NeurIPS 2025 Papers — Page 4
Conference on Neural Information Processing Systems · 5275 papers
Aligning Transformers with Continuous Feedback via Energy Rank Alignment
Shriram Chennakesavalu (Stanford University), Grant M. Rotskoff (Stanford University)
GenerationDrug DiscoveryTransformerReinforcement LearningSequentialBiomedical Data
🎯 What it does: The Energy Rank Alignment (ERA) algorithm is proposed, which utilizes an explicit reward function to optimize autoregressive models through gradient descent, achieving attribute-driven generation of molecular and protein sequences.
Aligning What Matters: Masked Latent Adaptation for Text-to-Audio-Video Generation
Jiyang Zheng (University of Sydney), Tongliang Liu (University of Sydney)
GenerationData SynthesisDiffusion modelVideoTextMultimodalityAudio
🎯 What it does: Proposes the SAVA framework, which uses learnable masks for selective alignment of audio and video latent spaces, achieving synchronized audio-video generation under text prompts.
Alignment of Large Language Models with Constrained Learning
Botong Zhang (University of Pennsylvania), Alejandro Ribeiro (University of Pennsylvania)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes and implements an iterative alignment method based on Lagrangian duality (CAID) to achieve the alignment problem of maximizing primary rewards while satisfying secondary constraints in large language models.
AlignVLM: Bridging Vision and Language Latent Spaces for Multimodal Document Understanding
Ahmed Masry (ServiceNow), Sai Rajeswar (Université de Montréal)
TransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: The ALIGNVLM framework is proposed, which maps visual features to a convex combination of LLM text embeddings to achieve cross-modal alignment.
ALINE: Joint Amortization for Bayesian Inference and Active Data Acquisition
Daolang Huang (Aalto University), Luigi Acerbi (University of Helsinki)
Hyperparameter SearchTransformerReinforcement LearningTabular
🎯 What it does: ALINE is proposed, a unified framework for amortized Bayesian inference and active data acquisition that allows for inference while querying data in real-time.
AliO: Output Alignment Matters in Long-Term Time Series Forecasting
Kwangryeol Park (Ulsan National Institute of Science and Technology), Seulki Lee (Ulsan National Institute of Science and Technology)
TransformerTime Series
🎯 What it does: The AliO method is proposed to address the output alignment problem in long sequence forecasting, and the TAM metric is introduced to quantify alignment quality.
All You Need is One: Capsule Prompt Tuning with a Single Vector
Yiyang Liu (University of Missouri Kansas City), Cheng Han (University of Missouri Kansas City)
TransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes the Capsule Prompt-Tuning (CaPT) framework, which efficiently fine-tunes large language models using a single learnable 'capsule' vector that combines instance-aware information and task-aware information.
Alleviating Hallucinations in Large Language Models through Multi-Model Contrastive Decoding and Dynamic Hallucination Detection
Chenyu Zhu (Zhejiang University), Kaifu Zhang (Zhejiang University)
GenerationOptimizationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: A multi-model contrastive decoding (MCD) framework is proposed, which suppresses the hallucination generation of large models by simultaneously using a real model and an easy hallucination model, and detects and corrects potential hallucination words in real-time during the generation process.
Alligat0R: Pre-Training through Covisibility Segmentation for Relative Camera Pose Regression
Thibaut Loiseau (Ecole des Ponts, University Gustave Eiffel), Vincent Lepetit (Ecole des Ponts, University Gustave Eiffel)
SegmentationPose EstimationAutonomous DrivingTransformerImage
🎯 What it does: A pre-training method based on covisible segmentation called Alligat0R is proposed for binocular vision tasks.
ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio–Language Models
Weifei Jin (Beijing University of Posts and Telecommunications), Derui Wang (CSIRO Data61 Responsible AI Research Centre)
Safty and PrivacyAdversarial AttackLarge Language ModelAudio
🎯 What it does: Proposes ALMGuard, which activates safety shortcuts in audio language models to protect the model from jailbreak attacks using universal audio perturbations.
AlphaBeta is not as good as you think: a simple random games model for a better analysis of deterministic game-solving algorithms
Raphael Boige, Bruno Scherrer (Université de Lorraine)
🎯 What it does: A forward random game tree model is proposed, which constructs the tree step by step through hierarchical conditional distributions while enforcing minimax constraints, to evaluate the average complexity of deterministic game solving algorithms.
AlphaDecay: Module-wise Weight Decay for Heavy-Tailed Balancing in LLMs
Di He (Shenzhen Institute of Advanced Technology), Lu Yin (University of Surrey)
TransformerLarge Language ModelText
🎯 What it does: The AlphaDecay method is proposed, which dynamically allocates different weight decay coefficients to different modules in large language models (LLMs), evaluating the spectral characteristics of each module based on the Heavy-Tail Self-Regularization (HT-SR) theory, thereby achieving module-level weight decay scheduling.
AlphaFold Database Debiasing for Robust Inverse Folding
Cheng Tan (Shanghai AI Laboratory), Stan Z. Li (Westlake University)
Protein Structure PredictionGraph Neural NetworkAuto EncoderGraph
🎯 What it does: This paper first discovers that the AlphaFold protein structure database (AFDB) has systematic biases in geometric details, leading to a decline in performance in highly sensitive inverse folding tasks. Subsequently, a debiasing method based on an SE(3) autoencoder, called DeSAE, is proposed. By introducing noise into the skeleton and learning to denoise, the AFDB structures are projected into a space that better aligns with experimental structures, resulting in a debiased AFDB dataset. Finally, it is validated on various inverse folding models (StructGNN, GraphTrans, GVP, PiFold, ProteinMPNN, etc.), finding that training with the debiased AFDB significantly improves sequence recovery rates and performs comparably to models trained with real PDB data.
AlphaZero Neural Scaling and Zipf's Law: a Tale of Board Games and Power Laws
Oren Neumann (Institute for Theoretical Physics Goethe University Frankfurt), Claudius Gros (Institute for Theoretical Physics Goethe University Frankfurt)
Reinforcement LearningSequential
🎯 What it does: This study investigates the performance scaling laws of the AlphaZero reinforcement learning algorithm under different model sizes and correlates them with Zipf's law; it also explores the mechanisms behind the reverse scaling phenomenon in large models.
ALTER: All-in-One Layer Pruning and Temporal Expert Routing for Efficient Diffusion Generation
Xiaomeng Yang (Northeastern University), Shangqian Gao (Florida State University)
GenerationOptimizationComputational EfficiencyKnowledge DistillationTransformerMixture of ExpertsDiffusion modelImage
🎯 What it does: Proposes the ALTER framework, transforming diffusion models into a multi-temporal expert mixture, unifying layer pruning, expert routing, and fine-tuning to achieve single-stage optimization.
Alternating Gradient Flows: A Theory of Feature Learning in Two-layer Neural Networks
Daniel Kunin (University of California Berkeley), Nina Miolane (KTH Royal Institute of Technology)
OptimizationRepresentation LearningTransformerTabularOrdinary Differential Equation
🎯 What it does: This paper proposes the Alternating Gradient Flows (AGF) framework to describe the feature learning dynamics of small-initialization two-layer networks and unifies the stepwise learning behavior of linear networks and nonlinear quantum networks.
AltLoRA: Towards Better Gradient Approximation in Low-Rank Adaptation with Alternating Projections
Xin Yu (Pennsylvania State University), Lingzhou Xue (Pennsylvania State University)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes AltLoRA, a parameter-efficient fine-tuning method that alternately updates the gradients and momentum of LoRA in a low-rank space, maintaining extremely low memory overhead.
ALTo: Adaptive-Length Tokenizer for Autoregressive Mask Generation
Lingfeng Wang (Uni-Ubi), Wuyue Zhao (Uni-Ubi)
SegmentationGenerationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningMultimodality
🎯 What it does: Proposes the ALTo adaptive length masking tokenizer and integrates it into a multimodal large language model, achieving dynamic generation of masked tokens based on target complexity;
AMBER: Adaptive Mesh Generation by Iterative Mesh Resolution Prediction
Niklas Freymuth (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
Graph Neural NetworkMeshPhysics Related
🎯 What it does: The AMBER method is proposed, which utilizes graph neural networks to predict the size field at each step and iteratively generates adaptive meshes through a mesh generator, thereby reducing the workload of manual mesh design and optimization.
Ambient Diffusion Omni: Training Good Models with Bad Data
Giannis Daras (Massachusetts Institute of Technology), Constantinos Costis Daskalakis
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: Proposes the Ambient Diffusion Omni (Ambient-o) framework, which enhances image generation quality and diversity in diffusion model training using low-quality, synthetic, or outlier distribution data.
Ambient Proteins - Training Diffusion Models on Noisy Structures
Giannis Daras (Massachusetts Institute of Technology), Daniel Jesus Diaz
Protein Structure PredictionDiffusion modelMesh
🎯 What it does: This paper proposes the Ambient Protein Diffusion framework, which utilizes low-confidence AlphaFold structures as noisy training samples to train a protein diffusion model that generates new proteins with high diversity and high designability.
Among Us: A Sandbox for Measuring and Detecting Agentic Deception
Satvik Golechha (MATS), Adrià Garriga-Alonso (FAR AI)
TransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: A sandbox environment based on the game 'Among Us' is proposed to induce and evaluate the deceptive behavior of large language models (LLMs) in open-ended, long-term agent deception; at the same time, an unlimited deception Elo metric is introduced to quantify the models' deception and detection capabilities.
AmorLIP: Efficient Language-Image Pretraining via Amortization
Haotian Sun (Georgia Institute of Technology), Bo Dai (Georgia Institute of Technology)
RetrievalComputational EfficiencyRepresentation LearningTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: The AMORLIP framework is proposed, which eliminates the need for a large number of negative samples in CLIP pre-training through a lightweight amortization network, achieving more efficient language-image alignment learning.
Amortized Active Generation of Pareto Sets
Daniel M. Steinberg (CSIRO Data61), Edwin V. Bonilla (CSIRO Data61)
GenerationOptimizationTransformerSequential
🎯 What it does: Proposes the A-GPS framework, an online learning conditional generative model for generating Pareto sets, and supports posterior user preference projection.
Amortized Sampling with Transferable Normalizing Flows
Charlie B. Tan (University of Oxford), Kirill Neklyudov (Université de Montréal)
Drug DiscoveryFlow-based ModelSequentialBiomedical Data
🎯 What it does: A transferable full-atom normalized flow model PROSE has been developed for efficient isothermal distribution sampling.
Amortized Variational Transdimensional Inference
Laurence Davies (University of New South Wales), Scott A Sisson
Flow-based ModelTabularTime Series
🎯 What it does: A CoSMIC regularization flow and VTI framework capable of adaptive variational inference in cross-dimensional (multi-model) space is proposed, which can approximate the posterior distribution of different dimensional models using a single variational density.
Amplifying Prominent Representations in Multimodal Learning via Variational Dirichlet Process
Tsai Hor Chan (University of Pennsylvania), Lequan Yu (University of Hong Kong)
GenerationAnomaly DetectionRepresentation LearningConvolutional Neural NetworkRecurrent Neural NetworkTransformerMultimodalityBiomedical DataElectronic Health Records
🎯 What it does: A multimodal learning framework DPMM based on the Dirichlet process is designed, which utilizes its richer-gets-richer property to amplify significant features in multimodal feature fusion and achieve adaptive alignment of modality distributions, while also being able to perform generative imputation for missing modalities.
An Adaptive Algorithm for Bilevel Optimization on Riemannian Manifolds
Xu Shi (Fudan University), Rujun Jiang (Fudan University)
OptimizationTabular
🎯 What it does: An adaptive Riemannian double-layer optimization algorithm, AdaRHD, is proposed to solve double-layer optimization problems with lower-layer strong convexity constraints on Riemannian manifolds, without the need to know parameters such as gradients and curvatures in advance.
An Adaptive Quantum Circuit of Dempster's Rule of Combination for Uncertain Pattern Classification
Fuyuan Xiao (Chongqing University), Witold Pedrycz (University of Alberta)
ClassificationComputational EfficiencyTabular
🎯 What it does: An adaptive quantum circuit AQC-DRC is proposed for pattern classification in uncertain environments.
An Analysis of Causal Effect Estimation using Outcome Invariant Data Augmentation
UZAIR AKBAR, Bo Dai (Max Planck Institute for Intelligent Systems)
Domain AdaptationContrastive LearningImageTabular
🎯 What it does: This paper proposes a method that treats data augmentation (DA) as a soft intervention to estimate unidentifiable causal effects and enhance the generalization performance of external interventions.
An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations
Seonghwan Park (POSTECH), Namhoon Lee (POSTECH)
ClassificationRecognitionExplainability and InterpretabilitySupervised Fine-TuningImage
🎯 What it does: The system studied the performance of the Concept Bottleneck Model (CBM) in the presence of label noise and proposed a two-stage robustness enhancement scheme.
An Analytical Theory of Spectral Bias in the Learning Dynamics of Diffusion Models
Binxu Wang (Harvard University), Cengiz Pehlevan (Harvard University)
RestorationGenerationData SynthesisConvolutional Neural NetworkDiffusion modelImageOrdinary Differential Equation
🎯 What it does: This paper derives the closed-form dynamics of the generative distribution during the training process of diffusion models by analyzing the gradient flow and probability flow ODEs of linear and convolutional denoisers, revealing the inverse variance spectrum bias law.
An Effective Levelling Paradigm for Unlabeled Scenarios
Fangming Cui (Shanghai Jiao Tong University), Xinmei Tian (University of Science and Technology of China)
ClassificationDomain AdaptationPrompt EngineeringVision Language ModelImage
🎯 What it does: This paper proposes a plugin method called Levelling Paradigm (LePa) to coordinate the CE loss and visual regularization in VLM fine-tuning, aiming to enhance the generalization performance of unlabeled tasks.
An Efficient Local Search Approach for Polarized Community Discovery in Signed Networks
Linus Aronsson (Chalmers University of Technology & University of Gothenburg), Morteza Haghir Chehreghani (Chalmers University of Technology & University of Gothenburg)
OptimizationGraph Neural NetworkGraph
🎯 What it does: A new polarization community detection (PCD) objective function is proposed, incorporating a square size penalty to avoid clustering imbalance, and the first scalable local search algorithm is designed, achieving linear convergence rate based on block coordinate Frank-Wolfe; an efficient implementation (Alg.3) is also provided, enabling rapid iterations on large-scale signed networks.
An Efficient Orlicz-Sobolev Approach for Transporting Unbalanced Measures on a Graph
Tam Le (Institute of Statistical Mathematics), Kenji Fukumizu (Institute of Statistical Mathematics)
OptimizationComputational EfficiencyGraph Neural NetworkTextGraph
🎯 What it does: This paper proposes two novel transport distances—Orlicz-EPT (entropy-perturbed transport utilizing the Orlicz geometric structure) and Orlicz-Sobolev Transport (OST)—to address optimal transport problems for measures with different total masses (unbalanced) on graphs; it also provides an efficient algorithm for solving OST through one-dimensional optimization and proves its relationship with traditional distances such as OW, GST, ST, and UST; theoretical proofs and experimental validations demonstrate its computational efficiency and performance.
An Ellipsoid Algorithm for Online Convex Optimization
Zakaria Mhammedi (Google Research)
Optimization
🎯 What it does: An online convex optimization (OCO) algorithm based on separation operators is proposed, utilizing dynamic ellipsoids for projection-free decision-making, significantly reducing dependence on the sphericality of the feasible set (κ).
An Evidence-Based Post-Hoc Adjustment Framework for Anomaly Detection Under Data Contamination
Sukanya Patra (University of Mons), Souhaib Ben Taieb (Mohamed bin Zayed University of Artificial Intelligence)
Anomaly DetectionContrastive LearningImageTabular
🎯 What it does: This work proposes a post-adjustment framework EPHAD for adaptive correction when testing unsupervised anomaly detection models under training data contamination.
An Improved Algorithm for Adversarial Linear Contextual Bandits via Reduction
Tim van Erven (University of Amsterdam), Chen-Yu Wei (University of Virginia)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: An efficient algorithm is proposed for handling linear contextual bandit problems with adversarial loss and a random action set. This algorithm simplifies the setting to an adversarial linear bandit problem with a fixed action set and achieves a better regret bound without knowledge of the context distribution or a context simulator.
An Information-theoretical Framework for Understanding Out-of-distribution Detection with Pretrained Vision-Language Models
Bo Peng (University of Technology Sydney), Zhen Fang (University of Technology Sydney)
Anomaly DetectionTransformerVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes an information-theoretic framework to explain and improve post-hoc OOD detection based on pre-trained vision-language models (VLMs). It achieves energy modeling by treating CLIP's OOD detection as random sampling of pointwise mutual information (PMI) and further decomposes PMI into conditional and unconditional components to reduce estimation error.
An Investigation of Memorization Risk in Healthcare Foundation Models
Sana Tonekaboni (Massachusetts Institute of Technology), Marzyeh Ghassemi (Massachusetts Institute of Technology)
Safty and PrivacyBiomedical DataElectronic Health Records
🎯 What it does: This study investigates the memory risks in the foundational model of structured electronic health records and proposes a black-box evaluation testing framework.
An Iterative Algorithm for Differentially Private $k$-PCA with Adaptive Noise
Johanna Düngler (University of Copenhagen), Amartya Sanyal (University of Copenhagen)
OptimizationSafty and PrivacyTabular
🎯 What it does: A random k-PCA algorithm satisfying differential privacy, k-DP-PCA and k-DP-Ojas, is proposed, supporting any k ≤ d, with sample complexity approximately O(d), and providing near-optimal error guarantees.
An Optimized Franz-Parisi Criterion and its Equivalence with SQ Lower Bounds
Siyu Chen (Yale University), Peiyuan Zhang (Yale University)
Anomaly DetectionOptimization
🎯 What it does: An optimized Franz-Parisi (GFP) difficulty criterion is proposed, and it is proven that under the lightweight assumption, this criterion is equivalent to the statistical query (SQ) lower bound, thereby providing a unified hardness analysis framework for multi-class detection tasks.
AnaCP: Toward Upper-Bound Continual Learning via Analytic Contrastive Projection
Saleh Momeni (University of Illinois Chicago), Bing Liu (University of Illinois Chicago)
ClassificationRepresentation LearningContrastive LearningImage
🎯 What it does: The AnaCP method is proposed, which utilizes analytical contrastive projection for continuous adaptation of features extracted from pre-trained models, achieving gradient-free training for class-incremental learning.
Analog Foundation Models
Julian Büchel (IBM Research), Abu Sebastian (IBM Research)
Computational EfficiencyKnowledge DistillationTransformerLarge Language ModelText
🎯 What it does: The study adapts large language models (such as Phi-3-mini-4k-instruct and Llama-3.2-1B-Instruct) to simulate memory computing (AIMC) hardware, achieving robustness against noise, quantization, and other non-idealities through hardware-aware training, and validates their performance on various benchmarks.
Analog In-memory Training on General Non-ideal Resistive Elements: The Impact of Response Functions
Zhaoxian Wu (Cornell University), Tianyi Chen (Cornell University)
OptimizationConvolutional Neural NetworkReinforcement LearningImage
🎯 What it does: This study investigates the dynamics of gradient descent training on AIMC hardware under non-ideal response functions and proposes a residual learning framework to eliminate biases caused by response asymmetry.
Analogy-based Multi-Turn Jailbreak against Large Language Models
Mengjie Wu (Wuhan University), Lina Wang (Wuhan University)
Safty and PrivacyAdversarial AttackTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper proposes an analogy-based multi-round jailbreak framework (AMA), which first constructs a completely secure dialogue context and then injects malicious semantics in the final step to induce the LLM to generate harmful content consistent with malicious intent.
Analytic Energy-Guided Policy Optimization for Offline Reinforcement Learning
Jifeng Hu (Jilin University), Dacheng Tao (Nanyang Technological University)
OptimizationReinforcement LearningDiffusion modelTabularBenchmark
🎯 What it does: A strategy optimization framework based on analytical energy guidance (AEPO) is proposed, which uses diffusion models in offline reinforcement learning to achieve more precise action sampling through analytical intermediate energy.
Analyzing Fine-Grained Alignment and Enhancing Vision Understanding in Multimodal Language Models
Jiachen Jiang (Ohio State University), Zhihui Zhu (Ohio State University)
RecognitionGenerationRetrievalCompressionTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: This paper analyzes the compression and alignment functions of the visual projector in multimodal large language models (MLLMs) and proposes a patch-aligned pre-training method. It constructs a dataset with detailed patch semantic labels (PAD) to enhance the granular alignment between visual embeddings and text word embeddings, significantly improving the performance of tasks such as caption generation, reference expression localization, visual question answering, and instruction following.
Analyzing Similarity Metrics for Data Selection for Language Model Pretraining
Dylan Sam (Carnegie Mellon University), Sanjiv Kumar (Google Research)
Large Language ModelText
🎯 What it does: A three-part evaluation framework is proposed and implemented to assess the applicability of text embedding models in the selection of pre-training data for language models, and comparative experiments are conducted with embedding models specifically trained for pre-training tasks.
Analyzing the Power of Chain of Thought through Memorization Capabilities
Lijia Yu (Institute of AI for Industries), Lijun Zhang (Institute of AI for Industries)
TransformerChain-of-Thought
🎯 What it does: This paper theoretically analyzes the impact of Chain of Thought (CoT) on autoregressive Transformers from the perspective of memorization. It provides the necessary and sufficient conditions for finite languages to be memorized and derives the upper and lower bounds of the parameters required for memory. By comparing the memory capabilities of CoT and non-CoT Transformers, it demonstrates that CoT does not extend the functionality of Transformers for all reasoning tasks and further explores the memory challenges of infinite languages.
Anatomically inspired digital twins capture hierarchical object representations in visual cortex
Emanuele Luconi (Bocconi University), Alessandro Sanzeni (Bocconi University)
RecognitionObject DetectionConvolutional Neural NetworkImage
🎯 What it does: This study investigates the digital twin model in capturing the hierarchical representation transformation of the visual cortex in rodents, particularly the robustness of object recognition.
Anchor-based Maximum Discrepancy for Relative Similarity Testing
Zhijian Zhou (University of Melbourne), Feng Liu (University of Melbourne)
ClassificationAnomaly DetectionConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: The anchor-based maximum discrepancy (AMD) method is proposed, which simultaneously learns the relative similarity hypothesis and the optimal kernel, addressing the issues of kernel selection and hypothesis prior in traditional relative similarity testing.
Anchored Diffusion Language Model
Litu Rout (University of Texas at Austin), Sanjay Shakkottai (University of Texas at Austin)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelTextBiomedical DataChain-of-Thought
🎯 What it does: Proposes the Anchored Diffusion Language Model (ADLM), which improves the generation and inference of diffusion language models through a two-stage denoising process (first predicting the distribution of important vocabulary, then using that distribution to guide subsequent denoising).
AngleRoCL: Angle-Robust Concept Learning for Physically View-Invariant Adversarial Patches
Wenjun Ji (Nankai University), Qing Guo (Nankai University)
Object DetectionGenerationAdversarial AttackDiffusion modelImage
🎯 What it does: The AngleRoCL method is proposed, achieving high attack effectiveness of text-to-image generation attack patches under different perspectives.
Angles Don’t Lie: Unlocking Training‑Efficient RL Through the Model’s Own Signals
Qinsi Wang (Duke University), Yiran Chen (Duke University)
Large Language ModelReinforcement LearningTabularSequential
🎯 What it does: Proposed a GAIN-RL framework based on the model's own perspective of signal concentration for dynamic scheduling of reinforcement learning fine-tuning data.
Angular Constraint Embedding via SpherePair Loss for Constrained Clustering
Shaojie Zhang (University of Manchester), Ke Chen (University of Manchester)
Representation LearningAuto EncoderImage
🎯 What it does: The SpherePair framework is proposed, achieving anchor-free constrained clustering by learning deep constrained embeddings in the angular space.
Angular Steering: Behavior Control via Rotation in Activation Space
Hieu M. Vu (Torilab), Tan Minh Nguyen
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelText
🎯 What it does: This paper proposes Angular Steering, a method for fine-grained, continuous adjustment of large language model behavior by rotating activation vectors within a two-dimensional subspace, and presents an adaptive variant.
AnimateQR: Bridging Aesthetics and Functionality in Dynamic QR Code Generation
Guangyang Wu (Shanghai Jiao Tong University), Xiaohong Liu (Shanghai Jiao Tong University)
GenerationDiffusion modelImage
🎯 What it does: The AnimateQR framework is proposed, capable of generating dynamic QR codes that balance aesthetics and readability.
Anomaly Detection by an Ensemble of Random Pairs of Hyperspheres
Walid Durani (LMU Munich), Christian Böhm (University of Vienna)
Anomaly DetectionTabular
🎯 What it does: This paper proposes an isolation-based anomaly detection method for hypersphere sets called ADERH, which is based on random pairing.
Anti-Aliased 2D Gaussian Splatting
Mae Younes (INRIA France), Adnane Boukhayma (INRIA France)
Gaussian SplattingImage
🎯 What it does: A two-dimensional Gaussian splitting method with anti-aliasing (AA-2DGS) is proposed, which can maintain high-quality rendering and preserve geometric accuracy at different sampling rates.
Antidistillation Sampling
Yash Savani (Carnegie Mellon University), J Zico Kolter
Knowledge DistillationTransformerLarge Language ModelText
🎯 What it does: Proposes an anti-distillation sampling method that adds a penalty term based on the gradient of a proxy model to the sampling distribution of the next token in LLM, maintaining high probability of generated text while significantly reducing the distillation effect of the student model.
Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector
Haoyan Yang (New York University), Taha Kass-Hout (GE Healthcare)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes an external module called the Reasoning-based Bias Detector (RBD) to detect biases in LLM evaluations and provide structured reasoning to help evaluators self-correct.
Any-stepsize Gradient Descent for Separable Data under Fenchel–Young Losses
Han Bao (Institute of Statistical Mathematics), Yuki Takezawa (Kyoto University and OIST)
Optimization
🎯 What it does: The study investigates the use of arbitrary step size gradient descent on linearly separable data and proves convergence under the Fenchel-Young loss.
Anytime-valid, Bayes-assisted, Prediction-Powered Inference
Valentin Kilian (University of Oxford), Francois Caron
Mixture of ExpertsTabular
🎯 What it does: When experimental data is scarce but there is a large amount of unlabeled data, machine learning is used to improve the accuracy of confidence intervals, and this framework is extended to online scenarios that are updated over time, constructing a uniformly valid confidence sequence over time.
AOR: Anatomical Ontology-Guided Reasoning for Medical Large Multimodal Model in Chest X-Ray Interpretation
Qingqiu Li (Fudan University), Shujun Wang (Hong Kong Polytechnic University)
RecognitionObject DetectionSegmentationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBiomedical DataChain-of-Thought
🎯 What it does: A multi-step reasoning framework AOR based on anatomical ontology is proposed, supporting region-level prompts for medical multimodal large models for chest X-ray interpretation.
APML: Adaptive Probabilistic Matching Loss for Robust 3D Point Cloud Reconstruction
Sasan Sharifipour (University of Oulu), Miguel Bordallo Lopez
GenerationData SynthesisPoint Cloud
🎯 What it does: This paper proposes a differentiable, near-quadratic complexity Adaptive Probability Matching Loss (APML) to replace the traditional Chamfer distance for supervised 3D point cloud reconstruction and generation tasks.
APOLLO: Automated LLM and Lean Collaboration for Advanced Formal Reasoning
Azim Ospanov (Huawei Hong Kong Research Center), Roozbeh Yousefzadeh (Chinese University of Hong Kong)
Large Language ModelTextBenchmark
🎯 What it does: This paper proposes a fully automated framework called Apollo, which utilizes LLMs, the Lean compiler, and automated solvers to collaboratively repair and generate formal proofs.
Approximate Domain Unlearning for Vision-Language Models
Kodai Kawamura (Tokyo University of Science), Go Irie (Tokyo University of Science)
Object DetectionDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: Proposed and implemented the task of 'Approximate Domain Unlearning' (ADU), which allows pre-trained vision-language models to recognize objects only within a specified domain while failing in other domains;
Approximate Gradient Coding for Distributed Learning with Heterogeneous Stragglers
Heekang Song (Korea Advanced Institute of Science and Technology), Wan Choi (Seoul National University)
Object DetectionOptimizationFederated LearningConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A gradient encoding scheme for heterogeneous slow nodes is proposed, which can reduce residual error and maintain unbiased gradient estimation in distributed learning.
Approximately Aligned Decoding
Daniel Melcer (Northeastern University), Anoop Deoras (Amazon)
GenerationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A new generation method called Approximately Aligned Decoding (AprAD) is proposed to avoid errors or constraint violations during the generation process of large language models while reducing computational overhead.
Approximating Shapley Explanations in Reinforcement Learning
Daniel Beechey (University of Bath), Özgür Şimşek (University of Bath)
Explainability and InterpretabilityReinforcement Learning
🎯 What it does: This paper proposes FastSVERL, a scalable parametric method for approximating Shapley values in reinforcement learning to explain the behavior, outcomes, and predictions of agents.
Approximation and Generalization Abilities of Score-based Neural Network Generative Models for Sub-Gaussian Distributions
Guoji Fu (National University of Singapore), Wee Sun Lee (National University of Singapore)
GenerationData SynthesisScore-based ModelTabular
🎯 What it does: This paper studies the approximation and generalization capabilities of score-based generative models (SGMs) in estimating an unknown distribution P0, particularly under the assumption that P0 is an α-th order Gaussian distribution. It proves that deep ReLU neural networks can achieve near-optimal rates of score estimation under specific conditions.
Approximation theory for 1-Lipschitz ResNets
Davide Murari (University of Cambridge), Carola-Bibiane Schönlieb (University of Cambridge)
Ordinary Differential Equation
🎯 What it does: This study investigates the approximation capability of 1-Lipschitz residual networks (ResNet) and proves that sufficiently wide and deep networks can densely approximate any continuous 1-Lipschitz scalar function. Furthermore, it demonstrates that by inserting constrained linear mappings between residual blocks, the same density can be maintained at a fixed width, thereby providing a theoretically realizable 1-Lipschitz ResNet architecture.
AR-RAG: Autoregressive Retrieval Augmentation for Image Generation
Jingyuan Qi (Virginia Tech), Lifu Huang (UC Davis)
GenerationRetrievalTransformerImageMultimodalityRetrieval-Augmented Generation
🎯 What it does: A self-regressive retrieval-enhanced framework AR-RAG is proposed and implemented, which dynamically retrieves nearest neighbor visual information in units of image blocks during the generation process.
ArchCAD-400K: A Large-Scale CAD drawings Dataset and New Baseline for Panoptic Symbol Spotting
Ruifeng Luo (Arcplus East China Architectural Design and Research Institute Co., Ltd.), Xianzhong Zhao (Tongji University)
Object DetectionSegmentationTransformerImageMultimodalityPoint CloudBenchmark
🎯 What it does: A large-scale CAD drawing dataset ArchCAD-400k was constructed, an efficient automatic annotation pipeline was designed, and a Dual Path Symbol Localization Framework (DPSS) was proposed for panoramic symbol localization.
Architectural and Inferential Inductive Biases for Exchangeable Sequence Modeling
Daksh Mittal (Columbia University), Hongseok Namkoong (Columbia University)
TransformerSequential
🎯 What it does: This paper studies the inductive biases of inference and architecture in modeling exchangeable sequences, proving that multi-step autoregressive inference outperforms traditional single-step inference, and systematically evaluates the impact of Transformer masking schemes on exchangeability.
Are Greedy Task Orderings Better Than Random in Continual Linear Regression?
Matan Tsipory (Technion), Daniel Soudry (Technion)
OptimizationImageTabular
🎯 What it does: This paper studies the impact of task order on learning effectiveness in continuous linear regression, particularly by arranging task order through greedy strategies (maximum distance, maximum residual) and comparing it with random order.
Are Language Models Efficient Reasoners? A Perspective from Logic Programming
Andreas Opedal (ETH Zürich), Bernhard Schölkopf (MPI for Intelligent Systems)
Large Language ModelPrompt EngineeringText
🎯 What it does: Evaluate the efficiency of large language models in reasoning tasks from a logical programming perspective, mapping natural language proofs to the shortest logical reasoning paths.
Are Large Language Models Sensitive to the Motives Behind Communication?
Addison J. Wu (Princeton University), Thomas L. Griffiths (Princeton University)
GenerationTransformerLarge Language ModelPrompt EngineeringVideoTextChain-of-Thought
🎯 What it does: This study evaluates the 'motivation awareness' capability of large language models (LLMs) in identifying and assessing the motivations of information sources, and verifies their performance through three types of experiments ranging from simple to ecological contexts.
Are Large Reasoning Models Good Translation Evaluators? Analysis and Performance Boost
Runzhe Zhan (University of Macau), Derek F. Wong (University of Macau)
Large Language ModelPrompt EngineeringText
🎯 What it does: The system studied the performance of large reasoning models (LRM) as machine translation evaluators and proposed the ThinMQM method for thinking calibration of LRM through artificially simulating MQM evaluation trajectories.
Are Pixel-Wise Metrics Reliable for Computerized Tomography Reconstruction?
Tianyu Lin (Johns Hopkins University), Zongwei Zhou (Johns Hopkins University)
RestorationSegmentationDiffusion modelImageBiomedical DataComputed Tomography
🎯 What it does: Proposed an evaluation metric for sparse-angle CT reconstruction based on anatomical segmentation and a corresponding supplementary framework called CARE, which can significantly enhance the integrity of anatomical structures;
AREAL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning
Wei Fu (Tsinghua University), Yi Wu (Ant Group)
Large Language ModelReinforcement LearningTextBenchmark
🎯 What it does: An entirely asynchronous RL system called AREAL is proposed for inference training of large-scale language models, completely decoupling the generation and training phases.
ARECHO: Autoregressive Evaluation via Chain-Based Hypothesis Optimization for Speech Multi-Metric Estimation
Jiatong Shi (Carnegie Mellon University), Shinji Watanabe (Carnegie Mellon University)
TransformerAudio
🎯 What it does: This paper researches and implements a multi-indicator speech evaluation framework called ARECHO, which unifies various scales and types of evaluation indicators by tokenizing them and modeling them in a dynamic classification chain, supporting flexible reasoning for any subset of indicators.
ARGenSeg: Image Segmentation with Autoregressive Image Generation Model
Xiaolong Wang (Ant Group), JUN ZHOU
SegmentationGenerationTransformerLarge Language ModelSupervised Fine-TuningImageMultimodality
🎯 What it does: A unified framework ARGenSeg based on autoregressive image generation is proposed, which directly predicts visual tokens using a multimodal large language model (MLLM) and obtains pixel-level segmentation masks through VQ-VAE decoding, without the need for an additional segmentation head.
ARIA: Training Language Agents with Intention-driven Reward Aggregation
Ruihan Yang (Fudan University), Yanghua Xiao (Fudan University)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the ARIA method, which improves reinforcement learning for language agents in open-ended language action tasks by aggregating rewards in a low-dimensional intention space.
ARM: Adaptive Reasoning Model
Siye Wu (Fudan University), Yanghua Xiao (Fudan University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: Proposes the Adaptive Reasoning Model (ARM), allowing large language models to adaptively select from four reasoning formats (Direct Answer, Short CoT, Code, Long CoT) based on task difficulty, and achieve efficient reasoning through two-stage training (SFT + Ada-GRPO); also supports Instruction-Guided Mode and Consensus-Guided Mode.
ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction
Jiabao Lei (Chinese University of Hong Kong), Kui Jia (Chinese University of Hong Kong)
GenerationData SynthesisTransformerMesh
🎯 What it does: ARMesh is proposed—a layer-by-layer refined 3D mesh generation framework based on autoregressive Transformer, which achieves coarse-to-fine generation from a single point to a complete mesh by reversing the mesh simplification process;
Ascent Fails to Forget
Ioannis Mavrothalassitis (Ecole Polytechnique Federale de Lausanne), Volkan Cevher (Ecole Polytechnique Federale de Lausanne)
OptimizationConvolutional Neural NetworkImage
🎯 What it does: This study investigates the reasons for the failure of gradient ascent/descent hybrid methods in machine unlearning, providing both theoretical and experimental analysis.
ASDSV: Multimodal Generation Made Efficient with Approximate Speculative Diffusion and Speculative Verification
Kaijun Zhou (Shanghai Jiao Tong University), Jinyu Gu (Shanghai Jiao Tong University)
GenerationComputational EfficiencyTransformerDiffusion modelImageVideo
🎯 What it does: The ASDSV method is proposed, which uses approximate speculative diffusion and speculative validation to accelerate the generation process in multimodal diffusion model inference.
ASGO: Adaptive Structured Gradient Optimization
Kang An (Rice University), Tong Zhang (University of Illinois Urbana-Champaign)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: An adaptive structure gradient optimization algorithm ASGO is proposed, along with theoretical convergence analysis and empirical validation.
Ask a Strong LLM Judge when Your Reward Model is Uncertain
Zhenghao Xu (Georgia Institute of Technology), Tuo Zhao (Georgia Institute of Technology)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: A framework for RLHF based on uncertainty routing is proposed, where samples are handed over to a powerful LLM judge when the reward model is uncertain, improving OOD performance and reducing costs.
Assessing the quality of denoising diffusion models in Wasserstein distance: noisy score and optimal bounds
Vahan Arsenyan (ENSAE Paris), Arnak S. Dalalyan (ENSAE Paris)
RestorationGenerationData SynthesisDiffusion modelImageStochastic Differential Equation
🎯 What it does: Theoretical analysis of the quality of Denoising Diffusion Probabilistic Models (DDPM) under the Wasserstein-2 distance, providing an optimal error upper bound.
Assignments for Congestion-Averse Agents: Seeking Competitive and Envy-Free Solutions
Jiehua Chen (Institute of Logic and Computation TU Wien), Yinghui Wen (Shandong Institute of Information Technology Industry Development)
Optimization
🎯 What it does: This study investigates the resource allocation problem of congestion avoidance agents, proposes a polynomial algorithm for determining competitive (CP) allocation, and proves that finding a fair or maximum competitive allocation that meets top guarantees is NP/W[1] hard.
Association-Focused Path Aggregation for Graph Fraud Detection
Tian Qiu (Zhejiang University), Yang Gao (Zhejiang University)
Anomaly DetectionGraph Neural NetworkGraphFinance Related
🎯 What it does: This paper studies the problem of fraud detection in graph structures and proposes a novel fraud detection framework based on Graph Path Aggregation (GPA).
Asymmetric Dual Self-Distillation for 3D Self-Supervised Representation Learning
Remco F. Leijenaar (University of Groningen), Hamidreza Kasaei (University of Groningen)
Knowledge DistillationRepresentation LearningTransformerContrastive LearningPoint Cloud
🎯 What it does: Proposes the AsymDSD framework, unifying mask point modeling and invariance learning into 3D point cloud self-supervised learning.
Asymmetric Dual-Lens Video Deblurring
Zeyu Xiao (National University of Singapore), Xinchao Wang (National University of Singapore)
RestorationConvolutional Neural NetworkOptical FlowVideo
🎯 What it does: A dual-branch structure-based AsLeD-Net is designed to achieve video deblurring by utilizing the complementary information from the wide/ultra-wide lenses of mobile asynchronous dual cameras.
Asymmetric Duos: Sidekicks Improve Uncertainty
Tim G. Zhou (University of British Columbia), Geoff Pleiss (University of British Columbia)
ClassificationSupervised Fine-TuningImage
🎯 What it does: Proposes the Asymmetric Duos method, pairing large models with small companion models and enhancing performance through temperature-weighted averaging.
Asymmetric REINFORCE for off-Policy Reinforcement Learning: Balancing positive and negative rewards
Charles Arnal (Meta), Remi Munos
Reinforcement LearningTabular
🎯 What it does: This paper studies a simple offline REINFORCE algorithm (called Asymmetric REINFORCE, abbreviated as AsymRE), analyzing the impact of the baseline V on the algorithm's convergence, policy support, and performance, and conducting experimental validation on multi-armed bandit and LLM cognitive reasoning tasks.
Asymptotic theory of SGD with a general learning-rate
Or Goldreich (University of Chicago), Wei Biao Wu (University of Chicago)
OptimizationTabular
🎯 What it does: This paper proposes a unified framework that quantifies the uncertainty of online stochastic gradient descent (SGD) under arbitrary learning rate choices, particularly providing a comprehensive description of the convergence characteristics for periodic learning rates and linearly decaying to zero learning rates.
Asymptotically exact variational flows via involutive MCMC kernels
Zuheng Xu (University of British Columbia), Trevor Campbell (University of British Columbia)
OptimizationComputational EfficiencyFlow-based ModelTabular
🎯 What it does: An asynchronous MixFlow variational flow framework based on Iterative Random Functions (IRF) with reversible measure preservation has been constructed.