ICML 2025 Papers — Page 4
International Conference on Machine Learning · 3257 papers
Avoiding spurious sharpness minimization broadens applicability of SAM
Sidak Pal Singh (Google Research), Yann Dauphin (Google DeepMind)
TransformerLarge Language ModelText
🎯 What it does: An improved method for Sharpness Aware Minimization (SAM) aimed at natural language processing is proposed, effectively reducing over-sharpening in language modeling tasks and avoiding pseudo-sharpening caused by the logit path.
AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
Zhengxuan Wu (Stanford University), Christopher Potts (Stanford University)
Data SynthesisRepresentation LearningTransformerLarge Language ModelSupervised Fine-TuningAuto EncoderTextBenchmark
🎯 What it does: This study proposes the AXBENCH benchmark for systematically evaluating concept detection and model-driven control methods in large language models.
B-score: Detecting biases in large language models using response history
An Vo (KAIST), Anh Totti Nguyen (Auburn University)
TransformerLarge Language ModelText
🎯 What it does: The B-score is proposed to detect and quantify bias by observing the distribution differences between single-turn and multi-turn responses of LLMs, and it verifies the effect of self-correction in multi-turn dialogues.
Backdoor Attacks in Token Selection of Attention Mechanism
Yunjuan Wang (Johns Hopkins University), Raman Arora (Johns Hopkins University)
OptimizationAdversarial AttackTransformerSequential
🎯 What it does: This paper studies how to implement backdoor attacks in the self-attention mechanism of Transformers by modifying token selection. Theoretically, it proves that single-head self-attention networks can interpolate contaminated data under gradient descent training and achieve backdoor functionality under certain conditions.
BackSlash: Rate Constrained Optimized Training of Large Language Models
Jun Wu, Yuxing Han (Tsinghua University)
CompressionOptimizationTransformerLarge Language ModelText
🎯 What it does: The BackSlash framework is proposed, which compresses LLM parameters directly through rate-distortion optimization during the training phase, resulting in models that are both small and efficient.
Balanced Learning for Domain Adaptive Semantic Segmentation
Wangkai Li (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
SegmentationDomain AdaptationImage
🎯 What it does: A balanced learning method BLDA is proposed for unsupervised domain adaptation semantic segmentation, addressing the prediction bias of different categories in the target domain;
BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing
Dongliang Guo (University of Virginia), Sheng Li (University of Virginia)
OptimizationData-Centric LearningTransformerSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: Proposes the BalancEdit method, which dynamically balances the universality and locality of multimodal model editing using a codebase;
Balancing Efficiency and Expressiveness: Subgraph GNNs with Walk-Based Centrality
Joshua Southern (Imperial College London), Fabrizio Frasca (Technion Israel Institute of Technology)
Graph Neural NetworkGraph
🎯 What it does: The paper proposes a Subgraph GNN method named HyMN, which combines walk-based centrality sampling and structural encoding to significantly reduce computational complexity and enhance expressiveness.
Balancing Interference and Correlation in Spatial Experimental Designs: A Causal Graph Cut Approach
Jin Zhu (London School of Economics and Political Science), Chengchun Shi (National University of Singapore)
OptimizationTabular
🎯 What it does: This paper proposes an algorithm based on causal graph cuts to minimize the mean squared error of the ATE estimator in experimental designs with spatial interference and correlation.
Balancing Model Efficiency and Performance: Adaptive Pruner for Long-tailed Data
Zhe Zhao (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
ClassificationCompressionComputational EfficiencyImage
🎯 What it does: An adaptive pruning framework for long-tail data, LTAP, is proposed, which achieves significant model compression and acceleration while maintaining or even improving the performance of tail classes.
Balancing Preservation and Modification: A Region and Semantic Aware Metric for Instruction-Based Image Editing
Zhuoying Li (Wangxuan Institute of Computer Technology Peking University), Yang Liu (Wangxuan Institute of Computer Technology Peking University)
Image TranslationSegmentationTransformerLarge Language ModelVision Language ModelImage
🎯 What it does: This paper proposes an evaluation metric for instruction-driven image editing called BPM, which can simultaneously assess the position and size of the edited area as well as semantic consistency, while preserving the content of the unedited areas.
Balancing the Scales: A Theoretical and Algorithmic Framework for Learning from Imbalanced Data
Corinna Cortes (Google Research), Yutao Zhong (Google Research)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A theoretical framework for class imbalance is proposed, defining a new class imbalance margin loss, and based on this loss, the IMMAX (Imbalanced Margin Maximization) algorithm is designed for binary and multi-class tasks.
BAME: Block-Aware Mask Evolution for Efficient N:M Sparse Training
Chenyi yang, Rongrong Ji (Xiamen University)
Computational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: Proposes the BAME method, achieving full sparse forward and backward propagation in N:M sparse training, utilizing block-aware mask evolution to reduce training costs.
BanditSpec: Adaptive Speculative Decoding via Bandit Algorithms
Yunlong Hou (National University of Singapore), Zhuoran Yang (Yale University)
OptimizationHyperparameter SearchTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes an adaptive speculative decoding framework called BANDITSPEC based on multi-armed bandits, which can dynamically select the best hyperparameters during LLM inference.
BAnG: Bidirectional Anchored Generation for Conditional RNA Design
Roman Klypa (University Grenoble Alpes), Sergei Grudinin (University Grenoble Alpes)
GenerationData SynthesisDrug DiscoveryTransformerBiomedical Data
🎯 What it does: The RNABAnG model and BAnG generation method are proposed, which can generate RNA sequences targeting specific proteins without the need for a large number of known RNA-protein interaction sequences or RNA structural information.
Banyan: Improved Representation Learning with Explicit Structure
Mattia Opper (University of Edinburgh), Siddharth N
Representation LearningRecurrent Neural NetworkContrastive LearningText
🎯 What it does: BANYAN learns sentence and word-level semantic representations through a self-structured encoder, utilizing an explicit hierarchical structure to construct embeddings.
BARK: A Fully Bayesian Tree Kernel for Black-box Optimization
Toby Boyne (Imperial College London), Ruth Misener (Imperial College London)
OptimizationTabular
🎯 What it does: We propose BARK—a fully Bayesian model that unifies the tree structure of BART with Gaussian process kernels, generating GP kernels through MCMC sampling of the tree structure for Bayesian optimization.
BARNN: A Bayesian Autoregressive and Recurrent Neural Network
Dario Coscia (International School of Advanced Studies), Gianluigi Rozza (International School of Advanced Studies)
GenerationData SynthesisDrug DiscoveryRecurrent Neural NetworkTime SeriesSequential
🎯 What it does: BARNN is proposed, a framework that converts any autoregressive or recurrent model into a Bayesian version, applicable to tasks such as time series, PDE solving, and molecular generation.
Batch List-Decodable Linear Regression via Higher Moments
Ilias Diakonikolas (University of Wisconsin Madison), Thanasis Pittas (University of Wisconsin Madison)
🎯 What it does: A new algorithm for list-decodable linear regression in batch form is proposed, utilizing higher-order moment information to achieve smaller batch sizes and lower estimation errors.
BaWA: Automatic Optimizing Pruning Metric for Large Language Models with Balanced Weight and Activation
Lian Liu (Advanced Micro Devices Inc), ying wang
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a new LLM quantization pruning metric, BaWA, to better balance weight and activation distributions in one-shot post-training pruning, thereby enhancing the performance of sparse models.
BaxBench: Can LLMs Generate Correct and Secure Backends?
Mark Vero (ETH Zurich), Martin Vechev (ETH Zurich)
Safty and PrivacyAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes and implements BAXBENCH, a benchmark for evaluating the functionality and security of large language models (LLMs) in generating complete backend applications, and conducts experimental evaluations on 11 mainstream LLMs using this benchmark.
Bayesian Active Learning for Bivariate Causal Discovery
Yuxuan Wang (Peking University), Yizhou Wang (Peking University)
TabularTime Series
🎯 What it does: This paper proposes an active learning framework based on Bayesian factors for bivariate causal discovery, aimed at effectively identifying the direction of causal relationships between variables through experimental interventions.
Bayesian Basis Function Approximation for Scalable Gaussian Process Priors in Deep Generative Models
Mehmet Yiğit Balık (Aalto University), Harri Lähdesmäki (Aalto University)
GenerationData SynthesisComputational EfficiencyTime SeriesSequentialBiomedical Data
🎯 What it does: A scalable Gaussian Process prior VAE model DGBFGP based on basis function approximation is proposed, which handles high-dimensional time series data with linear time complexity.
Bayesian Inference for Correlated Human Experts and Classifiers
Markelle Kelly (University of California), Padhraic Smyth (University of California)
ClassificationImage
🎯 What it does: Under the premise of given pre-trained classifier outputs, a Bayesian framework is proposed to accurately predict expert voting (consensus) with minimal human expert queries and to perform inference in online sequential queries.
Bayesian Neural Scaling Law Extrapolation with Prior-Data Fitted Networks
Dongwoo Lee (Yonsei University), Hae Beom Lee (Korea University)
TransformerImageText
🎯 What it does: This paper utilizes Prior-data Fitted Networks (PFN) to model and extrapolate uncertainty in neural network scaling laws within a Bayesian framework, addressing the limitations of traditional point estimation methods that cannot quantify prediction uncertainty.
Bayesian Optimization from Human Feedback: Near-Optimal Regret Bounds
Aya Kayal (University College London), Alberto Bernacchia (MediaTek Research)
OptimizationReinforcement Learning from Human FeedbackTabular
🎯 What it does: A multi-round learning strategy MR-LPF is proposed to find the globally optimal action from Bayesian optimization with only comparative feedback, and a tighter sub-linear regret upper bound is provided.
Bayesian Weight Enhancement with Steady-State Adaptation for Test-time Adaptation in Dynamic Environments
Jae-Hong Lee (Hankuk University of Foreign Studies)
Domain AdaptationImageStochastic Differential Equation
🎯 What it does: This study investigates Test-Time Adaptation (TTA) in dynamic environments, proposing a Bayesian Weight Enhancement Framework and a Steady-State Adaptation (SSA) algorithm to mitigate weight degradation caused by gradient noise.
BCE vs. CE in Deep Feature Learning
Qiufu Li (Shenzhen University), Linlin Shen (Shenzhen University)
ClassificationRepresentation LearningConvolutional Neural NetworkImage
🎯 What it does: This paper compares the roles of cross-entropy (CE) and binary cross-entropy (BCE) loss in deep feature learning across various classification tasks. It provides theoretical proof (showing that BCE can also lead to neural collapse) and validates through experiments that BCE can more quickly and effectively enhance the intra-class compactness and inter-class distinguishability of features, thereby improving classification performance.
BDC-CLIP: Brownian Distance Covariance for Adapting CLIP to Action Recognition
Fei Long (Dalian University of Technology), Peihua Li (Dalian University of Technology)
RecognitionTransformerVision Language ModelContrastive LearningVideoText
🎯 What it does: This paper proposes BDC-CLIP, a cross-modal alignment method for video action recognition using Brownian Distance Covariance (BDC) on CLIP.
Be a Goldfish: Forgetting Bad Conditioning in Sparse Linear Regression via Variational Autoencoders
Kuheli Pratihar (Indian Institute of Technology Kharagpur), Debdeep Mukhopadhyay (Indian Institute of Technology Kharagpur)
OptimizationAuto EncoderBiomedical Data
🎯 What it does: A variational autoencoder (VAE) framework is proposed to solve the NP-hard problem of sparse linear regression (SLR), with the core idea being to achieve sparse representation through a learnable sparse diagonal matrix and adaptive preprocessing.
Be Confident: Uncovering Overfitting in MLLM Multi-Task Tuning
Wenke Huang (Wuhan University), Dacheng Tao (Nanyang Technological University)
ClassificationRecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: In multi-task fine-tuning, we propose Noise Resilient Confidence Alignment (NRCA), which reduces the model's over-reliance on language priors and alleviates overfitting in open-ended response tasks by constructing visual inputs with Gaussian noise perturbations and aligning the noise with the word-level confidence of the normal visual branch.
BECAME: Bayesian Continual Learning with Adaptive Model Merging
Mei Li (Shanghai Jiao Tong University), Hongtao Lu (Shanghai Jiao Tong University)
ClassificationOptimizationImage
🎯 What it does: By linearly mixing the gradient projection model with the unconstrained model in continual learning, a two-stage training framework called BECAME is proposed to achieve a balance between stability and plasticity.
Behavior-agnostic Task Inference for Robust Offline In-context Reinforcement Learning
Long Ma (Peking University), Yizhou Wang (Peking University)
Reinforcement LearningSequential
🎯 What it does: A new behavior-agnostic task inference method (BATI) is proposed to improve the robustness of offline contextual reinforcement learning (ICRL), especially in the face of distribution changes.
Behavior-Regularized Diffusion Policy Optimization for Offline Reinforcement Learning
Chen-Xiao Gao (Nanjing University), Zongzhang Zhang (Nanjing University)
OptimizationReinforcement LearningDiffusion modelTabular
🎯 What it does: The BDPO algorithm is proposed, applying the behavior regularization framework to diffusion policies, achieving a combination of high expressiveness diffusion policies and behavior regularization in offline reinforcement learning.
Behavioral Exploration: Learning to Explore via In-Context Adaptation
Andrew Wagenmaker (University of California), Sergey Levine (University of California)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningSequential
🎯 What it does: The paper proposes a method called 'Behavioral Exploration (BE)', which trains a long-context generation model using offline expert demonstration data, allowing it to quickly explore and self-adjust during runtime through adaptive context (historical trajectories and coverage).
Bellman Unbiasedness: Toward Provably Efficient Distributional Reinforcement Learning with General Value Function Approximation
Taehyun Cho (Seoul National University), Jungwoo Lee (Seoul National University)
Reinforcement Learning
🎯 What it does: This paper proposes a new distributed reinforcement learning framework, defines Bellman Unbiasedness, and proves that under this framework, moment statistics are the only statistical functions that can be both closed and unbiasedly estimated. Based on this, the SF-LSVI algorithm is designed, which achieves near-optimal regret bounds in finite episodic MDPs with general value function approximation.
Benchmarking Abstract and Reasoning Abilities Through A Theoretical Perspective
Qingchuan Ma (Xiamen University), Rongrong Ji (Xiamen University)
Large Language ModelBenchmarkChain-of-Thought
🎯 What it does: This paper constructs a theoretically driven, quantifiable benchmark for rigorously assessing the capabilities of large language models in abstract reasoning.
Benchmarking Quantum Reinforcement Learning
Nico Meyer (Fraunhofer Institute for Integrated Circuits), Daniel Scherer
Reinforcement LearningBenchmark
🎯 What it does: A statistical benchmark evaluation method for quantum reinforcement learning has been designed and implemented, comparing the performance of classical and hybrid quantum models in the BeamManagement6G simulation environment.
Benefits of Early Stopping in Gradient Descent for Overparameterized Logistic Regression
Jingfeng Wu (University of California), Bin Yu (University of California)
OptimizationTabular
🎯 What it does: This study investigates the effect of early stopping in gradient descent on the regularization of models in over-parameterized logistic regression, proving that early stopping in gradient descent can eliminate excess logistic risk, while conventional gradient descent diverges in risk upon convergence.
Benign Overfitting in Token Selection of Attention Mechanism
Keitaro Sakamoto (University of Tokyo), Issei Sato (University of Tokyo)
TransformerImage
🎯 What it does: This paper provides a theoretical analysis of the training dynamics and generalization performance of Token selection in the attention mechanism of Transformers, and proves the existence of benign overfitting under label noise.
Benign Samples Matter! Fine-tuning On Outlier Benign Samples Severely Breaks Safety
Zihan Guan (University of Virginia), Anil Vullikanti (University of Virginia)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a sample selection method based on Self-Influence Score Normalization (Self-Inf-N) to identify the most destructive outlier samples for LLM safety alignment from seemingly benign datasets, and fine-tunes the model using only these 100 samples, significantly increasing the probability of generating harmful content.
Best of Both Worlds: Advantages of Hybrid Graph Sequence Models
Ali Behrouz (Google Research), Vahab Mirrokni (Google Research)
Graph Neural NetworkTransformerGraph
🎯 What it does: A unified graph sequence model framework GSM and its improved version GSM++ are proposed, achieving efficient learning of graph structures through hierarchical affinity clustering (HAC) tokenization, local GNN encoding, and a hybrid Transformer/SSM (Mamba) global encoding.
Best of Both Worlds: Regret Minimization versus Minimax Play
Adrian Müller, Volkan Cevher (EPFL)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes an online learning algorithm in zero-sum games with Bandit feedback that can maintain constant regret with respect to a given safe strategy while achieving ˜O(√T) regret relative to any fixed strategy;
Best Subset Selection: Optimal Pursuit for Feature Selection and Elimination
Zhihan Zhu (Beihang University), Yong Xia (Beihang University)
OptimizationTabularAudio
🎯 What it does: This study addresses the optimal subset selection problem, proposing optimal feature selection and elimination criteria based on the objective function, and embedding these into existing greedy algorithms to enhance recovery and interpretability performance.
BEST-Route: Adaptive LLM Routing with Test-Time Optimal Compute
Dujian Ding (University of British Columbia), Victor Rühle
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The BEST-Route framework is proposed, which dynamically selects LLMs of different scales and adjusts the best-of-n sampling number based on query difficulty to reduce inference costs while maintaining high quality.
Better to Teach than to Give: Domain Generalized Semantic Segmentation via Agent Queries with Diffusion Model Guidance
Fan Li (Northwestern Polytechnical University), Yuelei Xu (Northwestern Polytechnical University)
SegmentationDomain AdaptationAutonomous DrivingTransformerDiffusion modelImage
🎯 What it does: A diffusion model-guided agent query learning framework called QueryDiff is proposed to address the domain generalization semantic segmentation (DGSS) problem.
Beyond Atoms: Enhancing Molecular Pretrained Representations with 3D Space Modeling
Shuqi Lu (DP Technology), Guolin Ke (DP Technology)
Representation LearningDrug DiscoveryTransformerAuto EncoderGraph
🎯 What it does: The SpaceFormer framework is proposed, utilizing a grid-discretized 3D space and introducing virtual space points to enhance molecular pre-training representations.
Beyond Bradley-Terry Models: A General Preference Model for Language Model Alignment
Yifan Zhang (Tsinghua University), Quanquan Gu (University of California)
Recommendation SystemOptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: A preference embedding model (GPM) that embeds answers into a multidimensional latent space is proposed, enabling linear queries of complex non-transitive preferences.
Beyond Communication Overhead: A Multilevel Monte Carlo Approach for Mitigating Compression Bias in Distributed Learning
Ze'ev Zukerman (Technion), Kfir Yehuda Levy
OptimizationFederated LearningTransformerImageText
🎯 What it does: A gradient compression framework based on Multi-Layer Monte Carlo (MLMC) is proposed, utilizing biased compressors to generate unbiased gradient estimates, thereby eliminating compression bias in distributed learning.
Beyond Confidence: Exploiting Homogeneous Pattern for Semi-Supervised Semantic Segmentation
Rui Sun (Shenzhen International Graduate School Tsinghua University), Tianzhu Zhang (University of Science and Technology of China)
SegmentationAgentic AIImage
🎯 What it does: Proposes the AgScore scoring function, which utilizes homogeneous patterns in the embedding space to filter pseudo-labels, enhancing the robustness and accuracy of semi-supervised semantic segmentation.
Beyond Cropped Regions: New Benchmark and Corresponding Baseline for Chinese Scene Text Retrieval in Diverse Layouts
Gengluo Li (Institute of Information Engineering, Chinese Academy of Sciences), Yu Zhou (Nankai University)
RetrievalConvolutional Neural NetworkContrastive LearningImageTextBenchmark
🎯 What it does: A multi-layout Chinese scene text retrieval benchmark DL-CSVTR is designed, and a CSTR-CLIP model utilizing full image information and multi-granularity alignment is proposed.
Beyond CVaR: Leveraging Static Spectral Risk Measures for Enhanced Decision-Making in Distributional Reinforcement Learning
Mehrdad Moghimi (York University), Hyejin Ku (York University)
OptimizationReinforcement LearningTabularFinance Related
🎯 What it does: A distributed reinforcement learning-based algorithm QR-SRM is proposed, which optimizes the static spectral risk measure (SRM) and provides convergence guarantees.
Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation
Zixuan Hu (Peking University), LINGYU DUAN
Domain AdaptationGaussian SplattingImage
🎯 What it does: A method called ReCAP is proposed for Wild Testing Time Adaptation (WTTA), which replaces traditional entropy minimization with region confidence, thereby more effectively updating the model in data-scarce and multi-shift environments.
Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence
Gouki Minegishi (University of Tokyo), Yutaka Matsuo (University of Tokyo)
Meta LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes the In-Context Meta-Learning (ICML) task and clarifies how the model achieves In-Context Meta-Learning by analyzing the internal circuit evolution of a two-layer attention Transformer.
Beyond Log-Concavity and Score Regularity: Improved Convergence Bounds for Score-Based Generative Models in W2-distance
Marta Gentiloni Silveri, Antonio Ocello (Ecole Polytechnique)
GenerationData SynthesisOptimizationScore-based ModelTabularOrdinary Differential Equation
🎯 What it does: This paper proposes a new framework that utilizes the regularization of the OU process and HJB equation analysis to derive an upper bound for the convergence of the Stochastic Generative Model (SGM) in Wasserstein-2 distance, under the conditions of weak log-concavity and one-sided log-Lipschitz, without the need for traditional strong log-concavity or high-order smoothness of the fractional function.
Beyond Low-rank Decomposition: A Shortcut Approach for Efficient On-Device Learning
Le-Trung Nguyen (Institut Polytechnique de Paris), Enzo Tartaglione (Institut Polytechnique de Paris)
CompressionComputational EfficiencyConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an activation compression method based on single-shot subspace iteration (ASI), which significantly reduces the memory of activations during backpropagation and improves training efficiency.
Beyond Matryoshka: Revisiting Sparse Coding for Adaptive Representation
Tiansheng Wen (Xidian University), Chenyu You
RetrievalRepresentation LearningAuto EncoderContrastive LearningImageMultimodality
🎯 What it does: A post-processing sparse coding framework (Contrastive Sparse Representation, CSR) is proposed, which learns sparse representations on frozen pre-trained embeddings and simultaneously optimizes semantic retention and sparsity through contrastive learning and reconstruction loss.
Beyond Message Passing: Neural Graph Pattern Machine
Zehong Wang (University of Notre Dame), Yanfang Ye (University of Notre Dame)
ClassificationRepresentation LearningGraph Neural NetworkTransformerGraph
🎯 What it does: A message-passing-free graph learning framework GPM is proposed, which learns representations directly from graph substructure patterns.
Beyond Minimax Rates in Group Distributionally Robust Optimization via a Novel Notion of Sparsity
Quan M. Nguyen, Cristóbal A Guzmán
OptimizationTabular
🎯 What it does: Proposed a (λ,β)-sparsity framework and utilized sleep multi-armed bandit technology in group distributionally robust optimization to reduce sample complexity.
Beyond One-Hot Labels: Semantic Mixing for Model Calibration
Haoyang Luo (City University of Hong Kong), Chang Xu (University of Sydney)
ClassificationData SynthesisConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: Proposes the Calibration-aware Semantic Mixing (CSM) framework, which generates high-fidelity semantic mixed samples using diffusion models and re-labels them with soft labels to achieve precise calibration of deep networks.
Beyond Self-Interest: How Group Strategies Reshape Content Creation in Recommendation Platforms?
Yaolong Yu (Chinese University of Hong Kong), Sinno Jialin Pan (Chinese University of Hong Kong)
Recommendation System
🎯 What it does: A game theory analysis of the strategic behavior of content creators collaborating in groups on recommendation platforms, exploring its impact on content distribution and user welfare.
Beyond Self-Repellent Kernels: History-Driven Target Towards Efficient Nonlinear MCMC on General Graphs
Jie Hu (North Carolina State University), Do Young Eun (North Carolina State University)
OptimizationComputational EfficiencyGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: A historical driven target (HDT) framework is proposed, transferring the self-suppression mechanism from the transition kernel to the target distribution, applicable to any random walk (both reversible and irreversible) and achieving low computational cost;
Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions
Eray Erturk (University of Southern California), Joseph Futoma (Apple Inc.)
ClassificationAnomaly DetectionTransformerContrastive LearningTime SeriesBiomedical Data
🎯 What it does: A foundational model WBM based on wearable device behavior data was constructed and pre-trained, utilizing 2.5 billion hours of behavioral time series to detect and predict health status.
Beyond Task-Specific Reasoning: A Unified Conditional Generative Framework for Abstract Visual Reasoning
Fan Shi (Fudan University), Xiangyang Xue (Fudan University)
GenerationTransformerAuto EncoderImage
🎯 What it does: A Unified Conditional Generation Framework (UCGS) is proposed to address various abstract visual reasoning tasks through a single conditional generation model.
Beyond the Permutation Symmetry of Transformers: The Role of Rotation for Model Fusion
Binchi Zhang (University of Virginia), Jundong Li (University of Virginia)
ClassificationRecognitionTransformerImageText
🎯 What it does: This study investigates the rotational symmetry of the parameter space in Transformers, proposing a theoretically optimal parameter matching algorithm based on this symmetry, and applying it as a plugin for model fusion.
Beyond The Rainbow: High Performance Deep Reinforcement Learning on a Desktop PC
Tyler Clark (University Of Southampton), Jonathon Hare (University Of Southampton)
Reinforcement LearningVideo
🎯 What it does: BTR constructs an efficient and user-friendly reinforcement learning algorithm by combining six improvements of Rainbow DQN with a new architecture, regularization, distributed learning, adjustable exploration, parallelization, and adaptive pooling.
Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective
Steve Azzolin (University of Trento), Stefano Teso (University of Trento)
Explainability and InterpretabilityGraph Neural NetworkGraph
🎯 What it does: This paper conducts a formal analysis of the subgraph explanations generated by Self-Explaining Graph Neural Networks (SE-GNN) and proposes a new Dual-Channel GNN architecture to overcome the limitations of SE-GNN explanations.
Beyond Zero Initialization: Investigating the Impact of Non-Zero Initialization on LoRA Fine-Tuning Dynamics
Shiwei Li (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
TransformerSupervised Fine-TuningText
🎯 What it does: This study investigates the necessity of zero initialization during LoRA fine-tuning, proposing to initialize both A and B to non-zero values and analyzing their impact on training dynamics.
Bi-perspective Splitting Defense: Achieving Clean-Seed-Free Backdoor Security
Yangyang Shen (Southeast University), Beilun Wang (Southeast University)
ClassificationAnomaly DetectionConvolutional Neural NetworkImage
🎯 What it does: A dual-view segmentation defense (BSD) is proposed under the condition of not requiring additional sample cleaning, which trains a clean model by splitting the training set into a clean pool and a malicious pool, eliminating the impact of backdoor attacks.
BiAssemble: Learning Collaborative Affordance for Bimanual Geometric Assembly
Yan Shen (Peking University), Hao Dong (Peking University)
Robotic IntelligenceGraph Neural NetworkAuto EncoderPoint CloudBenchmark
🎯 What it does: Proposes the BiAssemble framework, which utilizes point-level collaborative utility perception to achieve long-sequence operations for two-handed geometric assembly.
Bifurcate then Alienate: Incomplete Multi-view Clustering via Coupled Distribution Learning with Linear Overhead
Shengju Yu (Hong Kong Baptist University), En Zhu (National University of Defense Technology)
OptimizationMultimodality
🎯 What it does: A new algorithm for incomplete multi-view clustering (IMC) called BACDL is proposed, which utilizes feature clustering dual separation and adversarial separation, combined with coupled distribution learning to achieve joint modeling of shared and view-specific discriminative features, and completes clustering embedding through linear time and space complexity update rules.
BILBO: BILevel Bayesian Optimization
Ruth Wan Theng Chew (National University of Singapore), Bryan Kian Hsiang Low (National University of Singapore)
Optimization
🎯 What it does: A novel two-layer Bayesian optimization algorithm BILBO is proposed, which can simultaneously optimize the upper and lower layers in noisy, constrained, and gradient-free black-box two-layer problems.
BiMaCoSR: Binary One-Step Diffusion Model Leveraging Flexible Matrix Compression for Real Super-Resolution
Kai Liu (Shanghai Jiao Tong University), Yulun Zhang (Shanghai Jiao Tong University)
RestorationSuper ResolutionCompressionDiffusion modelImage
🎯 What it does: This paper proposes BiMaCoSR, a first-order binary diffusion model that achieves extreme compression and acceleration for real super-resolution methods through low-rank and sparse matrix branches.
BiMark: Unbiased Multilayer Watermarking for Large Language Models
Xiaoyan Feng (Griffith University), Shirui Pan (Griffith University)
GenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelText
🎯 What it does: A multi-layer unbiased watermarking framework named BiMark is proposed, which can embed extractable multi-bit information while maintaining text quality, supporting model-independent detection, and meeting information capacity requirements during text generation by large language models.
Binary Hypothesis Testing for Softmax Models and Leverage Score Models
Yuzhou Gu (Institute for Advanced Study), Junze Yin (Boston University)
Information Theory
🎯 What it does: This study investigates the binary hypothesis testing problem of distinguishing between two softmax or leverage score models under energy/constraint conditions, providing precise upper and lower bounds on sample complexity.
BinauralFlow: A Causal and Streamable Approach for High-Quality Binaural Speech Synthesis with Flow Matching Models
Susan Liang (University of Rochester), Alexander Richard (Meta)
GenerationData SynthesisConvolutional Neural NetworkFlow-based ModelAudio
🎯 What it does: A causal flow-matching based binaural speech synthesis framework called BinauralFlow is proposed;
Bipartite Ranking From Multiple Labels: On Loss Versus Label Aggregation
Michal Lukasik (Google Research), Sanjiv Kumar (Google Research)
Tabular
🎯 What it does: This paper studies the binary ranking problem in a multi-label environment, providing theoretical analysis and comparing the Bayesian optimal solutions of loss aggregation and label aggregation methods.
Bivariate Causal Discovery with Proxy Variables: Integral Solving and Beyond
Yong Wu (Fudan University), Xinwei Sun (Fudan University)
Tabular
🎯 What it does: A non-parametric method is proposed to test integral equations in bivariate causal discovery using proxy variables (negative control outcomes) to address unmeasured confounding;
Black-Box Adversarial Attacks on LLM-Based Code Completion
Slobodan Jenko (ETH Zurich), Martin Vechev (ETH Zurich)
Adversarial AttackAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: A black-box adversarial attack targeting code completion engines based on large language models (LLMs) is proposed, where the attacker injects short comment strings into queries to induce the engine to generate code with security vulnerabilities.
Blink of an eye: a simple theory for feature localization in generative models
Marvin Li (Harvard University), Sitan Chen (Harvard University)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelText
🎯 What it does: A unified theoretical framework is proposed to explain the phenomenon of critical windows in generative models (including diffusion models and autoregressive language models), along with rigorous mathematical proofs and specific examples.
BlockDialect: Block-wise Fine-grained Mixed Format Quantization for Energy-Efficient LLM Inference
Wonsuk Jang (Stanford University), Thierry Tambe (Stanford University)
Computational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Proposes the BlockDialect method for block-level mixed-format quantization of large language models, supporting both weights and activations with 4-bit quantization.
BoA: Attention-aware Post-training Quantization without Backpropagation
Junhan Kim (Samsung Research), Yongkweon Jeon (Samsung Research)
OptimizationTransformerLarge Language ModelText
🎯 What it does: A gradient-free post-training quantization method called BOA is proposed, which uses the attention reconstruction error from the attention module to approximate the Hessian, thereby considering inter-layer dependencies when quantizing weights, significantly improving the quantization performance of large-scale LLMs.
Bongard in Wonderland: Visual Puzzles that Still Make AI Go Mad?
Antonia Wüst (Technical University of Darmstadt), Kristian Kersting (Technical University of Darmstadt)
TransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A systematic evaluation of the performance of current mainstream Vision-Language models on classic Bongard visual reasoning puzzles is conducted, proposing four task settings (open-ended solving, multiple-choice answering, concept recognition, hypothesis generation) to analyze the reasoning behavior of the models in depth.
BOOD: Boundary-based Out-Of-Distribution Data Generation
Qilin Liao (University of Hong Kong), Hengshuang Zhao (University of Hong Kong)
GenerationAnomaly DetectionDiffusion modelImage
🎯 What it does: Utilizing diffusion models to enhance OOD detection by generating high-quality OOD images that fall near the decision boundary by identifying ID feature decision boundaries and applying adversarial perturbations;
Boost-and-Skip: A Simple Guidance-Free Diffusion for Minority Generation
Soobin Um (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisDiffusion modelImageStochastic Differential Equation
🎯 What it does: Boost-and-Skip is proposed, a method for generating minority class samples without external guidance, which facilitates sampling in low-density regions by amplifying variance at the start of sampling and skipping the earliest time steps.
Boosting Adversarial Robustness with CLAT: Criticality Leveraged Adversarial Training
Bhavna Gopal (Duke University), Yiran Chen (University of Arizona)
Adversarial AttackConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This paper proposes the CLAT algorithm, which enhances robustness by fine-tuning only the critical layers of the model during adversarial training.
Boosting Masked ECG-Text Auto-Encoders as Discriminative Learners
Manh Pham Hung, Dong Ma (Singapore Management University)
Convolutional Neural NetworkTransformerAuto EncoderContrastive LearningMultimodalityBiomedical DataElectrocardiogram
🎯 What it does: The D-BETA framework is proposed, which jointly uses aligned ECG-text dual-modal self-supervised learning, combining generative masked autoencoders with contrastive learning.
Boosting Multi-Domain Fine-Tuning of Large Language Models through Evolving Interactions between Samples
Xize Liang (University of Science and Technology of China), Jianye HAO
Domain AdaptationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A dynamic learning curve based on sample mutual influence (EVIC) is proposed to address the performance trade-off in multi-domain fine-tuning of large language models.
Boosting Protein Graph Representations through Static-Dynamic Fusion
Pengkang Guo (École Polytechnique Fédérale de Lausanne), Daniel Probst (Wageningen University)
Drug DiscoveryProtein Structure PredictionGraph Neural NetworkGraphBiomedical Data
🎯 What it does: By combining the static structure of proteins with the dynamic correlations from molecular dynamics trajectories, a heterogeneous graph is constructed and a relational graph neural network is used for property prediction.
Boosting Virtual Agent Learning and Reasoning: A Step-Wise, Multi-Dimensional, and Generalist Reward Model with Benchmark
Bingchen Miao (Ant Group), Juncheng Li (Zhejiang University)
Large Language ModelReinforcement LearningAgentic AIMultimodalityBenchmark
🎯 What it does: A step-by-step, multi-dimensional reward model (Similar) and corresponding SRM benchmark are proposed and implemented. The automated MCTS-P algorithm is used to collect cross-task process-level annotated data across four major platforms, which is then utilized to train the reward model, providing fine-grained feedback during the training and inference phases of the virtual agent (GVA).
Bootstrapping Self-Improvement of Language Model Programs for Zero-Shot Schema Matching
Nabeel Seedat (University of Cambridge), Mihaela van der Schaar (Thomson Reuters)
RetrievalOptimizationTransformerLarge Language ModelPrompt EngineeringBiomedical DataElectronic Health RecordsChain-of-Thought
🎯 What it does: Proposes Matchmaker, a self-improving multi-stage language model program for zero-shot pattern matching tasks;
BOPO: Neural Combinatorial Optimization via Best-anchored and Objective-guided Preference Optimization
Zijun Liao (Sun Yat-sen University), Jiahai Wang (Sun Yat-sen University)
OptimizationGraph Neural NetworkTransformerReinforcement LearningTabular
🎯 What it does: A preference optimization training paradigm based on best anchoring and goal guidance (BOPO) is proposed to enhance the sample efficiency of neural combinatorial optimization models when solving NP-hard problems.
Bounded Rationality for LLMs: Satisficing Alignment at Inference-Time
Mohamad Fares El Hajj Chehade (University of Texas at Austin), Amrit Singh Bedi (University of Central Florida)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a method for aligning large language models with multi-dimensional user preferences during the inference phase through the 'satisficing' principle;
BounDr.E: Predicting Drug-likeness via Biomedical Knowledge Alignment and EM-like One-Class Boundary Optimization
Dongmin Bang (Seoul National University), Sun Kim (Seoul National University)
OptimizationDrug DiscoveryContrastive LearningBiomedical Data
🎯 What it does: A boundary optimization method based on knowledge graphs and structural information, called BOUNDR.E, is proposed for predicting the drug-likeness of compounds.
BoxLM: Unifying Structures and Semantics of Medical Concepts for Diagnosis Prediction in Healthcare
Yanchao Tan (Fuzhou University), Carl Yang (Emory University)
ClassificationRecommendation SystemGraph Neural NetworkTransformerLarge Language ModelTabularBiomedical DataElectronic Health Records
🎯 What it does: Proposes the BoxLM framework to implement diagnostic prediction in electronic health records (EHR), unifying the structure and semantics of medical concepts.
Branches: Efficiently Seeking Optimal Sparse Decision Trees via AO*
Ayman Chaouki (Ecole Polytechnique), Albert Bifet (University of Waikato)
ClassificationOptimizationComputational EfficiencyTabular
🎯 What it does: A search algorithm based on AO* called BRANCHES is proposed for finding optimal sparse decision trees.
Breaking Barriers: Combinatorial Algorithms for Non-Monotone Submodular Maximization with Sublinear Adaptivity and $1/e$ Approximation
Yixin Chen (Texas A&M University), Alan Kuhnle (Texas A&M University)
OptimizationGraph
🎯 What it does: This paper proposes two parallel non-monotonic submodular function maximization methods based on combinatorial algorithms, which can achieve an approximation ratio of 1/e-ε under size constraints, and run with logarithmic adaptive fitness and nearly linear query complexity.
Breaking Silos: Adaptive Model Fusion Unlocks Better Time Series Forecasting
Zhining Liu (University of Illinois Urbana-Champaign), Hanghang Tong (University of Illinois Urbana-Champaign)
OptimizationTime SeriesBenchmark
🎯 What it does: The TIMEFUSE framework is proposed, which achieves sample-level adaptive fusion of different time series prediction models through meta-feature extraction and a learnable fusor.
Breaking the $n^{1.5}$ Additive Error Barrier for Private and Efficient Graph Sparsification via Private Expander Decomposition
Anders Aamand (University of Copenhagen), Yinzhan Xu (University of California San Diego)
OptimizationSafty and PrivacyComputational EfficiencyGraph Neural NetworkGraph
🎯 What it does: A graph cut sparsification algorithm that satisfies differential privacy in polynomial time has been designed, capable of generating a sparse synthetic graph such that all cut values are preserved within a relative error of 1+γ and an absolute error of n^{1/25+o(1)}.
Breaking the Barrier of Hard Samples: A Data-Centric Approach to Synthetic Data for Medical Tasks
MAYNARA DONATO DE SOUZA, Cleber Zanchettin (Universidade Federal de Pernambuco)
Data SynthesisData-Centric LearningDiffusion modelFlow-based ModelGenerative Adversarial NetworkTabularBiomedical Data
🎯 What it does: The Profile2Gen framework is proposed to enhance the quality of synthetic data for medical regression tasks through data analysis and iterative generation.