NeurIPS 2025 Papers — Page 6
Conference on Neural Information Processing Systems · 5275 papers
Beyond Last-Click: An Optimal Mechanism for Ad Attribution
Nan An (Renmin University of China), Liang Zhang (Renmin University of China)
Recommendation SystemOptimizationTabular
🎯 What it does: This study investigates the impact of platform strategic reporting on attribution in multi-platform advertising, constructing a game theory model and designing a fair and incentive-compatible attribution mechanism.
Beyond Least Squares: Uniform Approximation and the Hidden Cost of Misspecification
Davide Maran (Politecnico di Milano), Csaba Szepesvari
OptimizationTabular
🎯 What it does: The study investigates the phenomenon of uniform error amplification in linear regression under random design, proving that it is determined by the Lebesgue constant, and proposes a method to reduce error amplification through auxiliary features and weighted ridge regression (BWR).
Beyond Masked and Unmasked: Discrete Diffusion Models via Partial Masking
Chen-Hao Chao (University of Toronto), Rahul Krishnan
GenerationData SynthesisComputational EfficiencyDiffusion modelImageText
🎯 What it does: MD-Prime is proposed, which transforms the discrete diffusion model to support subword-level masking with partial masking, significantly improving sampling efficiency and generation quality.
Beyond Modality Collapse: Representation Blending for Multimodal Dataset Distillation
Xin Zhang (Agency for Science Technology and Research), Joey Tianyi Zhou (Agency for Science Technology and Research)
CompressionKnowledge DistillationRepresentation LearningContrastive LearningImageTextMultimodalityAudio
🎯 What it does: A new multimodal dataset distillation framework called RepBlend is proposed, which compresses large image-text datasets into compact datasets while maintaining cross-modal learning effectiveness, addressing the modal collapse issue present in existing methods.
Beyond Node-Centric Modeling: Sketching Signed Networks with Simplicial Complexes
Wei Wu (Central South University), Chuan Luo (Beihang University)
Graph Neural NetworkGraph
🎯 What it does: Proposes EdgeSketch+, a method for edge-level signature network embedding based on simplified complexes and local sensitive hashing.
Beyond Oracle: Verifier-Supervision for Instruction Hierarchy in Reasoning and Instruction-Tuned LLMs
Sian-Yao Huang (CyCraft AI Lab), Cheng-Lin Yang (CyCraft AI Lab)
Large Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A unified instruction-level framework has been designed and implemented, utilizing an executable checker to supervise synthesized conflicting instruction instances, thereby achieving multi-level instruction alignment without the need for oracle labels.
Beyond Pairwise Connections: Extracting High-Order Functional Brain Network Structures under Global Constraints
Ling Zhan (Southwest University), Tao Jia (Chongqing Normal University)
Graph Neural NetworkContrastive LearningTime SeriesBiomedical DataMagnetic Resonance ImagingElectrocardiogram
🎯 What it does: A global constraint-based multi-resolution functional brain network learning framework GCM is proposed and implemented, which can directly mine high-order dependencies from multivariate time series and generate discrete functional brain networks.
Beyond Prediction: Managing the Repercussions of Machine Learning Applications
Aline Weber (University of Massachusetts), Bruno Castro da Silva (University of Massachusetts)
ClassificationTabularFinance Related
🎯 What it does: This paper proposes a classification algorithm named THEIA, which can ensure that the new model meets accuracy targets while maintaining a high-confidence upper bound on the actual repercussions produced after deployment, using only observational data from existing deployed models.
Beyond Random: Automatic Inner-loop Optimization in Dataset Distillation
Muquan Li (University of Electronic Science and Technology of China), Ke Qin (University of Electronic Science and Technology of China)
OptimizationKnowledge DistillationImage
🎯 What it does: An Automatic Truncation Backpropagation Through Time (AT-BPTT) method is proposed for dataset distillation, achieving efficient inner-loop optimization through dynamic truncation positions, window sizes, and low-rank Hessian approximations.
Beyond Scalar Rewards: An Axiomatic Framework for Lexicographic MDPs
Mehran Shakerinava (McGill University), Adam Oberman (McGill University)
OptimizationReinforcement Learning
🎯 What it does: This paper introduces a lexicographic reward mechanism in the theory of sequential decision-making, providing a complete theoretical framework for Lexicographic MDP (LMDP) under the assumption of memory independence, and proves that its optimal policy is similar to that of traditional scalar reward MDPs, ensuring the existence of a uniformly optimal static policy.
Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers
Johanna Vielhaben (Fraunhofer HHI), Nils Strodthoff (Carl von Ossietzky Universität Oldenburg)
TransformerImage
🎯 What it does: A method that combines concept discovery with representation alignment analysis is proposed, conducting fine-grained concept alignment evaluation on the intermediate layer features of Vision Transformers.
Beyond Scores: Proximal Diffusion Models
Zhenghan Fang (Johns Hopkins University), Jeremias Sulam (Johns Hopkins University)
GenerationData SynthesisOptimizationDiffusion modelImageTime SeriesStochastic Differential Equation
🎯 What it does: By using backward discretization and learning proximal operators, we propose Proximal Diffusion Models (ProxDM) as an alternative to traditional score-based diffusion sampling methods.
Beyond Single-Task: Robust Multi-Task Length Generalization for LLMs
Yi Hu (Institute for Artificial Intelligence, Peking University), Muhan Zhang (Institute for Artificial Intelligence, Peking University)
Meta LearningTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A multi-task post-training framework called Meta-RFFT is proposed to enhance the length generalization ability of large language models on unseen tasks.
Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning
Shenzhi Wang (Tsinghua University), Junyang Lin (Alibaba Inc.)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study investigates the role of RLVR (Reinforcement Learning with Verifiable Rewards) in LLM inference, focusing on analyzing and utilizing token entropy distribution, particularly high-entropy minority tokens (forking tokens), to enhance inference performance by performing policy gradient updates only on these tokens.
Beyond the Average: Distributional Causal Inference under Imperfect Compliance
Undral Byambadalai (CyberAgent), Shota Yasui (CyberAgent)
TabularBiomedical DataElectronic Health Records
🎯 What it does: This paper studies how to estimate Local Distributed Treatment Effects (LDTE) in experiments with incomplete compliance and covariate-adaptive randomization.
Beyond the Seen: Bounded Distribution Estimation for Open-Vocabulary Learning
Xiaomeng Fan (Beijing Institute of Technology), Yunde Jia (Shenzhen MSU-BIT University)
GenerationData SynthesisDomain AdaptationLarge Language ModelPrompt EngineeringDiffusion modelImage
🎯 What it does: A method is proposed to estimate environmental distribution in open vocabulary learning by generating unseen class data.
Beyond the Surface: Enhancing LLM-as-a-Judge Alignment with Human via Internal Representations
Peng Lai (Southern University of Science and Technology), Guanhua Chen (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The LAGER framework is proposed, which enhances the consistency between LLM-as-a-Judge and human ratings by aggregating the logits from multiple hidden layers and calculating the expected score, without altering the backbone of the LLM.
Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT
Guy Bar-Shalom (Technion), Haggai Maron (Nvidia Research)
TransformerLarge Language ModelText
🎯 What it does: A model based on Vision Transformer, ACT-ViT, was trained to detect hallucinations using the internal activation tensors of LLMs.
Beyond Value Functions: Single-Loop Bilevel Optimization under Flatness Conditions
Liuyuan Jiang (University of Rochester), Tianyi Chen (Cornell University)
OptimizationLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A new penalty-based bi-level optimization algorithm without value function evaluation (PBGD-Free) is proposed, which can be directly used for fine-tuning large-scale language model parameters.
Beyond Verifiable Rewards: Scaling Reinforcement Learning in Language Models to Unverifiable Data
Yunhao Tang (Meta Platforms Inc), Remi Munos
Supervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: A new reinforcement learning algorithm JEPO is proposed, which optimizes chain thinking under Jensen's evidence lower bound to achieve unified training on verifiable and unverifiable data.
BeyondMix: Leveraging Structural Priors and Long-Range Dependencies for Domain-Invariant LiDAR Segmentation
Yujia Chen (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
SegmentationDomain AdaptationPoint Cloud
🎯 What it does: A BeyondMix framework based on the Mamba sequence model is proposed for cross-domain LiDAR point cloud semantic segmentation, utilizing structural priors and long-range dependencies to achieve domain-invariant representation learning.
Bézier Splatting for Fast and Differentiable Vector Graphics Rendering
Xi Liu (Clemson University), Siyu Huang (Clemson University)
OptimizationComputational EfficiencyGaussian SplattingImage
🎯 What it does: This paper proposes Bézier Splatting, a differentiable vector graphics rendering method based on 2D Gaussian splatting, achieving fast and high-quality vectorization and rendering.
Bi-Directional Communication-Efficient Stochastic FL via Remote Source Generation
Maximilian Egger (Technical University of Munich), Deniz Gunduz
CompressionOptimizationFederated LearningConvolutional Neural NetworkImageStochastic Differential Equation
🎯 What it does: A bidirectional communication compression framework (BICOMPFL) based on remote source generation and minimum random coding (MRC) is proposed, which can simultaneously compress both uplink and downlink communication in Bayesian federated learning (stochastic FL);
Bi-Level Decision-Focused Causal Learning for Large-Scale Marketing Optimization: Bridging Observational and Experimental Data
Shuli Zhang (Nanjing University), Guihai Chen (Nanjing University)
OptimizationTabularFinance Related
🎯 What it does: Proposes the Bi-DFCL framework, which combines observational data and experimental data through a bi-level optimization approach for decision-focused causal learning, addressing the prediction-decision mismatch and the bias-variance dilemma.
Bi-Level Knowledge Transfer for Multi-Task Multi-Agent Reinforcement Learning
Junkai Zhang (Chinese Academy of Sciences), Jian Cheng (Chinese Academy of Sciences)
TransformerReinforcement LearningAuto EncoderSequential
🎯 What it does: The BiKT method is proposed in the offline multi-task multi-agent reinforcement learning (MARL) scenario, achieving zero-shot transfer of high-performance policies to unseen tasks from offline data of source tasks.
Bidirectional Motion Transformer for Safety-Critical Traffic Scenario Generation
Yuxin Liu (University of California), Bolei Zhou (University of California)
GenerationData SynthesisAutonomous DrivingTransformerReinforcement LearningPoint Cloud
🎯 What it does: A framework called Adv-BMT based on a bidirectional motion transformer is proposed for generating diverse and realistic safety-critical traffic collision scenarios from real-world data.
Bidirectional Representations Augmented Autoregressive Biological Sequence Generation: Application in De Novo Peptide Sequencing
Xiang Zhang (University of British Columbia), Siqi Sun (Fudan University)
GenerationData SynthesisKnowledge DistillationDrug DiscoveryTransformerBiomedical Data
🎯 What it does: A cross-vector model called CROSSNOVO is proposed, which integrates autoregressive (AR) and non-autoregressive (NAR) decoders to achieve de novo peptide sequence prediction through shared encoders and cross-decoder attention.
Bifrost-1: Bridging Multimodal LLMs and Diffusion Models with Patch-level CLIP Latents
Han Lin (University of North Carolina at Chapel Hill), Mohit Bansal (University of North Carolina at Chapel Hill)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelContrastive LearningImageMultimodality
🎯 What it does: The BIFROST-1 framework is proposed, bridging pre-trained multimodal LLMs and diffusion models through patch-level CLIP latent variables, achieving a unification of high-fidelity image generation and multimodal understanding.
Bigger, Regularized, Categorical: High-Capacity Value Functions are Efficient Multi-Task Learners
Michal Nauman, Pieter Abbeel
Robotic IntelligenceReinforcement Learning
🎯 What it does: The BRC method is proposed to achieve online TD learning scalability to 1B parameters by using a larger Q-value network, cross-entropy loss, and learnable task embeddings in multi-task reinforcement learning.
BiggerGait: Unlocking Gait Recognition with Layer-wise Representations from Large Vision Models
Dingqiang Ye (Southern University of Science and Technology), Xiaoming Liu (Michigan State University)
RecognitionComputational EfficiencyRepresentation LearningTransformerVision Language ModelContrastive LearningImage
🎯 What it does: Explores the role of hierarchical features in large visual models (LVM) for gait recognition and proposes the BiggerGait framework.
Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models
Tyler A. Chang (University of California San Diego), Ben Bergen
TransformerLarge Language ModelText
🎯 What it does: Identify and train the minimal subnetworks (bigram subnetworks) in Transformer language models that can predict the next word using only the current word, and study their structure and role in the residual flow.
Bilevel Network Learning via Hierarchically Structured Sparsity
Jiayi Fan (Shanghai University of Finance and Economics), Mengyun Wu (Shanghai University of Finance and Economics)
OptimizationGraph
🎯 What it does: A dual-layer network learning framework called NNBLNet based on neural networks is proposed to simultaneously recover the dependencies at the variable layer and the group layer.
Bilevel Optimization for Adversarial Learning Problems: Sharpness, Generation, and Beyond
Risheng Liu (Dalian University of Technology), Jin Zhang (Southern University of Science and Technology)
GenerationOptimizationAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a unified bi-level optimization framework that reformulates traditional min-max problems in adversarial learning (such as SAM and GAN) into a solvable lower-level problem, and designs a single-loop stochastic gradient algorithm for implementation.
Bilevel ZOFO: Efficient LLM Fine-Tuning and Meta-Training
Reza Shirkavand (University of Maryland), Heng Huang (University of Maryland)
OptimizationMeta LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A Bilevel-ZOFO dual-layer optimization framework is proposed, combining zero-order full model fine-tuning with first-order PEFT to achieve low memory, efficient LLM fine-tuning and meta-learning.
Binary Quadratic Quantization: Beyond First-Order Quantization for Real-Valued Matrix Compression
Kyo Kuroki (Institute of Science Tokyo), Masato Motomura (Institute of Science Tokyo)
CompressionTransformerImage
🎯 What it does: This paper proposes Binary Quadratic Quantization (BQQ) based on second-order binary matrix products for extremely low-bit matrix compression and post-training quantization of ViT models.
Bio-Inspired Image Restoration
Yuning Cui (Sun Yat-sen University), Alois Knoll (Technical University of Munich)
RestorationTransformerImage
🎯 What it does: This paper proposes an efficient and general image restoration framework called BioIR, which utilizes two bionic modules—Peripheral to Fovea (P2F) and Fovea to Peripheral (F2P)—to achieve bidirectional interaction between local details and global context, thereby enabling high-quality image recovery in various scenarios such as single denoising, universal restoration, and composite degradation.
BioCG: Constrained Generative Modeling for Biochemical Interaction Prediction
Amitay Sicherman (Technion Israel Institute of Technology), Kira Radinsky (Technion Israel Institute of Technology)
GenerationDrug DiscoveryGraph Neural NetworkTransformerBiomedical Data
🎯 What it does: Proposes the BioCG framework, transforming the prediction of biochemical entity interactions into a constrained sequence generation task, generating a discrete code sequence for the target entity.
BioCLIP 2: Emergent Properties from Scaling Hierarchical Contrastive Learning
Jianyang Gu (Ohio State University), Yu Su (Ohio State University)
ClassificationRecognitionTransformerContrastive LearningImageAgriculture Related
🎯 What it does: Trained the BIOCLIP 2 model using hierarchical contrastive learning on the TREEOFLIFE-200M dataset, which consists of 214M images and 952K species, and explored the emerging properties brought by scaling up.
BioOSS: A Bio-Inspired Oscillatory State System with Spatio-Temporal Dynamics
Zhongju Yuan (Ghent University), Dick B.M. Botteldooren
ClassificationOptimizationExplainability and InterpretabilitySpiking Neural NetworkTime SeriesOrdinary Differential Equation
🎯 What it does: A biologically inspired oscillatory state system named BioOSS is proposed, which implements wave propagation using pressure-type (p) and oscillatory-type (o) neurons in a 2D grid to simulate the spatiotemporal dynamics of neural circuits.
BioReason: Incentivizing Multimodal Biological Reasoning within a DNA-LLM Model
Adibvafa Fallahpour (University of Toronto), BO WANG
Drug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMultimodalityBiomedical Data
🎯 What it does: BIOREASON has been developed, a multimodal framework that integrates DNA foundational models with large language models, capable of performing multi-step biological reasoning and variant effect prediction on genomic sequences.
BIPNN: Learning to Solve Binary Integer Programming via Hypergraph Neural Networks
Sen Bai (Changchun University of Science and Technology), Zhengang Jiang (Changchun University of Science and Technology)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper proposes an unsupervised binary integer programming solving framework based on hypergraph neural networks, called BIPNN, which transforms nonlinear BIP into polynomial unconstrained optimization (PUBO) and trains the hypergraph CNN using this as a loss.
Bipolar Self-attention for Spiking Transformers
Shuai Wang (University of Electronic Science and Technology of China), Haizhou Li (Shenzhen Loop Area Institute)
Object TrackingSegmentationSpiking Neural NetworkTransformerReinforcement LearningImage
🎯 What it does: A bipolar self-attention (BSA) module is proposed for the self-attention mechanism in spiking neural networks (SNNs), utilizing ternary spiking neurons to achieve ternary matrix multiplication for query-key (Q-K) interactions, and introducing Shiftmax as an approximate Softmax to achieve low-entropy activation and row normalization constraints, significantly enhancing the performance of SNN Transformers.
Bisecle: Binding and Separation in Continual Learning for Video Language Understanding
Yue Tan (University of New South Wales), Flora D. Salim (University of New South Wales)
TransformerVision Language ModelContrastive LearningVideoText
🎯 What it does: This paper proposes a parameter-efficient continuous learning framework named Bisecle, aimed at enabling large visual-language models to continuously adapt to new tasks in video question answering without losing knowledge of old tasks.
Bit-swapping Oriented Twin-memory Multi-view Clustering in Lifelong Incomplete Scenarios
Shengju Yu (National University of Defense Technology), Xinwang Liu (National University of Defense Technology)
Optimization
🎯 What it does: A lifelong incomplete multi-view clustering method based on dual memory and bit swapping (BSTM) is proposed, which can achieve knowledge transfer in continuous incremental views while maintaining the clustering structure unchanged.
BitMark: Watermarking Bitwise Autoregressive Image Generative Models
Louis Kerner (CISPA Helmholtz Center for Information Security), Adam Dziedzic (CISPA Helmholtz Center for Information Security)
GenerationData SynthesisImage
🎯 What it does: This paper proposes BitMark, a watermark that can be embedded at the bit level and is detectable and radiative. It subtly shifts the generated bit sequence during the autoregressive generation process of images, allowing for the identification of model-generated content while maintaining image quality and generation speed.
Bits Leaked per Query: Information-Theoretic Bounds for Adversarial Attacks on LLMs
Masahiro Kaneko (MBZUAI), Timothy Baldwin (MBZUAI)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper proposes an information-theoretic framework that uses mutual information to measure the amount of information leaked by each query of LLM, and derives the lower bound on the minimum number of queries required for an attack, N_min = log(1/ε)/I(Z,T). Subsequently, experiments are conducted on seven mainstream LLMs and three typical attack types (system prompt leakage, jailbreak, relearning) to validate this theory.
Bivariate Matrix-valued Linear Regression (BMLR): Finite-sample performance under Identifiability and Sparsity Assumptions
Nayel Bettache (Cornell University)
OptimizationImage
🎯 What it does: The paper studies a double matrix-valued linear regression model and provides explicit optimization-unrelated estimators under both noise-free and noisy conditions, along with corresponding non-asymptotic error bounds.
Black-Box Membership Inference Attack for LVLMs via Prior Knowledge-Calibrated Memory Probing
Jinhua Yin (Tsinghua University), Tao Qi (Beijing University of Posts and Telecommunications)
Adversarial AttackTransformerVision Language ModelImageTextMultimodality
🎯 What it does: A black-box membership inference attack framework (KCMP) for large visual-language models is proposed and implemented, which can determine whether a training sample is included by only using the text output generated by the model.
Blackbox Model Provenance via Palimpsestic Membership Inference
Rohith Kuditipudi (Stanford University), Percy Liang (Stanford University)
TransformerLarge Language ModelText
🎯 What it does: A 'black-box source tracing' method based on language model memory is proposed to detect whether another model or text is based on its own training run;
Blameless Users in a Clean Room: Defining Copyright Protection for Generative Models
Aloni Cohen (University of Chicago)
GenerationData SynthesisSafty and PrivacyImageText
🎯 What it does: This paper proposes a provable copyright protection framework based on cleanroom design, defines the concepts of innocent copy protection and tainted models, and demonstrates that differential privacy can achieve (κ,β)-clean copy protection on golden datasets;
Blending Complementary Memory Systems in Hybrid Quadratic-Linear Transformers
Kazuki Irie (Harvard University), Samuel J. Gershman (Harvard University)
RetrievalTransformerLarge Language ModelReinforcement LearningTextSequential
🎯 What it does: Three types of Hybrid Quadratic-Linear Transformers are proposed and evaluated, which achieve a sequence model with long context, precise retrieval, and expressive power by integrating KV attention and DeltaNet's fast weight programming.
BLEUBERI: BLEU is a surprisingly effective reward for instruction following
Yapei Chang (University of Maryland), Mohit Iyyer (University of Maryland)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This study investigates the use of BLEU as a reward signal for aligning instruction following in large language models.
Blindfolded Experts Generalize Better: Insights from Robotic Manipulation and Videogames
Ev Zisselman (Technion Israel Institute of Technology), Aviv Tamar (Technion Israel Institute of Technology)
Robotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningVideo
🎯 What it does: Proposes to induce more exploratory demonstrations by masking information from experts (blindfold), thereby enhancing the generalization performance of behavior cloning in multi-task environments.
Block Coordinate Descent for Neural Networks Provably Finds Global Minima
Shunta Akiyama (CyberAgent)
OptimizationTabular
🎯 What it does: This paper proposes a Block Coordinate Descent (BCD) algorithm for training deep neural networks and proves its global convergence to the global optimal solution under strictly monotonically increasing activation functions (such as LeakyReLU); an improved version of BCD with skip connections and non-negative projections is introduced for ReLU activation, which also achieves global convergence; a generalization error upper bound based on Rademacher complexity is also provided.
Block-Biased Mamba for Long-Range Sequence Processing
Annan Yu (Cornell University), N. Benjamin Erichson (Lawrence Berkeley National Laboratory)
TextSequential
🎯 What it does: Analyze and address the performance shortcomings of Mamba in long sequence modeling, proposing the Block-Biased S6 (B2S6) improvement scheme.
Block-Diagonal LoRA for Eliminating Communication Overhead in Tensor Parallel LoRA Serving
Xinyu Wang (University of Warwick), Matthäus Kleindessner (Amazon Web Services)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: Proposes the Block-Diagonal LoRA (BD-LoRA) method to eliminate the communication overhead of S-LoRA in multi-device inference with multiple LoRA adapters;
BlockDecoder: Boosting ASR Decoders with Context and Merger Modules
Darshan Prabhu (Indian Institute of Technology Bombay), Preethi Jyothi (Indian Institute of Technology Bombay)
RecognitionTransformerAudio
🎯 What it does: Proposes BLOCKDECODER, which splits the decoder into a text encoder and a fusion module to achieve block-level autoregressive output.
BlockScan: Detecting Anomalies in Blockchain Transactions
Jiahao Yu (University of California Santa Barbara), Xinyu Xing (Northwestern University)
Anomaly DetectionTransformerLarge Language ModelTabularTime Series
🎯 What it does: Designed and implemented BlockScan, a Transformer-based framework for detecting anomalies in blockchain transactions, capable of identifying malicious transactions in the DeFi ecosystem.
Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation
Dogyun Park (Korea University), Hyunwoo J. Kim (KAIST)
GenerationData SynthesisComputational EfficiencyTransformerFlow-based ModelImageOrdinary Differential Equation
🎯 What it does: A Blockwise Flow Matching (BFM) framework is designed and implemented, dividing the generation process into several time segments, with each segment modeled using small dedicated velocity blocks, and improving generation quality and inference efficiency through semantic feature guidance and residual feature approximation techniques.
BlurDM: A Blur Diffusion Model for Image Deblurring
Jin-Ting He (National Yang Ming Chiao Tung University), Yen-Yu Lin (National Tsing Hua University)
RestorationDiffusion modelImage
🎯 What it does: A diffusion model that integrates the fuzzy formation process (BlurDM) is proposed for single image deblurring, incorporating it as a prior into the latent space of existing deblurring networks.
BlurGuard: A Simple Approach for Robustifying Image Protection Against AI-Powered Editing
Jinsu Kim (Korea University), Jongheon Jeong (Korea University)
Adversarial AttackDiffusion modelImage
🎯 What it does: A method for generating adversarial noise through adaptive Gaussian blur and spectral regularization is proposed to enhance the robustness of images under text-to-image model editing.
BMW: Bidirectionally Memory bank reWriting for Unsupervised Person Re-Identification
Xiaobin Liu (Nankai University), Jing Yuan (Nankai University)
RecognitionRetrievalConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A bidirectional memory pool rewriting (BMW) mechanism is proposed, which updates using gradient descent while considering intra-class compactness and inter-class separation, improving memory pool updates in unsupervised person re-identification.
BNMusic: Blending Environmental Noises into Personalized Music
Chi Zuo (Wuhan University), Ye Zhu (Princeton University)
GenerationData SynthesisOptimizationLarge Language ModelDiffusion modelAudio
🎯 What it does: The BNMusic framework is proposed, which mixes environmental noise with personalized music generated from user prompts to reduce the perceptibility of noise at low volumes.
Bohdi: Heterogeneous LLM Fusion with Automatic Data Exploration
Junqi Gao (Harbin Institute of Technology), Biqing Qi (Tsinghua University)
Data SynthesisDomain AdaptationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A completely synthetic data-based heterogeneous LLM fusion framework called Bohdi is proposed, which can automatically expand knowledge domains, generate multi-domain data, and inject knowledge through multi-source LLMs.
BoltzNCE: Learning likelihoods for Boltzmann Generation with Stochastic Interpolants and Noise Contrastive Estimation
Rishal Aggarwal (University of Pittsburgh), David Koes
GenerationData SynthesisDrug DiscoveryScore-based ModelFlow-based ModelContrastive LearningSequentialBiomedical Data
🎯 What it does: A model called BoltzNCE is proposed to accelerate the Boltzmann generator by learning the likelihood of an energy-based model to approximate a flow generator, enabling fast sampling.
Boosting Adversarial Transferability with Spatial Adversarial Alignment
Zhaoyu Chen (Fudan University), Wenqiang Zhang (Fudan University)
Adversarial AttackConvolutional Neural NetworkTransformerGenerative Adversarial NetworkImage
🎯 What it does: Proposes Spatial Adversarial Alignment (SAA), which aligns spatial perception and adversarial perception through alignment loss, and fine-tunes the proxy model using a witness model to enhance the transferability of adversarial samples across CNN and ViT.
Boosting Generative Image Modeling via Joint Image-Feature Synthesis
Theodoros Kouzelis (National Technical University of Athens), Nikos Komodakis (University of Crete)
GenerationData SynthesisTransformerDiffusion modelAuto EncoderImage
🎯 What it does: Proposes the ReDi framework, which jointly learns VAE low-level image latent representations and DINOv2 high-level semantic features in the diffusion model, achieving joint generation of images and semantics.
Boosting Knowledge Utilization in Multimodal Large Language Models via Adaptive Logits Fusion and Attention Reallocation
Wenbin An (Xi'an Jiaotong University), Shijian Lu (Nanyang Technological University)
GenerationRetrievalTransformerLarge Language ModelTextMultimodalityRetrieval-Augmented Generation
🎯 What it does: ALFAR is proposed, a plugin method that does not require additional training to enhance the knowledge utilization efficiency of multimodal large language models in retrieval-augmented generation (MRAG).
Boosting Resilience of Large Language Models through Causality-Driven Robust Optimization
Xiaoling Zhou (Peking University), Shikun Zhang (Peking University)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the CDRO framework, utilizing causal knowledge for targeted and enhanced reinforcement learning optimization, reducing the reliance of LLMs on false correlations, and lowering hallucinations and biases;
Boosting Skeleton-based Zero-Shot Action Recognition with Training-Free Test-Time Adaptation
Jingmin Zhu (Monash University), Qiuhong Ke (Monash University)
RecognitionDomain AdaptationGraph Neural NetworkLarge Language ModelVideoRetrieval-Augmented Generation
🎯 What it does: This paper proposes Skeleton-Cache, a training-free test-time adaptation framework that significantly enhances the recognition performance of skeleton-based zero-shot action recognition for unseen actions by constructing a non-parametric cache and dynamically updating it during inference.
Boosting the Uniqueness of Neural Networks Fingerprints with Informative Triggers
Zhuomeng Zhang (Shanghai Jiao Tong University), Shi-Lin Wang (Shanghai Jiao Tong University)
ClassificationOptimizationConvolutional Neural NetworkImage
🎯 What it does: A DNN fingerprint optimization method from an information-theoretic perspective is proposed, which enhances the uniqueness of model copyright tracking by calculating the conditional mutual information of triggers and using a greedy algorithm to select the most informative triggers.
Bootstrap Off-policy with World Model
Guojian Zhan (Tsinghua University), Shengbo Eben Li (Tsinghua University)
Reinforcement LearningWorld ModelSequential
🎯 What it does: A framework named BOOM is proposed, which combines online planning and offline learning through a bootstrap loop between the planner and the policy, eliminating actor-bias;
Bootstrap Your Uncertainty: Adaptive Robust Classification Driven by Optimal-Transport
Jiawei Huang (University of Science and Technology of China), Hu Ding (University of Science and Technology of China)
ClassificationOptimizationImage
🎯 What it does: This paper proposes AdaDRO, an adaptive distributionally robust optimization framework that utilizes optimal transport (OT) and its inverse process to dynamically learn semantic costs and gradually refine the uncertainty set.
Bootstrapping Hierarchical Autoregressive Formal Reasoner with Chain-of-Proxy-Autoformalization
Qi Liu (Shanghai Jiao Tong University), Junchi Yan (Shanghai Jiao Tong University)
Large Language ModelText
🎯 What it does: This paper proposes a hierarchical autoregressive formal reasoner HAR and a chain-based agent automatic formalization CoPA to address the issues of mismatched step granularity and data scarcity in formal problem solving.
Born a Transformer -- Always a Transformer? On the Effect of Pretraining on Architectural Abilities
Mayank Jobanputra (Saarland University), Michael Hahn (Saarland University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: By constructing a series of retrieval and replication tasks, combined with a theoretical framework (C-RASP[pos]) and experiments, this study investigates the performance of large-scale pre-trained Transformers in terms of length generalization, directional bias, and uniqueness bias, and explores how fine-tuning can eliminate directional bias.
Boundary-to-Region Supervision for Offline Safe Reinforcement Learning
Huikang Su (Harbin Institute of Technology), Qinghe Liu
TransformerReinforcement LearningTabularBenchmark
🎯 What it does: A Boundary-to-Region (B2R) framework is proposed to address the shortcomings of reward and cost symmetric conditionalization in offline safe reinforcement learning, achieving region-level supervision through trajectory filtering and CTG realignment.
Boundary-Value PDEs Meet Higher-Order Differential Topology-aware GNNs
Yunfeng Liao (Harbin Institute of Technology), Xiucheng Li (Harbin Institute of Technology)
Graph Neural NetworkGraphPhysics Related
🎯 What it does: A high-order graph neural network (DEC-HOGNN) based on Discrete Exterior Calculus (DEC) and Finite Element Exterior Calculus (FEEC) is proposed for efficiently solving boundary value problems in electromagnetics (such as Poisson, magnetostatics, electrostatic fields, etc.) and can be extended to other PDEs that can be expressed in differential form;
Bounds on the computational complexity of neurons due to dendritic morphology
Anamika Agrawal (Allen Institute), Michael A Buice
🎯 What it does: This study investigates the impact of the dendritic morphology of individual neurons on their computational complexity, using abstract dendritic networks to learn Boolean functions and discovering a phase transition phenomenon.
Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens
Zijian Dong (National University of Singapore), Juan Helen Zhou (National University of Singapore)
TransformerAuto EncoderMultimodalityBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: The first multimodal brain base model, Brain Harmony (BrainHarmonix), has been constructed to unify brain structural morphology and functional dynamics into a 1D marker.
Brain network science modelling of sparse neural networks enables Transformers and LLMs to perform as fully connected
Yingtao Zhang (Tsinghua University), Carlo Vittorio Cannistraci
TransformerLarge Language ModelText
🎯 What it does: A brain-inspired dynamic sparse training framework CHTs/CHTss is proposed, combining BRF initialization and Sigmoid dense decay, allowing Transformers and LLMs to maintain or exceed the performance of fully connected networks at extremely low connectivity rates.
Brain-Informed Fine-Tuning for Improved Multilingual Understanding in Language Models
Anuja Negi (Technical University of Berlin), Fatma Deniz (Technical University of Berlin)
TransformerSupervised Fine-TuningTextBiomedical DataMagnetic Resonance ImagingBenchmark
🎯 What it does: Using bilingual brain imaging data to fine-tune monolingual and multilingual Transformer language models, enhancing their performance in brain encoding and multilingual downstream NLP tasks;
Brain-Inspired fMRI-to-Text Decoding via Incremental and Wrap-Up Language Modeling
Wentao Lu (Nanjing University of Aeronautics and Astronautics), Xuyun Wen (Nanjing University of Aeronautics and Astronautics)
GenerationTransformerLarge Language ModelTextMultimodalityMagnetic Resonance Imaging
🎯 What it does: This paper proposes a cognitive-based segmented incremental fMRI-to-text decoding framework called CogReader.
Brain-Like Processing Pathways Form in Models With Heterogeneous Experts
Jack Cook (University of Oxford), Jascha Achterberg (University of Oxford)
Recurrent Neural NetworkMixture of ExpertsTime Series
🎯 What it does: This paper proposes the Mixture-of-Pathways model based on the Mixture-of-Experts architecture with heterogeneous experts, studying how to achieve self-organization of brain-like processing pathways through priors such as routing cost, task performance scaling, and expert dropout.
Brain-like Variational Inference
Hadi Vafaii (Redwood Center for Theoretical Neuroscience UC Berkeley), Jacob L. Yates (Redwood Center for Theoretical Neuroscience UC Berkeley)
GenerationData SynthesisComputational EfficiencyRecurrent Neural NetworkAuto EncoderImage
🎯 What it does: This paper proposes the FOND framework, deriving brain-like inference dynamics from variational inference and natural gradient, and applies it to construct an iterative variational autoencoder, particularly the iP-VAE recurrent neural network for sparse integer synaptic counts.
Brain-tuning Improves Generalizability and Efficiency of Brain Alignment in Speech Models
Omer Moussa (Max Planck Institute for Software Systems), Mariya Toneva (Max Planck Institute for Software Systems)
TransformerSupervised Fine-TuningBiomedical DataMagnetic Resonance ImagingAudio
🎯 What it does: This study proposes a multi-brain-tuning method that enhances the alignment of a pre-trained speech model with human brain responses by jointly fine-tuning it on fMRI data from multiple subjects.
BrainEC-LLM: Brain Effective Connectivity Estimation by Multiscale Mixing LLM
Wen Xiong (Beijing University of Technology), Jinduo Liu (Beijing University of Technology)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTime SeriesBiomedical DataMagnetic Resonance Imaging
🎯 What it does: This paper proposes a multi-scale hybrid method based on pre-trained LLMs to estimate brain effective connectivity from fMRI sequences.
BrainFlow: A Holistic Pathway of Dynamic Neural System on Manifold
Zhixuan Zhou (University of North Carolina), Guorong Wu (University of North Carolina)
GenerationData SynthesisComputational EfficiencyFlow-based ModelBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A reversible generative model called BrainFlow is proposed to learn the mapping of structural connectivity (SC) and functional connectivity (FC) on the SPD manifold.
BrainMoE: Cognition Joint Embedding via Mixture-of-Expert Towards Robust Brain Foundation Model
Ziquan Wei (University of North Carolina), Guorong Wu (University of North Carolina)
ClassificationRecognitionTransformerMixture of ExpertsMultimodalityBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: We constructed BrainMoE, a brain-based model framework based on a mixture of experts, by pre-training multiple experts under different cognitive states using fMRI and integrating their representations with cognitive adapters to support downstream tasks such as gender and age prediction, human behavior recognition, and early diagnosis of various brain diseases.
BrainODE: Neural Shape Dynamics for Age- and Disease-aware Brain Trajectories
Wonjung Park (KAIST), Jinah Park (KAIST)
SegmentationGenerationData SynthesisOptimizationRecurrent Neural NetworkMeshTime SeriesBiomedical DataMagnetic Resonance ImagingAlzheimer's DiseaseOrdinary Differential Equation
🎯 What it does: A BrainODE framework based on neural ODEs was constructed for continuous modeling of the evolution of brain shape with age and cognitive status, enabling predictions of future brain morphology from a single time point to assist in the early diagnosis of Alzheimer's disease.
BrainOmni: A Brain Foundation Model for Unified EEG and MEG Signals
Qinfan Xiao (Tsinghua University), Chao Zhang (Tsinghua University)
ClassificationRecognitionTransformerContrastive LearningMultimodalityBiomedical Data
🎯 What it does: This paper presents BrainOmni, a unified brain foundation model that can simultaneously handle various EEG and MEG signals.
BraVE: Offline Reinforcement Learning for Discrete Combinatorial Action Spaces
Matthew Landers (University of Virginia), Afsaneh Doryab (University of Virginia)
Reinforcement Learning
🎯 What it does: We propose BraVE, a value estimation method for discrete combinatorial action spaces in offline reinforcement learning, which utilizes a tree structure for efficient search and value evaluation of the action space.
BREAD: Branched Rollouts from Expert Anchors Bridge SFT & RL for Reasoning
Xuechen Zhang (University of Michigan), Samet Oymak (University of Michigan)
TransformerSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: The BREAD algorithm is proposed, integrating supervised fine-tuning (SFT) with reinforcement learning (GRPO) through expert anchors and branch replay, significantly enhancing the reasoning capabilities of small language models (SLM).
Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models
Gen Li (Chinese University of Hong Kong), Changxiao Cai (University of Michigan)
Diffusion model
🎯 What it does: This paper studies the sampling convergence properties of diffusion language models and provides upper and lower bounds on KL error based on information theory.
Breaking Latent Prior Bias in Detectors for Generalizable AIGC Image Detection
Yue Zhou (Shenzhen University), Bin Li (Shenzhen University)
Object DetectionAdversarial AttackGenerative Adversarial NetworkImage
🎯 What it does: A new adversarial training method is proposed to address the insufficient generalization ability of current AIGC detectors when facing unseen generator outputs, particularly by optimizing the initial latent noise to generate adversarial samples on the generator output manifold.
Breaking the Batch Barrier (B3) of Contrastive Learning via Smart Batch Mining
Raghuveer Thirukovalluru (Duke University), Bhuwan Dhingra (Duke University)
OptimizationRepresentation LearningContrastive LearningTextMultimodality
🎯 What it does: A batch mining method based on teacher model ranking and graph community detection (B3) is proposed, which enhances the training efficiency and performance of contrastive learning models by constructing batches that include strong negative samples.
Breaking the Compression Ceiling: Data-Free Pipeline for Ultra-Efficient Delta Compression
Xiaohui Wang (Fudan University), Tao Chen (Fudan University)
CompressionTransformerLarge Language ModelImageTextMultimodality
🎯 What it does: A fully data-free extreme compression framework called UltraDelta is proposed, which achieves ultra-high compression rates for the delta weights of fine-tuned models while maintaining model performance.
Breaking the Discretization Barrier of Continuous Physics Simulation Learning
Fan Xu (University of Science and Technology of China), Xibin Zhao (Tsinghua University)
Time SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: A purely data-driven continuous spatiotemporal dynamics modeling framework called CoPS is proposed to predict the evolution of continuous physical fields from sparse, unstructured observations.
Breaking the Frozen Subspace: Importance Sampling for Low-Rank Optimization in LLM Pretraining
Haochen Zhang (Rice University), Vladimir Braverman (Johns Hopkins University)
OptimizationTransformerLarge Language ModelText
🎯 What it does: This paper studies the problem of low-rank optimization subspace selection in the pre-training of large language models and proposes the SARA method based on importance sampling, breaking the bottleneck of traditional frozen principal subspaces.
Breaking the Gradient Barrier: Unveiling Large Language Models for Strategic Classification
Xinpeng Lv (National University of Defense Technology), Haotian Wang (National University of Defense Technology)
ClassificationOptimizationTransformerLarge Language ModelText
🎯 What it does: This paper proposes a gradient-free strategic classification framework GLIM based on large language models (LLM), utilizing in-context learning (ICL) to simulate strategic manipulation and decision rule optimization in a two-layer optimization without fine-tuning parameters;
Breaking the Order Barrier: Off-Policy Evaluation for Confounded POMDPs
Qi Kuang (Jiangxi University of Finance and Economics), Zhengling Qi (George Washington University)
Reinforcement LearningTabular
🎯 What it does: This paper studies the theory and methods of offline policy evaluation (OPE) in partially observable Markov decision processes (POMDP) with unmeasured confounding factors. It proposes a two-stage model estimator based on observable history for the proxy value function and provides an upper bound on its finite sample error.