ICML 2024 Papers — Page 3
International Conference on Machine Learning · 2610 papers
AnyTool: Self-Reflective, Hierarchical Agents for Large-Scale API Calls
Yu Du (Tsinghua University), Hongyang Zhang (University of Waterloo)
TransformerLarge Language ModelAgentic AIBenchmarkChain-of-Thought
🎯 What it does: A self-reflective hierarchical agent system named AnyTool has been developed, utilizing over 16,000 RapidAPI interfaces to address user queries without the need for additional training, employing GPT-4 function calls to implement API retrieval, problem-solving, and self-reflection loops.
Applying language models to algebraic topology: generating simplicial cycles using multi-labeling in Wu's formula
Kirill Brilliantov (ETH Zurich), German Magai (HSE University)
GenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper uses a language model to generate words for the cross regular closure in Wu's formula, thereby obtaining the simple cycles of the spherical homotopy groups.
Approximate Nearest Neighbor Search with Window Filters
Joshua Engels (Massachusetts Institute of Technology), Julian Shun (Massachusetts Institute of Technology)
RetrievalOptimizationImage
🎯 What it does: An efficient indexing and querying framework for approximate nearest neighbor search with window filtering (Window Search) is proposed and implemented, addressing scenarios where each data point has numerical labels and needs to be filtered by a window.
APT: Adaptive Pruning and Tuning Pretrained Language Models for Efficient Training and Inference
Bowen Zhao (University of Washington), Qingqing Cao (Apple)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The Adaptive Pruning and Tuning (APT) method is proposed for fine-tuning and inference processes of pre-trained language models, achieving adaptive pruning and tuning parameter adjustments during training, significantly enhancing training and inference efficiency.
AquaLoRA: Toward White-box Protection for Customized Stable Diffusion Models via Watermark LoRA
Weitao Feng (University of Science and Technology of China), Nenghai Yu (University of Science and Technology of China)
GenerationData SynthesisSafty and PrivacySupervised Fine-TuningDiffusion modelImageText
🎯 What it does: Proposes AquaLoRA, which utilizes the LoRA module to embed watermarks into the U-Net structure of Stable Diffusion, achieving white-box protection;
ArCHer: Training Language Model Agents via Hierarchical Multi-Turn RL
Yifei Zhou (University of California), Aviral Kumar (Google Deepmind)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: A hierarchical multi-round reinforcement learning framework called ArCHer is designed to train large language model agents for efficient RL in multi-round decision-making tasks.
Arrows of Time for Large Language Models
Vassilis Papadopoulos (EPFL), Clément Hongler (EPFL)
TransformerLarge Language ModelText
🎯 What it does: This paper studies the learning differences of autoregressive large language models in the temporal direction and discovers a systematic 'arrow of time' phenomenon under multilingual, multimodel, and varying context lengths;
ArtWhisperer: A Dataset for Characterizing Human-AI Interactions in Artistic Creations
Kailas Vodrahalli (Stanford University), James Zou (Stanford University)
GenerationDiffusion modelImageText
🎯 What it does: Collected and analyzed human interaction data in text-to-image models, gathering 51,026 interactions through an online game.
Assessing Large Language Models on Climate Information
Jannis Bulian (Google DeepMind), Nadine Strauss
TransformerLarge Language ModelPrompt EngineeringTextRetrieval-Augmented Generation
🎯 What it does: A multidimensional evaluation framework based on scientific communication research has been constructed and validated to measure the performance of large language models in climate information question answering.
Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications
Boyi Wei (Princeton University), Peter Henderson (Princeton University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: By pruning and low-rank modification of a large language model (Llama2-chat), this paper identifies and isolates sparse weight regions that are crucial for safety alignment at both the neuron and rank levels, and evaluates the impact of these regions on the model's safety and practicality.
AST-T5: Structure-Aware Pretraining for Code Generation and Understanding
Linyuan Gong (University of California), Alvin Cheung (University of California)
GenerationAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This paper proposes AST-T5, a pre-training framework based on Abstract Syntax Tree (AST) structure for code generation, translation, and understanding tasks.
Asymmetry in Low-Rank Adapters of Foundation Models
Jiacheng Zhu (Massachusetts Institute of Technology), Justin Solomon (Massachusetts Institute of Technology)
Domain AdaptationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the asymmetry of the two matrices A and B in LoRA low-rank adapters and proposes a method that only updates B while randomly fixing A to enhance parameter efficiency and generalization.
Asymptotically Optimal and Computationally Efficient Average Treatment Effect Estimation in A/B testing
VIKAS DEEP, Sandeep Kumar Juneja
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper studies how to minimize the required sample size in A/B testing under the premise of a given confidence interval width ε and confidence level 1-δ, in order to obtain a confidence interval for the average treatment effect (ATE) through adaptive experimental design. Two adaptive strategies, P1 and P2, are proposed, and it is proven that they achieve the lower bound of sample size in extreme cases. Numerical experiments validate that P2 still outperforms traditional uniform random allocation and the Clip-OGD strategy in non-extreme situations.
Asymptotics of feature learning in two-layer networks after one gradient-step
Hugo Cui (Ecole Polytechnique Federale de Lausanne), Bruno Loureiro (Ecole Normale Superieure)
🎯 What it does: The study investigates how a two-layer neural network learns features and significantly improves generalization performance in the high-dimensional scaling limit after a single large step of gradient descent.
Asymptotics of Learning with Deep Structured (Random) Features
Dominik Schröder, Bruno Loureiro (Ecole Normale Superieure)
TabularSequential
🎯 What it does: This paper studies the test error of ridge regression using deep structured random features (Gaussian rainbow networks) in the high-dimensional ratio limit and provides its precise asymptotic expression.
ATraDiff: Accelerating Online Reinforcement Learning with Imaginary Trajectories
Qianlan Yang (University of Illinois), Yu-Xiong Wang (University of Illinois)
Reinforcement LearningDiffusion modelTabular
🎯 What it does: Utilize offline data to train diffusion models to generate complete trajectories, injecting them into the online RL replay buffer as a data augmentation method, thereby significantly improving online training efficiency and sample utilization.
Attack-free Evaluating and Enhancing Adversarial Robustness on Categorical Data
Yujun Zhou (University of Notre Dame), Xiangliang Zhang (University of Notre Dame)
ClassificationAdversarial AttackTransformerTabular
🎯 What it does: This paper studies the evaluation and enhancement of adversarial robustness on categorical attributes (discrete features), proposing a non-attack evaluation metric IGSG and a regularization training method based on IGSG, called IGSG-reg.
Attention Meets Post-hoc Interpretability: A Mathematical Perspective
Gianluigi Lopardo (Universite Cote d'Azur), Damien Garreau (Julius Maximilians University of Wurzburg)
Explainability and InterpretabilityTransformerText
🎯 What it does: This paper presents a rigorous mathematical derivation of a single-layer multi-head attention network, proposing closed-form or approximate expressions for attention weights, gradient-based explanations, and LIME explanations, and systematically compares the essential differences among the three explanation methods.
AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers
Reduan Achtibat (Fraunhofer Heinrich-Hertz-Institute), Wojciech Samek (Fraunhofer Heinrich-Hertz-Institute)
Explainability and InterpretabilityTransformerLarge Language ModelMixture of ExpertsImageText
🎯 What it does: An Attention-aware Layer-wise Relevance Propagation (AttnLRP) method is proposed for the Transformer model, which provides reliable interpretability for the entire model (including attention layers and hidden layers) in a single backpropagation.
AttNS: Attention-Inspired Numerical Solving For Limited Data Scenarios
Zhongzhan Huang (Sun Yat-sen University), Liang Lin (Sun Yat-sen University)
OptimizationTime SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: An AI-Hybrid numerical solver based on an attention mechanism (AttNS) is designed to enhance the generalization and robustness of ODE solving in limited data scenarios.
Attribute Based Interpretable Evaluation Metrics for Generative Models
Dongkyun Kim (Yonsei University), Youngjung Uh (Yonsei University)
GenerationExplainability and InterpretabilityVision Language ModelDiffusion modelImage
🎯 What it does: An interpretable evaluation method based on attribute distribution is proposed to measure the difference in attribute intensity distribution between images generated by the model and the training set.
Auctionformer: A Unified Deep Learning Algorithm for Solving Equilibrium Strategies in Auction Games
Kexin Huang (Alibaba Group), Xiang Wang (University of Science and Technology of China)
TransformerTabular
🎯 What it does: This paper presents Auctionformer, a unified deep learning framework based on Transformer, designed to solve (approximate) Bayesian Nash equilibrium strategies in various auction games.
Audio Flamingo: A Novel Audio Language Model with Few-Shot Learning and Dialogue Abilities
Zhifeng Kong (NVIDIA), Bryan Catanzaro (NVIDIA)
RecognitionGenerationRetrievalTransformerLarge Language ModelMultimodalityRetrieval-Augmented GenerationAudio
🎯 What it does: Audio Flamingo is proposed, an audio language model that combines audio understanding, zero/few-shot learning, and multi-turn dialogue.
Auditing Private Prediction
Karan Chadha (Stanford University), Milad Nasr (Google DeepMind)
Safty and PrivacyImageText
🎯 What it does: An auditing framework for private prediction algorithms is proposed, and the privacy leakage is evaluated by instantiating different attackers (natural, poisoning, query control).
Augmenting Decision with Hypothesis in Reinforcement Learning
Nguyen Minh Quang (Singapore Management University), Hady W. Lauw (Singapore Management University)
Reinforcement LearningTabular
🎯 What it does: In value-based reinforcement learning, the authors found that early training stages have low utilization and are sensitive to critic bias. They subsequently proposed incorporating a 'hypothesis' representation (weak-model and adaptive rollout) into the actor/critic framework and built the ALH algorithm on top of TD3 to address these issues.
Autaptic Synaptic Circuit Enhances Spatio-temporal Predictive Learning of Spiking Neural Networks
Lihao Wang (Peking University), Zhaofei Yu (Peking University)
Spiking Neural NetworkVideo
🎯 What it does: A spatiotemporal self-connected recurrent neural network (STC-LIF) model based on autaptic structure is proposed to enhance the learning effectiveness of spiking neural networks in spatiotemporal prediction tasks.
Auto-Encoding Morph-Tokens for Multimodal LLM
Kaihang Pan (Zhejiang University), Hanwang Zhang (Skywork AI)
GenerationData SynthesisTransformerLarge Language ModelVision Language ModelAuto EncoderGenerative Adversarial NetworkImageTextMultimodality
🎯 What it does: Proposes Morph-Tokens, which first encodes images into abstract visual prompts and then decodes them into complete visual tokens, enabling multimodal LLMs to possess both understanding and generation capabilities;
Auto-Linear Phenomenon in Subsurface Imaging
Yinan Feng (University of North Carolina at Chapel Hill), Youzuo Lin (University of North Carolina at Chapel Hill)
Domain AdaptationRepresentation LearningSupervised Fine-TuningAuto EncoderImage
🎯 What it does: The Auto-Linear phenomenon is proposed, using self-supervised training of independent encoder/decoder and completing forward and inverse modeling of full waveform inversion through linear mapping.
Auto-Regressive Next-Token Predictors are Universal Learners
eran malach
GenerationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: The research demonstrates that autoregressive next-word prediction models (especially linear predictors) can approximate any Turing-computable function when trained on Chain-of-Thought (CoT) data, and introduces the concept of length complexity to measure the length of intermediate steps required for learning.
Autoencoding Conditional Neural Processes for Representation Learning
Victor Prokhorov (University of Edinburgh), Siddharth N
GenerationRepresentation LearningConvolutional Neural NetworkAuto EncoderImage
🎯 What it does: This paper proposes PPS-VAE, a variational autoencoding framework that adaptively infers a context pixel set, utilizing CNP for sparse observation and reconstruction of images.
Autoformalizing Euclidean Geometry
Logan Murphy (University of Toronto), Xujie Si (University of Toronto)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes a neural-symbolic framework that combines large language models (LLMs) with SMT solvers to automatically convert natural language proofs of Euclidean geometry into Lean proofs and assess their semantic correctness automatically.
Automated Evaluation of Retrieval-Augmented Language Models with Task-Specific Exam Generation
Gauthier Guinet (Amazon), Laurent Callot (Amazon)
RetrievalOptimizationTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: Automatically generate multiple-choice exams and use them to assess the task-specific accuracy of retrieval-augmented language models (RAG).
Automated Loss function Search for Class-imbalanced Node Classification
Xinyu Guo (Xidian University), Jing Liu (Xidian University)
ClassificationOptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: This paper proposes AutoLINC, an automatic loss function search framework for class-imbalanced node classification, which can automatically construct suitable loss functions based on graph structure and class distribution.
Automated Statistical Model Discovery with Language Models
Michael Y. Li (Stanford University), Noah Goodman
Large Language ModelTabularTime SeriesOrdinary Differential Equation
🎯 What it does: A statistical model discovery framework (BoxLM) is proposed that utilizes large language models to automatically generate and evaluate probabilistic programs, achieving a cycle of model building and critique.
Automating the Selection of Proxy Variables of Unmeasured Confounders
Feng Xie (Beijing Technology and Business University), Zhi Geng (Beijing Technology and Business University)
Biomedical Data
🎯 What it does: This study proposes a method for automatically identifying and utilizing proxy variables (negative control variables) to estimate causal effects in the presence of multiple unmeasured confounding variables, extending the proxy estimator for a single confounding variable, and provides identifiability conditions based on rank constraints and higher-order statistics along with corresponding algorithms.
Autonomous Sparse Mean-CVaR Portfolio Optimization
Yizun Lin (Jinan University), Cheng Li (Jinan University)
OptimizationTabularFinance Related
🎯 What it does: This paper proposes an Autonomous Sparse Mean-CVaR Portfolio Optimization Model (ASMCVaR), which efficiently solves NP-hard problems by transforming the 0-constraint into an indicator function and using tail approximation, combined with the PALM and nested FPPA algorithms.
AutoOS: Make Your OS More Powerful by Exploiting Large Language Models
Huilai Chen (Institute of Computing Technology, Chinese Academy of Sciences), Yunji Chen (Institute of Computing Technology, Chinese Academy of Sciences)
OptimizationTransformerLarge Language ModelTabular
🎯 What it does: Automating the customization and optimization of Linux kernel configurations using large language models (LLM) to enhance the performance of AIoT devices and general systems.
Averaging $n$-step Returns Reduces Variance in Reinforcement Learning
Brett Daley (University of Alberta), Marlos C. Machado (University of Alberta)
Reinforcement LearningSequential
🎯 What it does: This paper studies and proves that composite returns (such as λ returns) have lower variance under the same contraction modulus, and proposes a low-cost Piecewise λ return (Pilar) to approximate λ returns; it also provides a finite sample analysis under linear function approximation and experimentally validates its improvement in sample efficiency.
BadPart: Unified Black-box Adversarial Patch Attacks against Pixel-wise Regression Tasks
Zhiyuan Cheng (Purdue University), Xiangyu Zhang (Purdue University)
Depth EstimationAutonomous DrivingAdversarial AttackOptical FlowImage
🎯 What it does: A unified black-box adversarial patch attack framework is proposed for pixel-level regression tasks (such as monocular depth estimation and optical flow estimation), which can deploy a unified patch on any image through query optimization;
BAGEL: Bootstrapping Agents by Guiding Exploration with Language
Shikhar Murty (Stanford University), Kenton Lee (Google Deepmind)
TransformerLarge Language ModelReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: An unsupervised process is constructed to generate executable human demonstrations through multi-round labeling and resampling of randomly explored trajectories using a language model.
Bagged Deep Image Prior for Recovering Images in the Presence of Speckle Noise
Xi Chen (Rutgers University), Shirin Jalali (Rutgers University)
RestorationImage
🎯 What it does: This study investigates the recovery of images contaminated by multiple speckle noise in a multi-view low-sampling coherent imaging system using Deep Image Prior (DIP) and maximum likelihood estimation, and provides a theoretical upper bound on the mean square error.
Balanced Data, Imbalanced Spectra: Unveiling Class Disparities with Spectral Imbalance
Chiraag Kaushik (Georgia Institute of Technology), Eva L Dyer (Georgia Institute of Technology)
ClassificationConvolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: Proposes the concept of spectral imbalance and studies its impact on the inter-class error differences of classification models.
Balanced Resonate-and-Fire Neurons
Saya Higuchi (University of Lubeck), Sebastian Otte (University of Lubeck)
Spiking Neural NetworkTime SeriesSequentialBiomedical DataElectrocardiogram
🎯 What it does: Proposed and implemented a Balanced Resonance-Firing (BRF) neuron and its application in Recursive Spiking Neural Networks (RSNN) to address the divergence and high-frequency firing issues in traditional resonance-firing models during learning.
Balancing Feature Similarity and Label Variability for Optimal Size-Aware One-shot Subset Selection
Abhinab Acharya (Rochester Institute of Technology), Xumin Liu (Rochester Institute of Technology)
OptimizationData-Centric LearningImage
🎯 What it does: This paper proposes a size-aware method for one-size-fits-all subset selection called BOSS, which constructs an optimal subset by balancing feature similarity and label variability to enhance data-efficient training.
Balancing Similarity and Complementarity for Federated Learning
Kunda Yan (Tsinghua University), Changshui Zhang (Tsinghua University)
Federated LearningImageMultimodality
🎯 What it does: This paper studies the balance between model similarity and feature complementarity in federated learning and proposes the FedSaC framework to build a collaborative network.
Barrier Algorithms for Constrained Non-Convex Optimization
Pavel Dvurechensky (Weierstrass Institute for Applied Analysis and Stochastics), Mathias Staudigl (University of Mannheim)
Optimization
🎯 What it does: For non-convex optimization problems with general convex set constraints and linear equalities, first-order and second-order adaptive interior point algorithms are constructed. Utilizing the local geometry of self-concordant barrier functions, it is proven that under certain smoothness assumptions, ε-KKT and (ε, ε₁/₂)-2KKT approximate optimal points can be obtained in O(ε⁻²) and O(ε⁻³/²) iterations, respectively.
BAT: Learning to Reason about Spatial Sounds with Large Language Models
Zhisheng Zheng (University of Texas at Austin), David Harwath (University of Texas at Austin)
TransformerLarge Language ModelAudio
🎯 What it does: Designed and trained the BAT model, which integrates the binaural spatial audio perception encoder SPATIAL-AST with the large language model LLaMA-2, capable of detecting, locating, estimating distance, and reasoning about multi-source audio in a 3D environment.
Batch and match: black-box variational inference with a score-based divergence
Diana Cai (Flatiron Institute), Lawrence K. Saul (Columbia University)
OptimizationScore-based ModelImage
🎯 What it does: A black-box variational inference method based on score matching, called Batch and Match (BaM), is proposed, achieving efficient optimization of the complete covariance Gaussian variational family through closed-form updates and a stochastic proximal point algorithm.
Batch Singular Value Polarization and Weighted Semantic Augmentation for Universal Domain Adaptation
WangZiQi, Cong Wang (Sun Yat-sen University)
Domain AdaptationConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: A new framework BSP-WSA is proposed for the scenario of Universal Domain Adaptation (UniDA), which utilizes adversarial classifiers to identify unknown categories and align feature distributions, while further enhancing performance through Batch Singular Value Polarization (BSP) and Weighted Semantic Augmentation (WSA).
Bayesian Adaptation of Network Depth and Width for Continual Learning
Jeevan Thapa (Rochester Institute of Technology), Rui Li (Rochester Institute of Technology)
Convolutional Neural NetworkImage
🎯 What it does: This paper proposes a non-parametric Bayesian framework that jointly infers the depth and width of a network and dynamically expands the model structure in continual learning tasks, using a Beta process to control the number of layers and employing its conjugate Bernoulli process for drop-connect width regularization.
Bayesian Design Principles for Offline-to-Online Reinforcement Learning
Hao Hu (Tsinghua University), Chongjie Zhang (Washington University in St. Louis)
Reinforcement LearningTabular
🎯 What it does: This paper proposes an offline-to-online reinforcement learning framework based on Bayesian probability matching (BOORL). It achieves a balance between exploration and exploitation by using bootstrapped ensemble training to bias conservative policies during the offline phase, and then constructing posterior distributions and sampling actions based on softened Q-values during the online phase.
Bayesian Exploration Networks
Mattie Fellows (University of Oxford), Shimon Whiteson (University of Oxford)
Reinforcement Learning
🎯 What it does: A novel Bayesian non-parametric reinforcement learning method called Bayesian Exploration Network (BEN) is proposed, which can learn the true Bayesian optimal policy.
Bayesian Knowledge Distillation: A Bayesian Perspective of Distillation with Uncertainty Quantification
Luyang Fang (University of Georgia), Ping Ma (University of Georgia)
Knowledge DistillationImage
🎯 What it does: Proposes Bayesian Knowledge Distillation (BKD), viewing knowledge distillation as Bayesian inference with teacher information priors, and achieving uncertainty quantification of the student model through posterior sampling.
Bayesian Optimization of Function Networks with Partial Evaluations
Poompol Buathong (Cornell University), Peter I. Frazier (Cornell University)
OptimizationDrug DiscoveryTabular
🎯 What it does: A Bayesian optimization method for Partial Evaluation Function Networks (PFN) is proposed, which efficiently evaluates any node in the network through a new knowledge gradient (p-KGFN) function acquisition, significantly improving search efficiency within a limited budget.
Bayesian Power Steering: An Effective Approach for Domain Adaptation of Diffusion Models
Ding Huang (Hong Kong Polytechnic University), Jian Huang (Hong Kong Polytechnic University)
GenerationDomain AdaptationDiffusion modelImage
🎯 What it does: By introducing a Bayesian fine-tuning framework and a new network structure called Bayesian Dynamic Steering (BPS) into large-scale diffusion models, efficient transfer and adaptation to low-probability spaces for specific tasks are achieved.
Bayesian Program Learning by Decompiling Amortized Knowledge
Alessandro B. Palmarini (Santa Fe Institute), Siddharth N
Neural Architecture SearchBenchmark
🎯 What it does: Developed the DREAMDECOMPILER (Dream Decompiling) method, which uses the 'compiled' knowledge from a neural search strategy to guide the selection of library functions, thereby reducing both the breadth and depth of the search.
Bayesian Regret Minimization in Offline Bandits
Marek Petrik (University of New Hampshire), Mohammad Ghavamzadeh (Amazon AGI)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes an algorithm named BRMOB, which directly minimizes Bayesian risk in offline linear multi-armed bandits to achieve risk sensitivity in decision-making.
Bayesian Uncertainty for Gradient Aggregation in Multi-Task Learning
Idan Achituve (Sony Semiconductor Israel), Ethan Fetaya (Bar-Ilan University)
ClassificationOptimizationGaussian SplattingBiomedical Data
🎯 What it does: This paper treats the weights of the last linear layer of each task as a probability distribution in multi-task learning, thereby introducing Bayesian uncertainty into the gradients, and proposes a gradient aggregation method based on Gaussian distribution called BayesAgg-MTL.
BayOTIDE: Bayesian Online Multivariate Time Series Imputation with Functional Decomposition
Shikai Fang (University of Utah), Liang Sun (Alibaba Group)
Anomaly DetectionOptimizationTime SeriesStochastic Differential Equation
🎯 What it does: A Bayesian online multivariate time series imputation method named BayOTIDE is proposed, which utilizes function decomposition to split the series into trend and periodic factors, and achieves continuous time modeling through Gaussian process priors.
BBox-Adapter: Lightweight Adapting for Black-Box Large Language Models
Haotian Sun (Georgia Tech), Bo Dai (Georgia Tech)
Domain AdaptationOptimizationTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: A lightweight adapter BBOX-ADAPTER is proposed for black-box large language models (such as GPT-3.5, Mixtral) to enable adaptive improvements for specific tasks without model parameters or output probabilities.
Be Your Own Neighborhood: Detecting Adversarial Examples by the Neighborhood Relations Built on Self-Supervised Learning
Zhiyuan He (Chinese University of Hong Kong), Tsung-Yi Ho (Chinese University of Hong Kong)
Anomaly DetectionRepresentation LearningAdversarial AttackContrastive LearningImage
🎯 What it does: The BEYOND framework is proposed, which utilizes self-supervised learning (SSL) models to detect the consistency of features and labels of the input and its various augmented samples in order to identify adversarial examples.
BECoTTA: Input-dependent Online Blending of Experts for Continual Test-time Adaptation
Daeun Lee (Korea University), Sung Ju Hwang (Korea Advanced Institute of Science and Technology)
SegmentationDomain AdaptationAutonomous DrivingTransformerMixture of ExpertsImage
🎯 What it does: The BECoTTA framework is proposed, utilizing input-dependent low-rank expert Mixture-of-Domain Experts (MoDE) to achieve adaptive continuous testing, balancing memory retention and adaptation efficiency.
Behavior Generation with Latent Actions
Seungjae Lee (New York University), Lerrel Pinto (New York University)
GenerationAutonomous DrivingRobotic IntelligenceTransformerReinforcement LearningMultimodality
🎯 What it does: This paper proposes a behavior generation model based on residual vector quantization, VQ-BeT, which uses a Transformer to predict discretized action codes, thereby achieving high-quality generation of multimodal continuous actions.
BeigeMaps: Behavioral Eigenmaps for Reinforcement Learning from Images
Sandesh Adhikary (University of Washington), Byron Boots (University of Washington)
Representation LearningReinforcement LearningImage
🎯 What it does: A new RL representation learning method called BeigeMaps is proposed, which transforms behavioral distances into Laplacian feature mappings, replacing traditional isometric objectives to make representations more focused on local structures.
Benchmarking and Building Long-Context Retrieval Models with LoCo and M2-BERT
Jon Saad-Falcon (Stanford University), Christopher Re (Stanford University)
RetrievalTransformerLarge Language ModelTextBenchmark
🎯 What it does: A long text retrieval benchmark LoCoV1 is proposed, and an M2-BERT retrieval encoder capable of handling 32K tokens long text is constructed.
Benchmarking Deletion Metrics with the Principled Explanations
Yipei Wang (Purdue University), Xiaoqian Wang (Purdue University)
OptimizationExplainability and InterpretabilityConvolutional Neural NetworkImageBenchmark
🎯 What it does: A new TRACE framework is proposed to find the optimal feature deletion order under deletion/insertion metrics, thereby obtaining the most interpretable feature importance ranking, and it is used to comprehensively evaluate and benchmark various settings of deletion metrics.
Benign Overfitting in Adversarial Training of Neural Networks
Yunjuan Wang (Johns Hopkins University), Raman Arora (Johns Hopkins University)
OptimizationAdversarial AttackGenerative Adversarial NetworkTabular
🎯 What it does: This study investigates adversarial training of two-layer neural networks under high-dimensional mixed distributions with noise and demonstrates the existence of benign overfitting.
Benign Overfitting in Two-Layer ReLU Convolutional Neural Networks for XOR Data
Xuran Meng (University of Hong Kong), Yuan Cao (University of Hong Kong)
ClassificationOptimizationConvolutional Neural NetworkTabular
🎯 What it does: This study investigates the benign overfitting phenomenon of two-layer ReLU convolutional neural networks on XOR-type data (with label-flipping noise) and provides upper and lower bounds for training loss convergence and testing error.
Bespoke Non-Stationary Solvers for Fast Sampling of Diffusion and Flow Models
Neta Shaul (Weizmann Institute of Science), Yaron Lipman (Meta)
GenerationData SynthesisOptimizationKnowledge DistillationDiffusion modelImageTextOrdinary Differential EquationAudio
🎯 What it does: A Bespoke Non-Stationary (BNS) solver is proposed, which is a solver distillation method for diffusion and flow models that can efficiently sample within a small number of function evaluations.
Best Arm Identification for Stochastic Rising Bandits
Marco Mussi (Politecnico di Milano), Alberto Maria Metelli (Politecnico di Milano)
🎯 What it does: This paper studies the fixed budget best arm identification (BAI) problem in the Stochastic Rising Bandit (SRB) environment, proposing two algorithms R-UCBE (UCB-based) and R-SR (Successive Reject) and providing theoretical upper bounds for error probability and simple reward.
Best of Both Worlds Guarantees for Smoothed Online Quadratic Optimization
Neelkamal Bhuyan (Georgia Institute of Technology), Adam Wierman (California Institute of Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper studies the online quadratic optimization problem (SOQO) in the presence of smooth costs (switching cost), proposing a lazy adaptive interpolation algorithm (LAI) that achieves near-optimal performance in random environments, and the best-dual world algorithm LAI(γ) that maintains an approximately optimal competitive ratio in adversarial environments;
Better & Faster Large Language Models via Multi-token Prediction
Fabian Gloeckle (Meta), Gabriel Synnaeve (Meta)
GenerationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: Improvements have been made to the training method of large language models by adopting multi-step prediction (predicting multiple future tokens at once) instead of the traditional single-step prediction, thereby enhancing sample efficiency and model performance.
Better Locally Private Sparse Estimation Given Multiple Samples Per User
Yuheng Ma (Renmin University of China), Hanfang Yang (Renmin University of China)
OptimizationSafty and PrivacyTabular
🎯 What it does: This paper studies sparse linear regression under user-level local differential privacy (ULDP) and proposes a new framework that eliminates the linear dependence on dimension d by selecting candidate variables and estimating in a low-dimensional space.
Better Safe than Sorry: Pre-training CLIP against Targeted Data Poisoning and Backdoor Attacks
Wenhan Yang (University of California), Baharan Mirzasoleiman (University of California)
Adversarial AttackData-Centric LearningContrastive LearningImageText
🎯 What it does: This paper proposes SAFECLIP, which can defend against targeted data poisoning and backdoor attacks during the pre-training of CLIP while maintaining model performance.
BetterV: Controlled Verilog Generation with Discriminative Guidance
Zehua PEI, Bei Yu (Chinese University of Hong Kong)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: A BetterV framework is constructed, which achieves automatic generation of Verilog code and optimization of downstream EDA tasks by fine-tuning LLMs with instructions and combining a generative discriminator.
Beyond Chinchilla-Optimal: Accounting for Inference in Language Model Scaling Laws
Nikhil Sardana (MosaicML), Jonathan Frankle
TransformerLarge Language ModelText
🎯 What it does: Considering the inference cost, we improve the Chinchilla scaling law to calculate the optimal parameters and data scale for the total cost of training and inference, and validate that the model continues to improve through experiments with 47 extreme token/parameter ratio models.
Beyond ELBOs: A Large-Scale Evaluation of Variational Methods for Sampling
Denis Blessing (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
OptimizationDiffusion modelFlow-based ModelTabularBenchmark
🎯 What it does: This paper constructs a unified evaluation framework and conducts large-scale experiments on 17 variational sampling methods across 12 target distributions.
Beyond Implicit Bias: The Insignificance of SGD Noise in Online Learning
Nikhil Vyas (Harvard University), Boaz Barak (Harvard University)
OptimizationConvolutional Neural NetworkTransformerImage
🎯 What it does: Compare the implicit bias effects of different batch sizes (SGD noise levels) in online learning (one round of training) versus offline learning (multiple rounds of training), and propose and validate the 'golden path' hypothesis, which states that in online training, SGD only adds noise along the gradient descent path without altering the final model.
Beyond Individual Input for Deep Anomaly Detection on Tabular Data
Hugo Thimonier (University Paris-Saclay), Bich-Liên DOAN
Anomaly DetectionTransformerTabular
🎯 What it does: A deep anomaly detection method based on Non-Parametric Transformer (NPT) is proposed, which generates anomaly scores by reconstructing the features of normal samples through masking in the training set.
Beyond Point Prediction: Score Matching-based Pseudolikelihood Estimation of Neural Marked Spatio-Temporal Point Process
Zichong Li (Georgia Institute of Technology), Hongyuan Zha (Chinese University of Hong Kong)
TransformerScore-based ModelTime SeriesSequentialBiomedical DataFinance RelatedStochastic Differential Equation
🎯 What it does: This paper proposes a pseudo-likelihood estimation method based on Score Matching, called SMASH, for learning labeled spatiotemporal point processes. It can provide confidence intervals/regions for event times and locations, as well as calibrated probabilities for event labels.
Beyond Regular Grids: Fourier-Based Neural Operators on Arbitrary Domains
Levi E. Lingsch (ETH Zurich), Siddhartha Mishra (ETH Zurich)
Computational EfficiencyPoint Cloud
🎯 What it does: A Direct Spectral Evaluation (DSE) method is proposed, which efficiently computes truncated Fourier (or spherical harmonic) transforms using matrix multiplication on arbitrary point distributions, thereby extending neural operators (such as FNO, UFNO, FFNO, SFNO) to handle non-uniform grids, point clouds, and spherical data.
Beyond Sole Strength: Customized Ensembles for Generalized Vision-Language Models
Zhihe Lu (National University of Singapore), Xinchao Wang (National University of Singapore)
ClassificationRecognitionDomain AdaptationTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: The paper proposes three ensemble strategies based on pre-trained vision-language models (CLIP), utilizing the collaboration of weak and strong models to enhance the zero-shot, few-shot, and cross-dataset generalization performance.
Beyond the Calibration Point: Mechanism Comparison in Differential Privacy
Georgios Kaissis (Technical University of Munich), Daniel Rueckert (Technical University of Munich)
Safty and PrivacyImage
🎯 What it does: A new method for comparing privacy mechanisms is proposed, which can quantify the 'additional privacy risk' when two mechanisms are the same at a single (p, ε, δ) point but have different overall privacy performances, and this method is used to evaluate and select DP-SGM (DP-SGD) parameters.
Beyond the Federation: Topology-aware Federated Learning for Generalization to Unseen Clients
Mengmeng Ma (University of Delaware), Xi Peng (University of Delaware)
OptimizationFederated LearningBiomedical DataElectronic Health Records
🎯 What it does: A client topology-based federated learning framework TFL is proposed, which utilizes a client relationship graph to identify representative influential clients, thereby enhancing the model's generalization ability on unseen clients (OOF).
Beyond the Norms: Detecting Prediction Errors in Regression Models
Andres Altieri (Universite Paris-Saclay), Pablo Piantanida (McGill University)
Anomaly DetectionOptimizationTabular
🎯 What it does: This paper studies the detection of unreliable behavior in regression models, particularly the identification of prediction errors, and proposes a data-driven method to assess the reliability of regressors.
Beyond the ROC Curve: Classification Trees Using Cost-Optimal Curves, with Application to Imbalanced Datasets
Magzhan Gabidolla (University of California), Miguel Á. Carreira-Perpiñán (University of California)
ClassificationOptimizationTabular
🎯 What it does: A 'Cost Optimal Curve (COC)' method is proposed and implemented for the imbalanced binary classification problem, constructing interpretable skewed decision trees by directly optimizing the weighted 0/1 loss, and the first algorithm capable of iteratively globally reducing this loss is provided.
Bias of Stochastic Gradient Descent or the Architecture: Disentangling the Effects of Overparameterization of Neural Networks
Amit Peleg (University of Tübingen), Matthias Hein (University of Tübingen)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: The study investigates the generalization differences between Stochastic Gradient Descent (SGD) and random sampling (G&C) under zero training error when neural networks are over-parameterized;
Bidirectional Reciprocative Information Communication for Few-Shot Semantic Segmentation
Yuanwei Liu (Northwestern Polytechnical University), Fahad Shahbaz Khan (Linkoping University)
SegmentationConvolutional Neural NetworkImage
🎯 What it does: The IFRNet framework is proposed, which uses bidirectional recursive information transmission to address the intra-class diversity problem in few-shot semantic segmentation.
BiE: Bi-Exponent Block Floating-Point for Large Language Models Quantization
Lancheng Zou (Chinese University of Hong Kong), Bei Yu (Chinese University of Hong Kong)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: This paper proposes the Bi-Exponent Block Floating-Point (BiE) numerical format and designs a corresponding Post-Training Quantization (PTQ) process to quantize large language models (LLMs) to 4/3 bits while maintaining nearly unchanged accuracy.
Bifurcated Attention for Single-Context Large-Batch Sampling
Ben Athiwaratkun (Together.ai), Bing Xiang (Goldman Sachs)
GenerationComputational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: A context-aware bifurcated attention method is proposed for the incremental decoding phase in single-context large batch sampling, significantly reducing KV cache memory I/O and thereby decreasing inference latency.
Biharmonic Distance of Graphs and its Higher-Order Variants: Theoretical Properties with Applications to Centrality and Clustering
Mitchell Black (Oregon State University), Amir Nayyeri (Oregon State University)
Graph Neural NetworkGraphTabular
🎯 What it does: This study investigates the biharmonic distance and its k-order generalization for measuring the importance of edges and the theoretical properties of global topology in graphs, and applies it to edge centrality and clustering.
BiLLM: Pushing the Limit of Post-Training Quantization for LLMs
Wei Huang (University of Hong Kong), XIAOJUAN QI
TransformerLarge Language ModelText
🎯 What it does: Proposes BiLLM, a post-training 1-bit quantization method for large language models;
Binary Decomposition: A Problem Transformation Perspective for Open-Set Semi-Supervised Learning
Jun-Yi Hang (Southeast University), Min-Ling Zhang (Southeast University)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: This paper proposes a method to transform open set semi-supervised learning into a binary decomposition approach, and based on this, designs the BDMatch method.
Binning as a Pretext Task: Improving Self-Supervised Learning in Tabular Domains
Kyungeun Lee (LG AI Research), Woohyung Lim (LG AI Research)
ClassificationRepresentation LearningTransformerAuto EncoderTabular
🎯 What it does: A binning-based self-supervised pre-training task is proposed, using discretized bin indices to replace the original continuous values for reconstruction, in order to learn effective representations of tabular data.
Bipartite Matching in Massive Graphs: A Tight Analysis of EDCS
Amir Azarmehr (Northeastern University), Mohammad Roghani (Stanford University)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper reveals a factor through the construction of a factor revealing linear programming (LP) and conducts a rigorous analysis of the performance of edge-degree constrained subgraphs (EDCS) in maximum matching approximation, providing the exact approximation ratios for any parameters β and β-.
BiSHop: Bi-Directional Cellular Learning for Tabular Data with Generalized Sparse Modern Hopfield Model
Chenwei Xu (Northwestern University), Han Liu (Northwestern University)
ClassificationData-Centric LearningTabular
🎯 What it does: A bidirectional sparse Hopfield network framework, BiSHop, has been designed and implemented for deep tabular data learning.
Bivariate Causal Discovery using Bayesian Model Selection
Anish Dhir (Imperial College), Mark van der Wilk (University of Oxford)
TabularBenchmark
🎯 What it does: This paper proposes a method for identifying bivariate causal relationships using Bayesian model selection, removing the rigid constraints on data generation models found in traditional methods.
BLO-SAM: Bi-level Optimization Based Finetuning of the Segment Anything Model for Overfitting-Preventing Semantic Segmentation
Li Zhang (University of California), Pengtao Xie (University of California)
SegmentationOptimizationSupervised Fine-TuningImageBiomedical DataComputed Tomography
🎯 What it does: A dual-layer optimization-based fine-tuning method called BLO-SAM is proposed, which automatically learns prompt embeddings and mitigates overfitting through training on separate subsets, achieving semantic segmentation without manual prompts.
Block Acceleration Without Momentum: On Optimal Stepsizes of Block Gradient Descent for Least-Squares
Liangzu Peng (University of Pennsylvania), Wotao Yin (Alibaba Group)
OptimizationTabular
🎯 What it does: This paper studies the use of the Block Gradient Descent (BGD) algorithm in two decomposed least squares problems, deriving its optimal step size under the assumption of block orthogonality, and proving that under this condition, the asymptotic convergence rate of BGD is at least twice that of the Heavy Ball method (HB).
Boosting Offline Optimizers with Surrogate Sensitivity
Manh Cuong Dao (Hanoi University of Science and Technology), Trong Nghia Hoang (Washington State University)
OptimizationTabularBenchmark
🎯 What it does: This paper proposes a model sensitivity-based regularization framework called BOSS to enhance the performance of various offline optimizers;