NeurIPS 2025 Papers — Page 5
Conference on Neural Information Processing Systems · 5275 papers
Asymptotically Stable Quaternion-valued Hopfield-structured Neural Network with Periodic Projection-based Supervised Learning Rules
Tianwei Wang (University of Edinburgh), Wei Pang (Heriot-Watt University)
OptimizationRobotic IntelligenceReinforcement Learning
🎯 What it does: A quaternion-based supervised Hopfield structured neural network (QSHNN) is proposed, and the weight matrix is ensured to maintain the quaternion multiplication structure through periodic orthogonal projection. The existence uniqueness, asymptotic stability, and trajectory smoothness of the system are theoretically proven, and rapid convergence and high precision control are achieved in simulations of random quaternion targets and robot motion planning.
Asymptotics of SGD in Sequence-Single Index Models and Single-Layer Attention Networks
Luca Arnaboldi (EPFL), Lenka Zdeborova
OptimizationTransformerSequentialOrdinary Differential Equation
🎯 What it does: This study investigates the dynamics of the Sequence Single Index model (SSI) and the stochastic gradient descent (SGD) of single-layer attention networks, deriving a closed-form expression for population loss and analyzing the high-dimensional learning phase.
ATLAS: Autoformalizing Theorems through Lifting, Augmentation, and Synthesis of Data
Xiaoyang Liu (Shanghai Jiao Tong University), Tao Luo (Shanghai Jiao Tong University)
Data SynthesisKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: The ATLAS framework is proposed, which automatically generates large-scale high-quality natural language-formal language pairs of mathematical theorems and implements automated translation on Lean 4;
AtlasGS: Atlanta-world Guided Surface Reconstruction with Implicit Structured Gaussians
Xiyu Zhang, Guofeng Zhang
SegmentationGenerationGaussian SplattingPoint Cloud
🎯 What it does: Introducing AtlasGS: an implicit structured Gaussian light scattering method based on the Atlanta-world constraint, capable of achieving smooth and detail-rich 3D reconstruction in indoor and urban scenes.
Atom of Thoughts for Markov LLM Test-Time Scaling
Fengwei Teng (Hong Kong University of Science and Technology), Zhijiang Guo (Hong Kong University of Science and Technology)
TransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: A framework based on Markov processes called Atom of Thoughts (AOT) is proposed, utilizing memoryless state transitions to minimize historical dependencies in the reasoning process, and achieving gradual simplification of complex problems through a two-phase decomposition-contraction approach.
Atomic Diffusion Models for Small Molecule Structure Elucidation from NMR Spectra
Ziyu Xiong (Princeton University), Ellen D Zhong
GenerationData SynthesisDrug DiscoveryTransformerDiffusion modelTabularMagnetic Resonance Imaging
🎯 What it does: We propose CHEFNMR, an end-to-end framework based on a conditional atomic diffusion model that can directly predict the 3D structure of unknown small molecules solely from 1D NMR spectra and chemical formulas.
Atomic Thinking of LLMs: Decoupling and Exploring Mathematical Reasoning Abilities
Jiayi Kuang (Sun Yat-sen University), Philip S. Yu (University of Illinois Chicago)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: This paper proposes to decompose the mathematical reasoning capabilities of large language models into 'atomic' abilities and constructs corresponding datasets for evaluation and interactive experiments.
Attack by Yourself: Effective and Unnoticeable Multi-Category Graph Backdoor Attacks with Subgraph Triggers Pool
Jiangtong Li (Tongji University), Changjun Jiang (Tongji University)
OptimizationAdversarial AttackGraph Neural NetworkGraph
🎯 What it does: A multi-class graph neural network backdoor attack method based on a subgraph trigger pool, EUMC, is proposed to achieve controllable and low-detectability attacks on node classification models.
Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs
Zhixin Xie (Nanyang Technological University), Jun Luo (Nanyang Technological University)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A two-stage fine-tuning attack is proposed, utilizing 10 completely harmless question-answer pairs to jailbreak LLMs.
Attention (as Discrete-Time Markov) Chains
Yotam Erel (Tel Aviv University), Amit Haim Bermano
SegmentationGenerationTransformerImage
🎯 What it does: Treating the attention matrix as a discrete-time Markov chain, this paper reinterprets the attention mechanism of visual Transformers using methods such as high-order propagation, multi-step iteration, and TokenRank, enhancing its performance in multi-task scenarios.
Attention Mechanism, Max-Affine Partition, and Universal Approximation
Hude Liu (Northwestern University), Han Liu (Northwestern University)
Transformer
🎯 What it does: It is proven that a single-layer single-head self-attention and cross-attention, along with a linear preprocessing layer, can achieve arbitrary precision approximation of any continuous or integrable function on a compact domain under L∞ and Lp norms.
Attention on the Sphere
Boris Bonev (NVIDIA Corporation), Thorsten Kurth (NVIDIA Corporation)
SegmentationDepth EstimationTransformerImage
🎯 What it does: An attention mechanism for spherical data is proposed and implemented, allowing the Transformer to naturally handle functions on a two-dimensional sphere, achieving rotational equivariance;
Attention Sinks: A 'Catch, Tag, Release' Mechanism for Embeddings
Stephen Zhang (University of Toronto), Vardan Papyan (University of Toronto)
TransformerLarge Language ModelText
🎯 What it does: The paper systematically analyzes the phenomenon of 'attention sink points' in large language models and proposes a three-step mechanism of 'capture-label-release', explaining how sink points inject label information into word embeddings and propagate it through subsequent layers.
Attention with Trained Embeddings Provably Selects Important Tokens
Diyuan Wu (Institute of Science and Technology Austria), Marco Mondelli
ClassificationOptimizationTransformerText
🎯 What it does: This paper studies how token embeddings capture the importance of words in text during the gradient descent training process, and clarifies the training dynamics and implicit biases of a single-layer softmax attention model in binary classification tasks.
Attention-based clustering
Rodrigo Maulen-Soto (Sorbonne University), Claire Boyer (Université Paris Saclay)
OptimizationTransformerSequential
🎯 What it does: This paper studies the capabilities of Transformer in unsupervised clustering, proposing a simplified linear attention layer and validating through theoretical risk analysis and PGD/SGD experiments that it can learn cluster centers from mixed Gaussian data; it also demonstrates the remarkable effect of a non-parametric attention layer in achieving approximate quantization in contextual clustering.
Attention! Your Vision Language Model Could Be Maliciously Manipulated
Xiaosen Wang (Huazhong University of Science and Technology), Shudong Zhang (Xidian University)
Adversarial AttackVision Language ModelImageText
🎯 What it does: Proposed and implemented a precise manipulation attack on visual-language models (VMA), which precisely controls the output sequence of VLMs by applying invisible perturbations to images.
AttentionPredictor: Temporal Patterns Matter for KV Cache Compression
Qingyue Yang (University of Science and Technology of China), Bin Li (University of Science and Technology of China)
CompressionOptimizationConvolutional Neural NetworkLarge Language ModelText
🎯 What it does: Proposes AttentionPredictor, which predicts the next attention score through a learned spatiotemporal convolution model to achieve KV cache compression while maintaining LLM performance.
Attractive Metadata Attack: Inducing LLM Agents to Invoke Malicious Tools
Kanghua Mo (Guangzhou University), Zhihao li
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringTextFinance Related
🎯 What it does: This paper proposes and implements an attack method (AMA) that induces agents to call malicious tools by manipulating the metadata (name, description, parameter schema) of LLM agent tools.
Attribution-Driven Adaptive Token Pruning for Transformers
Yaoyao Yan (Shandong Normal University), Weizhi Xu (Shandong Normal University)
Computational EfficiencyKnowledge DistillationTransformerText
🎯 What it does: This paper proposes an attribution-based adaptive token pruning method called AD-TP, which uses Integrated Gradients to evaluate token importance and dynamically determines the retention ratio based on information content, significantly reducing the FLOPs of Transformers.
Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models
Sreyan Ghosh (NVIDIA), Bryan Catanzaro (NVIDIA)
TransformerLarge Language ModelMultimodalityChain-of-ThoughtAudio
🎯 What it does: Developed Audio Flamingo 3, a fully open-source large audio-language model that supports audio reasoning for up to 10 minutes, multi-turn multi-audio chatting, and voice-to-voice interaction, with on-demand chain reasoning capabilities.
Audio Super-Resolution with Latent Bridge Models
Chang Li (Tsinghua University), Jun Zhu (Tsinghua University)
GenerationSuper ResolutionAuto EncoderStochastic Differential EquationAudio
🎯 What it does: A novel audio super-resolution system called AudioLBM based on the Latent Bridge Model (LBM) is proposed, utilizing the continuous latent space of audio waveforms for latent-to-latent generation from low-resolution (LR) to high-resolution (HR);
Audio-Sync Video Generation with Multi-Stream Temporal Control
Shuchen Weng (Beijing Academy of Artificial Intelligence), Xinlong Wang (Beijing Academy of Artificial Intelligence)
GenerationData SynthesisDiffusion modelVideoAudio
🎯 What it does: This paper proposes the MTV framework, which enables controllable video generation based on audio, achieving precise synchronization and multi-dimensional control through audio splitting.
Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models
Haolang Lu (Beijing University of Posts and Telecommunications), Kun Wang (Nanyang Technological University)
Large Language ModelTextChain-of-Thought
🎯 What it does: This paper audits the chain-of-thought (CoT) process of reasoning large language models (RLLM) by constructing a controlled knowledge domain, systematically analyzing the generation, propagation, and correction mechanisms of hallucinations.
Audits Under Resource, Data, and Access Constraints: Scaling Laws For Less Discriminatory Alternatives
Sarah H. Cen (Stanford University), Daniel E. Ho (Stanford University)
Tabular
🎯 What it does: This paper proposes a framework to verify the existence of 'Less Discriminatory Alternatives' (LDA) under resource-constrained conditions, with the core idea being to derive a scalable scaling law from the closed upper bound of the error-fairness Pareto frontier (PF).
AudSemThinker: Enhancing Audio-Language Models Through Reasoning over Semantics of Sound
Gijs Wijngaard (Maastricht University), Michel Dumontier (Maastricht University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVideoTextMultimodalityAudio
🎯 What it does: A structured reasoning phase audio-language model AUDSEMTHINKER has been constructed, and a new semantic description dataset AUDSEM has been released.
AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition
Parsa Rahimi, Sébastien Marcel
RecognitionData SynthesisAdversarial AttackHyperparameter SearchDiffusion modelImage
🎯 What it does: In the task of facial recognition with limited data, AugGen is proposed to generate synthetic samples using a condition diffusion model trained solely on the target dataset, which are then mixed with the original data to enhance recognition performance.
AuroRA: Breaking Low-Rank Bottleneck of LoRA with Nonlinear Mapping
Haonan Dong (Peking University), Liang Wang (Alibaba Group)
Supervised Fine-TuningImageText
🎯 What it does: AuroRA is proposed, which incorporates adaptive nonlinear layers into the LoRA structure to enhance parameter efficiency and break through the low-rank bottleneck;
Auto-Compressing Networks
Vaggelis Dorovatas (National Technical University of Athens), Alexandros Potamianos (National Technical University of Athens)
CompressionOptimizationTransformerImageText
🎯 What it does: Proposes Auto-Compressing Networks (ACN), a network structure that automatically compresses information during training by adding long skip connections directly to the output at each layer.
Auto-Connect: Connectivity-Preserving RigFormer with Direct Preference Optimization
Guojingfeng, Chunchao Guo (Tencent Hunyuan)
GenerationOptimizationTransformerMesh
🎯 What it does: Auto-Connect is an automated rigging method that generates animatable skeletons from 3D meshes while maintaining skeletal connectivity, and further generates high-quality skin weights using geographical distance information.
Auto-Search and Refinement: An Automated Framework for Gender Bias Mitigation in Large Language Models
Yue Xu (ShanghaiTech University), Wenjie Wang (ShanghaiTech University)
GenerationOptimizationTransformerLarge Language ModelText
🎯 What it does: An automated, model-agnostic framework named FaIRMaker is proposed to generate 'Fairwords' and transform them into interpretable instructions through a sequence-to-sequence model, thereby mitigating gender bias in large language models while maintaining the original model's task performance.
AutoData: A Multi-Agent System for Open Web Data Collection
Tianyi Ma (University of Notre Dame), Yanfang Ye (University of Notre Dame)
TransformerLarge Language ModelAgentic AITextBenchmarkFinance Related
🎯 What it does: AutoData has been developed—a multi-agent system that automatically completes open web data collection using natural language instructions. The core includes two main queues for research and development, as well as a dedicated task manager, and achieves efficient collaboration through directed hypergraph caching (OHCache). Additionally, an open data collection benchmark Instruct2DS covering academic, financial, and sports fields has been established.
AutoDiscovery: Open-ended Scientific Discovery via Bayesian Surprise
Dhruv Agarwal (University of Massachusetts Amherst), Peter Clark (Allen Institute for AI)
Large Language ModelReinforcement LearningTabular
🎯 What it does: This paper presents AUTODISCOVERY, a target-free open-ended automated scientific discovery method based on Bayesian surprise.
AutoEdit: Automatic Hyperparameter Tuning for Image Editing
Chau Pham (University at Buffalo), David Doermann (University at Buffalo)
GenerationOptimizationHyperparameter SearchReinforcement LearningDiffusion modelImageBenchmark
🎯 What it does: This paper proposes AutoEdit, a reinforcement learning-based framework that automatically searches for and dynamically adjusts the best hyperparameters during the image editing process of diffusion models, avoiding the high costs of manual tuning and brute-force searching.
Autoencoding Random Forests
Binh Duc Vu (King's College London), David Watson
CompressionRepresentation LearningAuto EncoderTabularBiomedical Data
🎯 What it does: This paper presents a complete method for implementing autoencoders using Random Forests (RF), including learning low-dimensional embeddings from RF and various decoding strategies;
AutoJudge: Judge Decoding Without Manual Annotation
Roman Garipov (Higher School of Economics University), Max Ryabinin (Together AI)
OptimizationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelText
🎯 What it does: This paper proposes an automated task-specific accelerated inference method called AutoJudge, which uses a semi-greedy search to automatically identify which mismatched token pairs significantly impact the final answer during the inference process. This allows for the rapid generation of non-critical tokens in Speculative Decoding, reducing the number of calls to large models.
Automated Composition of Agents: A Knapsack Approach for Agentic Component Selection
Michelle Yuan (Oracle AI), Yi Zhang (Amazon)
OptimizationAgentic AIPrompt EngineeringText
🎯 What it does: This paper proposes an automated combination framework that treats the selection of agent components as an online knapsack problem, enabling the system to dynamically evaluate and select the optimal tools or sub-agents under budget constraints.
Automated Detection of Visual Attribute Reliance with a Self-Reflective Agent
Christy Li (Massachusetts Institute of Technology), Tamar Rott Shaham (Massachusetts Institute of Technology)
RecognitionObject DetectionGenerationTransformerLarge Language ModelAgentic AIImageMultimodality
🎯 What it does: Designed and implemented a self-reflective intelligent agent SAIA for the automatic detection of the dependency relationships of visual attributes in pre-trained visual models;
Automated Model Discovery via Multi-modal & Multi-step Pipeline
Lee Jung-Mok (POSTECH), Tae-Hyun Oh (KAIST)
Large Language ModelVision Language ModelTime Series
🎯 What it does: A multi-modal, multi-step automatic model discovery pipeline is proposed, utilizing AnalyzerVLM and EvaluatorVLM for model suggestion and evaluation;
Automatic Auxiliary Task Selection and Adaptive Weighting Boost Molecular Property Prediction
Zhiqiang Zhong (University of Luxembourg), Davide Mottin (Aarhus University)
Drug DiscoveryGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringGraphBenchmark
🎯 What it does: Proposes the AUTAUT framework, which uses LLM to automatically retrieve and filter auxiliary tasks, while dynamically integrating them into the main molecular property prediction model through a gradient-aligned adaptive weighting mechanism;
Automatic Synthetic Data and Fine-grained Adaptive Feature Alignment for Composed Person Retrieval
Delong Liu (Beijing University of Posts and Telecommunications), Yuan Dong (Beijing University of Posts and Telecommunications)
Data SynthesisRetrievalTransformerLarge Language ModelDiffusion modelImageTextMultimodality
🎯 What it does: A new task for person retrieval combining visual and textual information is proposed—Composed Person Retrieval, along with the construction of a large-scale synthetic dataset SynCPR and a manually annotated test set ITCPR.
Automatic Visual Instrumental Variable Learning for Confounding-Resistant Domain Generalization
Fuyuan CAO, Jiye Liang (Hefei University of Technology)
ClassificationDomain AdaptationImage
🎯 What it does: A domain generalization method for automatically learning visual instrumental variables, VIV-DG, is proposed to counteract the effects of observed and unobserved confounding;
Automaton Constrained Q-Learning
Anastasios Manganaris (Purdue University), Ahmed H Qureshi
Robotic IntelligenceReinforcement LearningSequential
🎯 What it does: This study proposes a new reinforcement learning algorithm ACQL, aimed at solving complex tasks for robots under temporal logic (LTL) constraints, particularly considering multi-stage sub-goals and dynamic safety limits.
AutoPartGen: Autoregressive 3D Part Generation and Discovery
Minghao Chen, Andrea Vedaldi
GenerationData SynthesisTransformerDiffusion modelPoint CloudMesh
🎯 What it does: Proposes AutoPartGen, a self-regressive 3D component generation model that can generate objects, scenes, or cities composed of combinable 3D components from a single image, 2D mask, or existing 3D models;
AutoRedTeamer: Autonomous Red Teaming with Lifelong Attack Integration
Andy Zhou, Bo Li
OptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A fully automated, lifelong learning multi-agent red team framework called AutoRedTeamer is proposed and implemented, capable of generating test cases from high-level risk descriptions, continuously discovering and integrating new attack vectors, and executing attack assessments on black-box LLMs.
Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation
Shanchuan Lin (ByteDance), Lu Jiang (ByteDance)
GenerationData SynthesisKnowledge DistillationTransformerDiffusion modelGenerative Adversarial NetworkVideo
🎯 What it does: Perform autoregressive adversarial fine-tuning on a pre-trained latent video diffusion model to obtain an efficient model that can generate videos with only one forward pass per frame in real-time interactive scenarios.
Autoregressive Motion Generation with Gaussian Mixture-Guided Latent Sampling
Linnan Tu (Huazhong University of Science and Technology), Shijuan Huang (Huazhong University of Science and Technology)
GenerationData SynthesisTransformerAuto EncoderTextMultimodality
🎯 What it does: A text-driven action generation framework based on Gaussian Mixture Model (GMM), named GMMotion, is proposed, which completes action encoding and autoregressive generation through a two-stage process using VAE + Causal Transformer.
AutoSciDACT: Automated Scientific Discovery through Contrastive Embedding and Hypothesis Testing
Samuel Bright-Thonney (Massachusetts Institute of Technology), Philip Harris (Massachusetts Institute of Technology)
Anomaly DetectionData-Centric LearningTransformerContrastive LearningImageTime SeriesPhysics Related
🎯 What it does: AutoSciDACT provides a complete set of automated scientific discovery pipelines, first compressing high-dimensional experimental data into low-dimensional embeddings using contrastive learning, and then utilizing NPLM two-sample tests to discover and statistically detect distribution shifts and novel structures in the embedding space.
AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling
Zhining Zhang (Peking University), Tianmin Shu (Johns Hopkins University)
TransformerLarge Language ModelAgentic AITextMultimodalityBenchmark
🎯 What it does: This paper presents AutoToM, a scalable modeling framework for mental reasoning that can automatically construct and tune agent models for Bayesian inverse planning, capable of inferring any psychological variable in any context.
AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning
Zewei Zhou (University of California Los Angeles), Jiaqi Ma (University of California Los Angeles)
Autonomous DrivingTransformerReinforcement LearningVision Language ModelVision-Language-Action ModelMultimodalityChain-of-Thought
🎯 What it does: We propose AutoVLA, an end-to-end visual-language-action model that integrates semantic reasoning and trajectory planning, directly generating executable action tokens and supporting both fast and slow thinking modes.
Aux-Think: Exploring Reasoning Strategies for Data-Efficient Vision-Language Navigation
Shuo Wang (Beijing Advanced Innovation Center for Future Blockchain), Zhaoxin Fan (Beijing Advanced Innovation Center for Future Blockchain)
TransformerVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: In the Visual Language Navigation (VLN) task, an Aux-Think framework is proposed, which introduces Chain of Thought (CoT) as an auxiliary task during the training phase to learn a more structured reasoning pattern, and directly predicts actions during the testing phase, thereby improving navigation success rates.
Availability-aware Sensor Fusion via Unified Canonical Space
Dong-Hee Paek (KAIST), Seung-Hyun Kong (KAIST)
Object DetectionAutonomous DrivingTransformerPoint Cloud
🎯 What it does: A usability-aware fusion method (ASF) based on unified norm projection and cross-sensor patch attention is proposed, achieving collaborative perception among cameras, LiDAR, and 4D Radar, while maintaining robustness in the event of sensor failure or degradation.
AVCD: Mitigating Hallucinations in Audio-Visual Large Language Models through Contrastive Decoding
Chaeyoung Jung (Korea Advanced Institute of Science and Technology), Joon Son Chung (Korea Advanced Institute of Science and Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelContrastive LearningVideoTextMultimodalityAudio
🎯 What it does: A training-free and human intervention-free inference decoding framework called AVCD is proposed to alleviate the hallucination problem in audio-video large language models.
Avoiding exp(R) scaling in RLHF through Preference-based Exploration
Mingyu Chen (Boston University), Xuezhou Zhang (Boston University)
Reinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: An online RLHF algorithm SE-POPO is proposed, utilizing preference-based exploration and a self-updating sampler, achieving for the first time a polynomial dependence of sample complexity on the reward scale Rmax, thus solving the exponential growth problem present in previous algorithms.
Axial Neural Networks for Dimension-Free Foundation Models
Hyunsu Kim (Korea Advanced Institute of Science and Technology), Juho Lee (Korea Advanced Institute of Science and Technology)
Graph Neural NetworkTransformerTime SeriesBenchmarkPhysics RelatedOrdinary Differential Equation
🎯 What it does: A dimension-independent Axial Neural Network (XNN) is proposed in the context of multidimensional PDE solutions, enabling the model to be uniformly trained and inferred across different spatial dimensions such as 1D, 2D, and 3D.
Backdoor Cleaning without External Guidance in MLLM Fine-tuning
Xuankun Rong (Wuhan University), Mang Ye (Nanyang Technological University)
TransformerLarge Language ModelSupervised Fine-TuningVision Language ModelMultimodality
🎯 What it does: In the fine-tuning process of multimodal large language models, the BYE framework is proposed, which utilizes the self-supervised signal of attention entropy to perform unsupervised filtering on the training set, automatically detecting and removing samples carrying backdoor triggers, thereby enhancing the model's security in attack scenarios.
Backdoor Mitigation via Invertible Pruning Masks
Kealan Dunnett (Queensland University of Technology), Raja Jurdak (Queensland University of Technology)
OptimizationAdversarial AttackTransformerImage
🎯 What it does: A model pruning method based on reversible masks and component selection (IMS) is proposed to eliminate backdoor attacks in deep learning models with only a small number of clean samples.
Backpropagation-Free Test-Time Adaptation via Probabilistic Gaussian Alignment
Youjia Zhang (Sungkyunkwan University), Sungeun Hong (Sungkyunkwan University)
Domain AdaptationGaussian SplattingImage
🎯 What it does: This paper proposes ADAPT, a gradient-free and iterative-free adaptive method for testing time, modeling the features of CLIP as a Gaussian distribution with shared covariance, and continuously updating class means through a high-confidence knowledge base.
Backward Conformal Prediction
Etienne Gauthier (INRIA-ENS-PSL Paris), Michael I. Jordan (UC Berkeley)
ClassificationConvolutional Neural NetworkImage
🎯 What it does: A Backward Conformal Prediction framework is proposed, allowing for the pre-setting of constraints on the size of the prediction set and adaptively determining the error coverage rate, thereby controlling the size of the prediction set while ensuring conformity.
BADiff: Bandwidth Adaptive Diffusion Model
Xi Zhang (Nanyang Technological University), Weisi Lin (Nanyang Technological University)
GenerationCompressionDiffusion modelImage
🎯 What it does: This paper presents BADiff, a diffusion model that can adaptively generate image quality based on real-time network bandwidth.
BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization
Xueyang Zhou (Huazhong University of Science and Technology), Lichao Sun (Lehigh University)
OptimizationAdversarial AttackVision-Language-Action ModelContrastive LearningMultimodality
🎯 What it does: This paper studies backdoor attacks on the Vision-Language-Action (VLA) model and proposes a target decoupling two-stage optimization framework named BadVLA.
Balanced Active Inference
Boyu Chen (Xiamen University), Zhonglei Wang (Xiamen University)
Tabular
🎯 What it does: A balanced active inference framework is proposed, combining cube balanced sampling with active inference to achieve higher statistical efficiency by balancing model uncertainty under a limited labeling budget.
Balanced Conic Rectified Flow
Shin seong Kim (Yonsei University), Youngjung Uh (Yonsei University)
RestorationGenerationRectified FlowImageOrdinary Differential Equation
🎯 What it does: A Balanced Conic Reflow method is proposed, utilizing inverse mapping of real images and spherical linear interpolation (Slerp) to supervise the reflow steps of Rectified Flow, thereby reducing reliance on a large number of pseudo samples and suppressing distribution drift.
Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization
kaiyuan Li, Xinlei Chen (Tsinghua University)
CompressionOptimizationComputational EfficiencyTransformerVision Language ModelMultimodality
🎯 What it does: A balanced visual token pruning method BTP is proposed, which can significantly compress the number of visual tokens while maintaining the performance of the visual language model.
Balancing Gradient and Hessian Queries in Non-Convex Optimization
Deeksha Adil (Institute for Theoretical Studies ETH Zurich), Chenyi Zhang (Stanford University)
Optimization
🎯 What it does: The research balances the number of gradient and Hessian queries in non-convex optimization, proposing a new algorithm and providing a trade-off for the complexity of gradient and Hessian queries.
Balancing Multimodal Training Through Game-Theoretic Regularization
Konstantinos Kontras (KU Leuven), Maarten De Vos (KU Leuven)
OptimizationContrastive LearningMultimodality
🎯 What it does: A multi-modal competitive regularization method MCR is proposed, which utilizes information theory to decompose unique and shared information and dynamically balances multi-modal contributions through game theory.
Balancing Performance and Costs in Best Arm Identification
Michael O Harding, Kirthevasan Kandasamy (University of Wisconsin Madison)
OptimizationDrug DiscoveryTabularBiomedical Data
🎯 What it does: A novel optimal arm identification framework is proposed, which combines the sampling cost with the risk function of performance loss, and designs a dynamic budget elimination algorithm.
Balancing Positive and Negative Classification Error Rates in Positive-Unlabeled Learning
Ximing Li (Jilin University), Renchu Guan (Jilin University)
ClassificationOptimizationImageBiomedical DataAlzheimer's Disease
🎯 What it does: This paper proposes a new PU learning risk estimator DC-PU, which balances the error rates of positive and negative classes with dual constraints and improves training stability.
BAM-ICL: Causal Hijacking In-Context Learning with Budgeted Adversarial Manipulation
Rui Chu (Tufts University), Yingjie Lao (Tufts University)
ClassificationAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: This paper proposes a budget allocation-based two-stage attack framework (BAM-ICL), which hijacks model output by finely perturbing context examples during few-shot learning (ICL) in large language model inference.
Bandit and Delayed Feedback in Online Structured Prediction
Yuki Shibukawa (University of Tokyo), Kenji Yamanishi (University of Tokyo)
Reinforcement Learning
🎯 What it does: An online structured prediction algorithm is proposed that can handle weak feedback, including delayed feedback and bandwidth feedback.
Bandit Guided Submodular Curriculum for Adaptive Subset Selection
Prateek Chanda (Indian Institute of Technology Bombay), Ganesh Ramakrishnan (Indian Institute of Technology Bombay)
OptimizationReinforcement LearningImageText
🎯 What it does: An online submodular subset selection framework called ONLINESUBMOD is proposed, enabling adaptive curriculum learning based on multi-armed bandits.
BaRISTA: Brain Scale Informed Spatiotemporal Representation of Human Intracranial Neural Activity
Lucine L Oganesian, Maryam M. Shanechi (University of Southern California)
ClassificationRepresentation LearningTransformerTime SeriesBiomedical Data
🎯 What it does: A self-supervised Transformer model for intracranial potentials, BaRISTA, is proposed, which can freely encode and mask at different spatial scales (channels, brain regions, brain lobes), and is pre-trained using a latent reconstruction task with spatial-level masks.
Bayes optimal learning of attention-indexed models
Fabrizio Boncoraglio (École Polytechnique Fédérale de Lausanne), Lenka Zdeborova
Transformer
🎯 What it does: This paper proposes the Attention-Indexed Model (AIM) and solves its Bayes optimal learning performance in the high-dimensional limit, providing precise generalization error and phase transition of the learning stage.
Bayesian Concept Bottleneck Models with LLM Priors
Jean Feng (University of California San Francisco), Yan Shuo Tan (National University of Singapore)
ClassificationExplainability and InterpretabilityLarge Language ModelImageTextTabularElectronic Health Records
🎯 What it does: A concept bottleneck model (BC-LLM) based on a Bayesian framework has been developed, utilizing large language models (LLM) for iterative search, generation, and correction of interpretable concepts, which are then used as feature inputs for transparent predictive models.
Bayesian Ego-graph inference for Networked Multi-Agent Reinforcement Learning
Wei Duan (Australian Artificial Intelligence Institute University of Technology Sydney), Junyu Xuan (Australian Artificial Intelligence Institute University of Technology Sydney)
Graph Neural NetworkReinforcement LearningGraph
🎯 What it does: A decentralized actor-critic framework named BayesG is proposed, which utilizes Bayesian variational inference to learn local interaction structures, allowing each agent to sample sparse subgraphs on its ego-graph and make decisions.
Bayesian Optimization with Preference Exploration using a Monotonic Neural Network Ensemble
Hanyang Wang (University of Warwick), Matthias Poloczek (Amazon)
OptimizationContrastive Learning
🎯 What it does: This study combines Bayesian optimization with preference exploration to address multi-objective black-box optimization problems, learning the decision maker's utility preferences and quickly locating the optimal solution.
BayeSQP: Bayesian Optimization through Sequential Quadratic Programming
Paul Brunzema (RWTH Aachen University), Sebastian Trimpe (RWTH Aachen University)
OptimizationBenchmark
🎯 What it does: A high-dimensional black-box optimization algorithm named BayeSQP has been developed, which combines the structure of Sequential Quadratic Programming (SQP) with Bayesian Optimization (BO). It can utilize Gaussian Processes (GP) to jointly model the values, gradients, and Hessian matrices of the objective function and its constraints with only zero-order observations, thereby constructing uncertainty-aware quadratic subproblems to obtain search directions. It then selects step sizes through constrained posterior sampling to achieve efficient local search.
BEAST: Efficient Tokenization of B-Splines Encoded Action Sequences for Imitation Learning
Hongyi Zhou (Karlsruhe Institute of Technology), Rudolf Lioutikov (Microsoft Research)
Computational EfficiencyRobotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: This paper proposes a B-spline-based action sequence tokenizer called BEAST, which can compress continuous robotic actions into fixed-length discrete or continuous tokens and can be seamlessly integrated with various pre-trained models (such as Florence-2, ACT, Transformer, etc.), supporting parallel decoding.
BecomingLit: Relightable Gaussian Avatars with Hybrid Neural Shading
Jonathan Schmidt (Technical University of Munich), Matthias Nießner (Technical University of Munich)
GenerationData SynthesisConvolutional Neural NetworkNeural Radiance FieldImageVideo
🎯 What it does: A method is proposed for reconstructing re-illuminable high-resolution facial avatars from low-cost multi-view video captured on a lighting platform, achieving real-time rendering and animation.
Behavior Injection: Preparing Language Models for Reinforcement Learning
Zhepeng Cen (Carnegie Mellon University), Ding Zhao (Salesforce AI Research)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningAgentic AIText
🎯 What it does: A behavior injection (BRIDGE) method is proposed, which enhances data through the incorporation of exploration and exploitation behaviors before supervised fine-tuning, making large language models more suitable for subsequent reinforcement learning fine-tuning.
Belief-Calibrated Multi-Agent Consensus Seeking for Complex NLP Tasks
Wentao Deng (Shandong University), Pengjie Ren (Shandong University)
Large Language ModelAgentic AIText
🎯 What it does: A belief calibration-based multi-agent consensus seeking framework (BCCS) is proposed, achieving stable consensus in multi-agent systems through belief calibration consensus judgment, collaborator allocation, and leader selection.
BeliefMapNav: 3D Voxel-Based Belief Map for Zero-Shot Object Navigation
Zibo Zhou (Shanghai Jiao Tong University), Siheng Chen (Shanghai Jiao Tong University)
OptimizationRobotic IntelligenceTransformerLarge Language ModelVision Language ModelSimultaneous Localization and MappingPoint Cloud
🎯 What it does: This paper presents BeliefMapNav, a belief map system based on 3D voxels for zero-shot object navigation.
Benford’s Curse: Tracing Digit Bias to Numerical Hallucination in LLMs
Jiandong Shao (Nanyang Technological University), Jianfei Yang (Nanyang Technological University)
Large Language ModelTextBenchmark
🎯 What it does: This paper systematically analyzes the numerical bias exhibited by large language models in numerical reasoning tasks by constructing a unified benchmark for digital distribution. It demonstrates that the digit distribution similar to Benford's law in the pre-training corpus is internalized by the model, leading to numerical hallucinations.
Benign Overfitting in Single-Head Attention
Roey Magen (Weizmann Institute of Science), Gal Vardi (Weizmann Institute of Science)
ClassificationOptimizationTransformerImage
🎯 What it does: This study investigates the benign overfitting phenomenon of single-head softmax attention models on data with label noise, proving that with an appropriate signal-to-noise ratio, gradient descent can achieve perfect fitting of the entire training set in just two steps while maintaining good generalization.
Bernstein–von Mises for Adaptively Collected Data
Kevin Du (Harvard University), Lucas Janson (Harvard University)
TabularTime Series
🎯 What it does: This paper studies Bayesian uncertainty quantification under adaptive sampling data (such as multi-armed bandits, contextual multi-armed bandits, and linear quadratic regulators) and proves that the Bernstein–von Mises theorem holds in such adaptive sampling scenarios, thereby demonstrating the equivalence of Bayesian inference with Wald-type frequentist methods in large samples.
Best-of-N Jailbreaking
John Hughes (Anthropic), Mrinank Sharma (Anthropic)
Adversarial AttackTransformerLarge Language ModelVision Language ModelImageTextMultimodalityAudio
🎯 What it does: Proposed and implemented the Best-of-N (BoN) Jailbreaking algorithm, which utilizes randomly augmented text, images, and audio inputs to repeatedly sample under black-box conditions until the model produces harmful outputs.
Better Estimation of the Kullback--Leibler Divergence Between Language Models
Afra Amini (ETH Zurich), Ryan Cotterell (ETH Zurich)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A Rao-Blackwellized KL divergence estimator is proposed and implemented to assess the differences between language models.
Better Language Model Inversion by Compactly Representing Next-Token Distributions
Murtaza Nazir (University of Southern California), Swabha Swayamdipta (University of Southern California)
GenerationCompressionTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a technique for reverse engineering language models using a compressed sequence of next-word probabilities, which can recover hidden prompt words or system messages from a small amount of model output.
Better NTK Conditioning: A Free Lunch from (ReLU) Nonlinear Activation in Wide Neural Networks
Chaoyue Liu (Purdue University), Xiao Liu (Purdue University)
Recurrent Neural NetworkAudio
🎯 What it does: This study investigates the role of nonlinear activation (using ReLU as an example) in wide neural networks, finding that it can better separate similar samples in the model's gradient feature space and reduce the condition number of the neural tangent kernel (NTK); it also proves that increasing network depth further amplifies this effect.
Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging
Ibrahim Ethem Hamamci (University of Zurich), Bjoern Menze (University of Zurich)
GenerationData SynthesisConvolutional Neural NetworkTransformerVision Language ModelDiffusion modelImageTextBiomedical DataComputed Tomography
🎯 What it does: The BTB3D framework is proposed, utilizing a causal convolution encoder-decoder to generate fine 3D medical image labels, achieving new performance breakthroughs in CT report generation and text-to-CT synthesis tasks.
Better Training Data Attribution via Better Inverse Hessian-Vector Products
Andrew Wang (University of Toronto), Roger Baker Grosse
OptimizationData-Centric LearningConvolutional Neural NetworkSupervised Fine-TuningImageTabular
🎯 What it does: An algorithm named ASTRA is proposed, which accelerates the random Neumann series iteration using the EKFAC preprocessor to improve the performance of training data attribution (TDA).
BevSplat: Resolving Height Ambiguity via Feature-Based Gaussian Primitives for Weakly-Supervised Cross-View Localization
Qiwei Wang (ShanghaiTech University), Yujiao Shi (ShanghaiTech University)
Pose EstimationDepth EstimationAutonomous DrivingConvolutional Neural NetworkGaussian SplattingSimultaneous Localization and MappingImage
🎯 What it does: A weakly supervised cross-view localization method called BevSplat is proposed, which generates bird's-eye view (BEV) using feature-based 3D Gaussian primitives to address the issue of high ambiguity and improve pose estimation between ground cameras and satellite images.
Beyond $\tilde{O}(\sqrt{T})$ Constraint Violation for Online Convex Optimization with Adversarial Constraints
Abhishek Sinha (Tata Institute of Fundamental Research), Rahul Vaze (Tata Institute of Fundamental Research)
Optimization
🎯 What it does: This paper studies the opponent-constrained (COCO) model in online convex optimization and proposes a strategy for adjustable cumulative constraint violation (CCV) and returns;
Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning
Jiayu Wang (University of Wisconsin Madison), Frederic Sala (Salesforce AI Research)
TransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: This paper proposes the SPARKLE framework and the SparkleRL-PSS multi-stage RL pipeline, conducting a fine-grained analysis and enhancement of large language models in three aspects: plan following, knowledge integration, and subproblem decomposition in mathematical reasoning.
Beyond Attention or Similarity: Maximizing Conditional Diversity for Token Pruning in MLLMs
Qizhe Zhang (Peking University), Shanghang Zhang (ByteDance)
CompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelImageVideoMultimodality
🎯 What it does: A training-free, model-agnostic visual token pruning method called CDPruner is proposed, which significantly reduces inference costs while maintaining the performance of multimodal large language models.
Beyond Average Value Function in Precision Medicine: Maximum Probability-Driven Reinforcement Learning for Survival Analysis
Jianqi Feng (Shandong University), Wei Zhao (Shandong University)
Reinforcement LearningBiomedical Data
🎯 What it does: This study investigates a reinforcement learning framework aimed at alternating repeated event data, with the goal of maximizing the probability that the interval between two events exceeds a predetermined threshold.
Beyond Benign Overfitting in Nadaraya-Watson Interpolators
Daniel Barzilai (Weizmann Institute of Science), Ohad Shamir (Weizmann Institute of Science)
ClassificationImage
🎯 What it does: This study investigates the Nadaraya-Watson interpolator and explores how different values of the single hyperparameter β lead to varying generalization behaviors: benign overfitting occurs when β=d, tempered overfitting occurs when β>d, and catastrophic overfitting occurs when β<d.
Beyond Components: Singular Vector-Based Interpretability of Transformer Circuits
Areeb Ahmad (Indian Institute of Technology Kanpur), Ashutosh Modi (Indian Institute of Technology Kanpur)
Explainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: Decomposes the attention heads and MLP layers within the Transformer into orthogonal singular vector directions to achieve fine-grained explanations of the model's internal computations; and selectively adjusts these directions using learned diagonal masks.
Beyond Expectations: Quantile-Guided Alignment for Risk-Calibrated Language Models
Xinran Wang (University of Minnesota), Ali Anwar (University of Minnesota)
OptimizationReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes Quantile-Guided Alignment (QA), a framework for alignment that achieves risk calibration by constraining the quantiles of the output distribution based on RLHF.
Beyond Greedy Exits: Improved Early Exit Decisions for Risk Control and Reliability
Divya Jyoti Bajpai (Indian Institute of Technology Bombay), Manjesh Kumar Hanawal (Indian Institute of Technology Bombay)
ClassificationComputational EfficiencyText
🎯 What it does: This study focuses on risk control in Early-Exit deep networks, proposing the UAT framework that uses a multi-armed bandit dynamic adaptive threshold to balance speed and accuracy during inference.
Beyond Higher Rank: Token-wise Input-Output Projections for Efficient Low-Rank Adaptation
Shiwei Li (Huazhong University of Science and Technology), Ruixuan Li (Huazhong University of Science and Technology)
TransformerSupervised Fine-TuningText
🎯 What it does: Proposes Token-wise Projected Low-Rank Adaptation (TopLoRA), a parameter-efficient fine-tuning method that dynamically generates projection matrices for each input token while maintaining low rank.