects.
LEVI: Generalizable Fine-tuning via Layer-wise Ensemble of Different Views
Yuji Roh (Google Inc), Zhe Zhao (Google DeepMind)
Domain AdaptationRecommendation SystemSupervised Fine-TuningImageText
🎯 What it does: This paper proposes a hierarchical ensemble framework named LEVI, which combines large-scale pre-trained models with small task-specific models to enhance the model's generalization performance in out-of-distribution (OOD) environments through hierarchical fusion during the fine-tuning phase.
Libra: Building Decoupled Vision System on Large Language Models
Yifan Xu (Chinese Academy of Sciences), Changsheng Xu (Chinese Academy of Sciences)
RecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsVision Language ModelImageTextMultimodality
🎯 What it does: A decoupled visual system Libra is built on large language models, achieving multimodal reasoning from images to text.
LIDAO: Towards Limited Interventions for Debiasing (Large) Language Models
Tianci Liu (Purdue University), Jing Gao (Purdue University)
GenerationOptimizationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a limited intervention framework named LIDAO for debiasing large language models (LLMs) while maintaining generative fluency, and extends the method to adversarial prompts.
Lie Neurons: Adjoint-Equivariant Neural Networks for Semisimple Lie Algebras
Tzu-Yuan Lin (University of Michigan), Maani Ghaffari (University of Michigan)
Point Cloud
🎯 What it does: A neural network for adjoint equivariance of semisimple Lie algebras, called Lie Neurons, is proposed, which can handle inputs from any finite-dimensional semisimple Lie algebra.
Light and Optimal Schrödinger Bridge Matching
Nikita Gushchin (Skolkovo Institute of Science and Technology), Alexander Korotin (Artificial Intelligence Research Institute)
Image TranslationOptimizationImageBenchmarkStochastic Differential Equation
🎯 What it does: A one-time bridge matching Schrödinger bridge learning method is proposed, and a lightweight solver LightSB-M is implemented.
Lightweight Image Super-Resolution via Flexible Meta Pruning
Yulun Zhang (Shanghai Jiao Tong University), Fisher Yu (ETH Zurich)
RestorationSuper ResolutionConvolutional Neural NetworkImage
🎯 What it does: A flexible meta-pruning method for lightweight image super-resolution (FMP) is proposed, which can perform both structured (channel) and unstructured (weight) pruning during training.
Limited Preference Aided Imitation Learning from Imperfect Demonstrations
Xingchen Cao (Nanjing University), Yang Yu (Nanjing University)
Robotic IntelligenceReinforcement Learning from Human FeedbackReinforcement LearningTabular
🎯 What it does: This paper proposes a preference-assisted imitation learning algorithm called PAIL, which can learn strategies that surpass the performance of the original demonstrations in scenarios with only imperfect demonstrations and limited human preference feedback.
Linear Alignment: A Closed-form Solution for Aligning Human Preferences without Tuning and Feedback
Songyang Gao (Shanghai Artificial Intelligence Laboratory), Dahua Lin (Shanghai Artificial Intelligence Laboratory)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Proposes the Linear Alignment method, which aligns LLMs with human preferences through a single inference step without model training or manual labeling;
Linear Explanations for Individual Neurons
Tuomas Oikarinen (University of California San Diego), Tsui-Wei Weng (University of California San Diego)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerImageMultimodality
🎯 What it does: A method called Linear Explanations is proposed, which views the activation of a single neuron as a linear combination of concepts, generating explanations through a concept activation matrix and sparse linear regression, while also introducing a simulation-based automatic evaluation framework.
Linguistic Calibration of Long-Form Generations
Neil Band (Stanford University), Tatsunori Hashimoto (Stanford University)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a self-calibration method for language models aimed at long text generation, enabling the model to naturally express confidence in various statements during the generation process, thereby assisting users in making better decisions.
Liouville Flow Importance Sampler
Yifeng Tian (Los Alamos National Laboratory), Yen Ting Lin (Los Alamos National Laboratory)
Flow-based ModelOrdinary Differential Equation
🎯 What it does: A pure flow model (LFIS) based on the Liouville equation is proposed, which efficiently samples non-normalized density functions and estimates the log-normalization constant by deterministically transporting a simple initial distribution along a given temperature path to the target distribution through learning a time-dependent velocity field.
Listenable Maps for Audio Classifiers
Francesco Paissan (Fondazione Bruno Kessler), Cem Subakan (Laval University)
ClassificationExplainability and InterpretabilityConvolutional Neural NetworkAudio
🎯 What it does: This paper proposes a post-hoc interpretability method called L-MAC, which utilizes the latent representations of a pre-trained audio classifier to train a decoder that generates binary masks. These masks are then applied to the STFT magnitudes to produce audible explanation audio, helping humans understand the classification results.
Listening to the noise: Blind Denoising with Gibbs Diffusion
David Heurtel-Depeiges (Ecole Polytechnique), Bruno Régaldo-Saint Blancard (Flatiron Institute)
RestorationDiffusion modelImage
🎯 What it does: A blind denoising framework named Gibbs Diffusion (GDiff) is proposed, which combines diffusion models and Gibbs sampling to simultaneously recover signal and noise parameters.
Listwise Reward Estimation for Offline Preference-based Reinforcement Learning
Heewoong Choi (Seoul National University), Taesup Moon (Seoul National University)
Reinforcement LearningSequential
🎯 What it does: This paper proposes a new offline preference-based reinforcement learning method called LiRE, which constructs a complete trajectory ranking list (RLT) and effectively utilizes second-order preference information through the same ternary feedback to achieve more accurate reward function estimation.
LLaGA: Large Language and Graph Assistant
Runjin Chen (University of Texas at Austin), Zhangyang Wang (University of Texas at Austin)
ClassificationGraph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: This paper proposes the LLaGA framework, which maps graph structures to the word embedding space of LLMs through node sequence templates and projectors, achieving a seamless integration of graph tasks with large models.
LLark: A Multimodal Instruction-Following Language Model for Music
Joshua P Gardner, Rachel M Bittner
GenerationData SynthesisTransformerLarge Language ModelSupervised Fine-TuningMultimodalityAudio
🎯 What it does: Construct and train a multimodal instruction-following language model (LLARK) that can generate natural language answers based on audio and text instructions, covering tasks such as music understanding, title generation, and reasoning.
LLM and Simulation as Bilevel Optimizers: A New Paradigm to Advance Physical Scientific Discovery
Pingchuan Ma (Massachusetts Institute of Technology), Wojciech Matusik (Massachusetts Institute of Technology)
OptimizationDrug DiscoveryAI Code AssistantTransformerLarge Language ModelTextTabularPhysics Related
🎯 What it does: A dual-layer optimization framework combining large language models and differentiable physical simulations (Scientific Generative Agent, SGA) is proposed for automated physical science discovery, including constitutive law search and molecular design.
LLM Maybe LongLM: SelfExtend LLM Context Window Without Tuning
Hongye Jin (Texas A&M University), Xia Hu (Rice University)
TransformerLarge Language ModelText
🎯 What it does: The SelfExtend method is proposed, which expands the context window of large language models without fine-tuning during the inference phase by mapping unseen relative positional information and combining hierarchical attention (grouped attention and adjacent normal attention).
LLM-Empowered State Representation for Reinforcement Learning
Boyuan Wang (Tsinghua University), Xiangyang Ji (Tsinghua University)
Large Language ModelReinforcement LearningTabularBenchmark
🎯 What it does: Utilizing LLM to automatically generate task-related state representations and intrinsic reward functions, improving the quality of state representation in RL and accelerating learning.
Local Causal Structure Learning in the Presence of Latent Variables
Feng Xie (Beijing Technology and Business University), Zhi Geng (Beijing Technology and Business University)
GraphTabular
🎯 What it does: An algorithm is proposed to accurately discover the direct causal relationships of a given target variable using local structure learning (MMB-by-MMB) in the presence of latent variables.
Local Feature Selection without Label or Feature Leakage for Interpretable Machine Learning Predictions
Harrie Oosterhuis (Radboud University), Avishek Anand (TU Delft)
OptimizationExplainability and InterpretabilityReinforcement LearningImageTabular
🎯 What it does: The theoretical definitions of no-label leakage and no-feature leakage are proposed, along with a completely no-leak local feature selection method—linear programming solver and a practical Sequential Unmasking Without Reversion (SUWR).
Local vs. Global Interpretability: A Computational Complexity Perspective
Shahaf Bassan (Hebrew University of Jerusalem), Guy Katz (Hebrew University of Jerusalem)
Explainability and InterpretabilityTabular
🎯 What it does: A systematic theoretical analysis of the local and global interpretability of machine learning models from the perspective of computational complexity is conducted, providing a complexity classification of various forms of explanations (such as minimal sufficient reasons, necessary/redundant features, and completion counts).
Locality-Sensitive Hashing-Based Efficient Point Transformer with Applications in High-Energy Physics
Siqi Miao (Georgia Institute of Technology), Pan Li (Georgia Institute of Technology)
TransformerPoint CloudPhysics Related
🎯 What it does: An efficient point Transformer (HEPT) based on Locality Sensitive Hashing (LSH) is designed for real-time processing of large-scale point clouds.
Localizing Task Information for Improved Model Merging and Compression
Ke Wang (Ecole Polytechnique Federale de Lausanne), Pascal Frossard (Ecole Polytechnique Federale de Lausanne)
CompressionOptimizationTransformerSupervised Fine-TuningImageText
🎯 What it does: This paper proposes the TALL-masks method, which locates and extracts key information for each task from a multi-task vector using binary masks; at the same time, it constructs Consensus Merging based on the masks to eliminate selfish/disastrous weights, significantly improving the model fusion effect; and it achieves efficient compression using the masks, requiring only the storage of zero-shot models, merged vectors, and masks to recover nearly complete performance.
Locally Differentially Private Decentralized Stochastic Bilevel Optimization with Guaranteed Convergence Accuracy
Ziqin Chen (Clemson University), Yongqiang Wang (Clemson University)
OptimizationFederated LearningSafty and PrivacyMeta LearningImage
🎯 What it does: A decentralized random dual-layer optimization algorithm with local differential privacy is proposed, avoiding the nested consistency loops in traditional algorithms while balancing accurate convergence and privacy protection.
Locally Estimated Global Perturbations are Better than Local Perturbations for Federated Sharpness-aware Minimization
Ziqing Fan (Shanghai Jiao Tong University), Yanfeng Wang (Shanghai Jiao Tong University)
OptimizationFederated LearningImage
🎯 What it does: A SAM algorithm for local estimation of global perturbation in federated learning (FedLESAM) is proposed, which avoids the misleading effects of traditional SAM on local loss surfaces under multi-step local updates and data heterogeneity;
Locally Interdependent Multi-Agent MDP: Theoretical Framework for Decentralized Agents with Dynamic Dependencies
Alex DeWeese (Carnegie Mellon University), Guannan Qu (Carnegie Mellon University)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: A distributed Markov decision process model named Locally Interdependent Multi-Agent MDP is proposed, along with three closed-form approximately optimal decentralized strategies.
LoCoCo: Dropping In Convolutions for Long Context Compression
Ruisi Cai (University of Texas at Austin), Beidi Chen (Meta AI)
CompressionTransformerLarge Language ModelText
🎯 What it does: The LoCoCo method is proposed, which inserts a one-dimensional convolution fusion layer into the self-attention module of LLMs to compress the KV cache, supporting efficient inference and fine-tuning for long contexts.
Log Neural Controlled Differential Equations: The Lie Brackets Make A Difference
Benjamin Walker (University of Oxford), Terry Lyons (University of Oxford)
OptimizationComputational EfficiencyTime SeriesOrdinary Differential Equation
🎯 What it does: Introduces Log-NCDEs, which improve multivariate time series modeling methods by explicitly constructing Lie brackets in the vector field of NCDE.
Logistic Variational Bayes Revisited
Michael Komodromos (Imperial College London), Sarah Lucie Filippi
ClassificationOptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes a new upper bound for the expected softplus function and applies it to variational logistic regression and Gaussian process classification, introducing the VI-PER method.
Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning
Hao Zhao (École Polytechnique Fédérale de Lausanne), Nicolas Flammarion (École Polytechnique Fédérale de Lausanne)
Supervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes a simple data selection method for instruction fine-tuning that uses only the longest 1,000 replies, and further enhances model performance through self-revision and noise augmentation.
Long Range Propagation on Continuous-Time Dynamic Graphs
Alessio Gravina (University of Pisa), Claas Grohnfeldt (Huawei Technologies)
Graph Neural NetworkGraphTime SeriesBenchmarkOrdinary Differential Equation
🎯 What it does: A continuous-time antisymmetric network (CTAN) based on stable non-dissipative ODEs is proposed to achieve long-range information propagation in C-TDG.
Long-Tail Learning with Foundation Model: Heavy Fine-Tuning Hurts
Jiang-Xin Shi (Nanjing University), Yu-Feng Li (Nanjing University)
ClassificationData-Centric LearningTransformerSupervised Fine-TuningContrastive LearningImage
🎯 What it does: This paper demonstrates that full fine-tuning leads to a decline in tail class performance in long-tail learning tasks by comparing full fine-tuning and lightweight fine-tuning, and proposes a lightweight fine-tuning framework called LIFT, which can achieve high accuracy with very few learnable parameters in a short time.
Longitudinal Targeted Minimum Loss-based Estimation with Temporal-Difference Heterogeneous Transformer
Toru Shirakawa (Osaka University), Mark J. van der Laan
TransformerTime SeriesSequential
🎯 What it does: This paper proposes a deep longitudinal TMLE method based on Transformer (Deep LTMLE) for estimating causal average outcomes under dynamic interventions.
LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens
Yiran Ding (Microsoft Research), Mao Yang (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By optimizing RoPE positional encoding and evolutionary search, the context window of the pre-trained LLM is expanded from 4k to 2048k, achieving long text inference while maintaining original performance.
Look Ahead or Look Around? A Theoretical Comparison Between Autoregressive and Masked Pretraining
Qi Zhang (Peking University), Yisen Wang (Peking University)
ClassificationGenerationTransformerSupervised Fine-TuningImageTextMultimodality
🎯 What it does: Theoretical comparison of autoregressive and masked self-supervised pre-training, and proposed improvement objectives.
Lookbehind-SAM: k steps back, 1 step forward
Goncalo Mordido (Mila Quebec AI Institute), Sarath Chandar (Canada CIFAR AI Chair)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: An optimizer named Lookbehind is proposed, which improves the traditional Sharpness-Aware Minimization (SAM) by searching for worse perturbations through multi-step ascent and using linear interpolation to reduce gradient variance, thereby further reducing the sharpness of the loss while maintaining a low loss.
LoRA Training in the NTK Regime has No Spurious Local Minima
Uijeong Jang (Seoul National University), Ernest K. Ryu (University of California)
OptimizationTransformerSupervised Fine-TuningImageTextAudio
🎯 What it does: The theoretical analysis of LoRA (Low-Rank Adaptation) fine-tuning under the NTK (Neural Tangent Kernel) framework proves the existence of a global optimal solution with a rank of √N, and when the rank r ≳ √N, the optimization landscape has no spurious local minima, allowing gradient descent to converge to a low-rank global optimum, which also exhibits good generalization performance.
LoRA+: Efficient Low Rank Adaptation of Large Models
Soufiane Hayou (Simons Institute), Bin Yu (University of California Berkeley)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper studies the suboptimal problem encountered when fine-tuning large-width pre-trained language models using LoRA, and proposes the LoRA+ method, which improves feature learning efficiency and fine-tuning speed by setting different learning rate ratios for two layers of adapters.
LoRAP: Transformer Sub-Layers Deserve Differentiated Structured Compression for Large Language Models
Guangyan Li (Chinese Academy of Sciences), Wensheng Zhang (Guangzhou University)
CompressionTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes LoRAP, which applies weighted low-rank approximation to the MHA layer of the Transformer and uses gradient-free structured channel pruning for the FFN layer, restoring performance through LoRA fine-tuning after compression.
Loss Shaping Constraints for Long-Term Time Series Forecasting
Ignacio Hounie (University of Pennsylvania), Alejandro Ribeiro (University of Pennsylvania)
TransformerTime SeriesSequential
🎯 What it does: A long-term time series forecasting framework based on constraint learning, called Loss Shaping Constraints, is proposed, which can set error upper bounds for each prediction time step during training and automatically adjust through recoverable soft constraints.
Low-Cost High-Power Membership Inference Attacks
Sajjad Zarifzadeh (National University of Singapore), Reza Shokri (National University of Singapore)
Adversarial AttackImage
🎯 What it does: A low-cost, high-power membership inference attack method called RMIA is proposed, which can perform membership inference on the target model using only a small number of pre-trained reference models.
Low-Rank Bandits via Tight Two-to-Infinity Singular Subspace Recovery
Yassir Jedra (Massachusetts Institute of Technology), Alexandre Proutiere (KTH Royal Institute of Technology)
Recommendation SystemOptimizationReinforcement Learning
🎯 What it does: For the contextual Bandit problem with low-rank structure, a two-stage algorithm is proposed: the first stage uses spectral methods to estimate the singular subspace of the reward matrix, and the second stage transforms the problem into a linear Bandit with missing parameters, implementing policy evaluation (PE), best policy identification (BPI), and scheduling minimization respectively.
Low-Rank Similarity Mining for Multimodal Dataset Distillation
Yue Xu (Shanghai Jiao Tong University), Yong-Lu Li (Shanghai Jiao Tong University)
Data SynthesisRetrievalCompressionKnowledge DistillationTransformerContrastive LearningImageTextMultimodality
🎯 What it does: Proposes Low-Rank Similarity Mining (LoRS) for distilling image-text datasets, enabling synthetic data to contain both images and text while learning their corresponding similarity matrix;
LPGD: A General Framework for Backpropagation through Embedded Optimization Layers
Anselm Paulus (Max Planck Institute for Intelligent Systems), Vít Musil (Masaryk University)
OptimizationTabularFinance Related
🎯 What it does: This paper proposes the Lagrangian Proximal Gradient Descent (LPGD) framework for embedding parameterized optimization layers in neural networks and achieving effective backpropagation, addressing the issues of traditional gradient information loss or degradation.
LQER: Low-Rank Quantization Error Reconstruction for LLMs
Cheng Zhang (Imperial College London), Yiren Zhao (Imperial College London)
CompressionOptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: A new post-training quantization framework LQER is proposed, which combines quantization and low-rank approximation to reconstruct the quantization error of weights, thereby restoring the performance of large language models (LLMs) without additional training or iterative optimization.
LSEnet: Lorentz Structural Entropy Neural Network for Deep Graph Clustering
Li Sun (North China Electric Power University), Philip S. Yu (University of Illinois at Chicago)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A graph clustering method called LSEnet is proposed, which does not require a predefined number of clusters. It utilizes differentiable structural information (DSI) to learn a hierarchical partition tree in the hyperbolic space of the Lorentz model and complete clustering.
Lyapunov-stable Neural Control for State and Output Feedback: A Novel Formulation
Lujie Yang (Massachusetts Institute of Technology), Huan Zhang (University of Illinois at Urbana-Champaign)
Reinforcement LearningTime Series
🎯 What it does: Design and train neural network controllers and observers, and provide formal stability guarantees for the closed-loop system through Lyapunov functions (including state feedback and output feedback).
Machine Vision Therapy: Multimodal Large Language Models Can Enhance Visual Robustness via Denoising In-Context Learning
Zhuo Huang (University of Sydney), Tongliang Liu (University of Sydney)
Domain AdaptationTransformerLarge Language ModelSupervised Fine-TuningImageMultimodality
🎯 What it does: Utilize multimodal large language models for error diagnosis and correction of visual models to enhance their robustness in out-of-distribution scenarios.
MADA: Meta-Adaptive Optimizers Through Hyper-Gradient Descent
Kaan Ozkara (University of California Los Angeles), Volkan Cevher (Amazon Web Services)
OptimizationTransformerLarge Language ModelImageText
🎯 What it does: A framework for dynamically learning optimal optimizers in deep learning is proposed - the Meta-Adaptive Optimizer (MADA). It performs parametric interpolation among a set of known optimizers (such as Adam, Adan, Yogi, Lion, etc.) and utilizes hyper-gradient descent to update the interpolation coefficients in real-time during training, automatically finding the most suitable optimizer. Additionally, AVGrad is introduced as an improved version of AMSGrad, replacing the maximum operation with a time average to enhance hyper-gradient flow.
Maestro: Uncovering Low-Rank Structures via Trainable Decomposition
Samuel Horváth (Mohamed bin Zayed University of Artificial Intelligence), Hongyi Wang (Carnegie Mellon University)
CompressionOptimizationConvolutional Neural NetworkTransformerSupervised Fine-TuningImageText
🎯 What it does: The MAESTRO framework is proposed, which dynamically learns the optimal rank of each layer during training through Low-Rank Ordered Decomposition (LOD), achieving low-rank compression of the model.
MAGDi: Structured Distillation of Multi-Agent Interaction Graphs Improves Reasoning in Smaller Language Models
Justin Chen, Mohit Bansal (University of North Carolina Chapel Hill)
Knowledge DistillationGraph Neural NetworkLarge Language ModelContrastive LearningTextGraph
🎯 What it does: Construct a Multi-Agent Generated Graph (MAG) and distill its structure into a small language model to enhance reasoning capabilities.
MagicLens: Self-Supervised Image Retrieval with Open-Ended Instructions
Kai Zhang (Ohio State University), Ming-Wei Chang (Google DeepMind)
RetrievalTransformerLarge Language ModelContrastive LearningImageText
🎯 What it does: The MagicLens model is proposed to achieve image retrieval based on acceptable open-text instructions.
Magicoder: Empowering Code Generation with OSS-Instruct
Yuxiang Wei (University of Illinois), LINGMING ZHANG
GenerationAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Inspired by open-source code snippets, the Magicoder series of 7B-level LLMs has been developed specifically for code generation.
MagicPose: Realistic Human Poses and Facial Expressions Retargeting with Identity-aware Diffusion
Di Chang (University of Southern California), Mohammad Soleymani
GenerationPose EstimationDiffusion modelImageVideo
🎯 What it does: MagicPose is a two-dimensional human pose and facial expression redirection method based on diffusion models, capable of generating new images while maintaining identity, controlling the target pose and expression.
MAGNOLIA: Matching Algorithms via GNNs for Online Value-to-go Approximation
Alexandre Hayderi (Stanford University), Anders Wikum (Stanford University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes an online Bayesian bipartite matching algorithm called MAGNOLIA based on graph neural networks (GNN), which learns to estimate value-to-go (VTG) using GNN to make approximately optimal matching decisions upon the arrival of each online node.
Major-Minor Mean Field Multi-Agent Reinforcement Learning
Kai Cui (Technische Universitaet Darmstadt), Heinz Koeppl (Technische Universitaet Darmstadt)
Reinforcement Learning
🎯 What it does: The Major-Minor Mean Field Control (M3FC) framework is proposed, which integrates mean field modeling of the main agent and multiple agents, and based on this, the M3FMARL algorithm is designed to solve large-scale cooperative multi-agent reinforcement learning using centralized training and decentralized execution.
Make-A-Shape: a Ten-Million-scale 3D Shape Model
Ka-Hei Hui (Chinese University of Hong Kong), Chi-Wing Fu (Chinese University of Hong Kong)
GenerationData SynthesisDiffusion modelMultimodalityPoint Cloud
🎯 What it does: Trained a 3D shape generation model called Make-A-Shape based on 100 trillion voxel-level data, capable of quickly generating high-quality, detail-rich 3D models from single views, multiple views, point clouds, voxels, and other multimodal conditions.
Making Old Things New: A Unified Algorithm for Differentially Private Clustering
Max Dupre la Tour (McGill University), David Saulpic (CNRS)
OptimizationSafty and Privacy
🎯 What it does: A unified private k-means clustering method based on the greedy algorithm by Mettu-Plaxton from 20 years ago is proposed, which can achieve approximate clustering under various differential privacy models such as central, local, shuffled, continuous observation, and MPC.
MALIBO: Meta-learning for Likelihood-free Bayesian Optimization
Jiarong Pan (Bosch Center for Artificial Intelligence), Joaquin Vanschoren (Eindhoven University of Technology)
OptimizationHyperparameter SearchMeta LearningTabularBenchmark
🎯 What it does: This paper presents MALIBO, a likelihood-free Bayesian optimization method based on meta-learning that directly learns acquisition functions for different tasks, bypassing traditional surrogate models.
Manifold Integrated Gradients: Riemannian Geometry for Feature Attribution
Eslam Zaher (University of Queensland), Fred Roosta (University of Queensland)
Explainability and InterpretabilityAdversarial AttackConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: This paper proposes a manifold-based Integral Gradient method (MIG), which generates feature attribution maps by calculating gradients along geodesics in the Riemannian latent space learned by VAE.
Mapping the Multiverse of Latent Representations
Jeremy Wayland (Helmholtz Munich), Bastian Rieck (Helmholtz Munich)
GenerationData SynthesisRepresentation LearningTransformerAuto EncoderImageText
🎯 What it does: The PRESTO framework is proposed to map and compare the latent space structures of different machine learning models (especially those using latent representations) in a multiverse.
Masked Face Recognition with Generative-to-Discriminative Representations
Shiming Ge (Institute of Information Engineering Chinese Academy of Sciences), Dan Zeng (Shanghai University)
RecognitionKnowledge DistillationRepresentation LearningConvolutional Neural NetworkTransformerGenerative Adversarial NetworkContrastive LearningImage
🎯 What it does: A unified deep network is proposed, which first learns occlusion-robust general representations using a generative encoder, and then transforms them into identity recognition features through a discriminative rewriter and knowledge distillation, ultimately training a classification head to complete occluded face recognition.
MaSS: Multi-attribute Selective Suppression for Utility-preserving Data Transformation from an Information-theoretic Perspective
Yizhuo Chen (University of Illinois Urbana-Champaign), Tarek F. Abdelzaher
OptimizationSafty and PrivacyContrastive LearningMultimodalityAudio
🎯 What it does: A multi-attribute selective suppression framework based on information theory (MaSS) is proposed and implemented, which can suppress specified sensitive attributes while maintaining data availability, and also manage unlabeled general features.
Mastering Robot Manipulation with Multimodal Prompts through Pretraining and Multi-task Fine-tuning
Jiachen Li (University of California), William Yang Wang (University of California)
Robotic IntelligenceTransformerLarge Language ModelPrompt EngineeringImageTextMultimodality
🎯 What it does: Train robots to complete manipulation tasks through pre-trained inverse dynamics and multi-task fine-tuning, using multimodal (text + image) prompts.
Mastering Text-to-Image Diffusion: Recaptioning, Planning, and Generating with Multimodal LLMs
Ling Yang (Peking University), Bin CUI
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageText
🎯 What it does: A training-free framework called RPG (Recaption, Plan, Generate) is proposed, utilizing multimodal large language models (MLLM) for text re-labeling, planning regions, and achieving region-level generation and editing within diffusion models.
Mastering Zero-Shot Interactions in Cooperative and Competitive Simultaneous Games
Yannik Mahlau (Leibniz University Hannover), Bodo Rosenhahn (Leibniz University Hannover)
Robotic IntelligenceReinforcement LearningAgentic AITabularBenchmark
🎯 What it does: In a simultaneous cooperation and competition game, the Albatross method is proposed, which can adaptively make decisions based on the opponent's bounded rationality under zero-shot interaction.
MathScale: Scaling Instruction Tuning for Mathematical Reasoning
Zhengyang Tang (Chinese University of Hong Kong), Furu Wei (Microsoft Research Asia)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes a scalable mathematical reasoning data generation method called MathScale, which utilizes GPT-3.5 to automatically generate two million high-quality question-answer pairs and performs instruction tuning on LLMs using MWPBENCH.
Matrix Information Theory for Self-Supervised Learning
Yifan Zhang (Tsinghua University), Yang Yuan (Tsinghua University)
Object DetectionRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningImageText
🎯 What it does: A self-supervised learning framework based on matrix information theory, Matrix-SSL, is proposed, which combines maximum entropy coding with matrix uniformity and alignment loss to improve non-contrastive learning.