NeurIPS 2025 Papers — Page 16
Conference on Neural Information Processing Systems · 5275 papers
Evaluating Robustness of Monocular Depth Estimation with Procedural Scene Perturbations
John Nugent, Jia Deng (Princeton University)
Depth EstimationPoint CloudBenchmark
🎯 What it does: Systematically evaluate the robustness of monocular depth estimation models under different environmental changes through programmatically generated 3D scenes and various perturbations (camera, objects, materials, lighting, etc.).
Evaluating the Inductive Abilities of Large Language Models: Why Chain-of-Thought Reasoning Sometimes Hurts More Than Helps
Haibo Jin (University of Illinois), Haohan Wang (University of Illinois)
TransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: Evaluated the ability of large language models in inductive reasoning (implicit rule learning) through four controlled game tasks, and analyzed why chain reasoning sometimes backfires.
Eve3D: Elevating Vision Models for Enhanced 3D Surface Reconstruction via Gaussian Splatting
Jiawei Zhang (Beihang University), Matteo Poggi (University of Bologna)
Depth EstimationOptimizationGaussian SplattingPoint CloudMesh
🎯 What it does: A dense 3D reconstruction framework called Eve3D based on 3D Gaussian Splatting (3DGS) is proposed, which can achieve high-quality surface meshes and view synthesis results in a relatively short time.
Event-based HDR Structured Light
Jiacheng Fu (University of Science and Technology of China), Zhiwei Xiong (University of Science and Technology of China)
Depth EstimationPoint Cloud
🎯 What it does: The first HDR structured light 3D measurement framework specifically designed for event cameras is proposed, achieving accurate 3D reconstruction in extremely high dynamic range scenes.
Event-Driven Dynamic Scene Depth Completion
Zhiqiang Yan (National University of Singapore), Gim Hee Lee (National University of Singapore)
RestorationDepth EstimationConvolutional Neural NetworkPoint CloudBenchmark
🎯 What it does: Proposes the EventDC framework, which utilizes event cameras to complete the sparse to dense depth in dynamic scenes;
Event-Guided Consistent Video Enhancement with Modality-Adaptive Diffusion Pipeline
Kanghao Chen (Hong Kong University of Science and Technology), Ying-Cong Chen (Hong Kong University of Science and Technology)
RestorationDiffusion modelVideoMultimodalityStochastic Differential Equation
🎯 What it does: In the low-light video enhancement task, the EVDiffuser framework is proposed, which integrates RGB and event data to achieve consistent enhancement.
EventMG: Efficient Multilevel Mamba-Graph Learning for Spatiotemporal Event Representation
Sheng Wu (Fudan University), Bo Hu (Fudan University)
RecognitionObject DetectionRepresentation LearningGraph Neural NetworkVideo
🎯 What it does: Designed and implemented EventMG, a multi-level Mamba-Graph lightweight spatiotemporal representation learning framework for efficient processing and representation of event camera data.
Every Rollout Counts: Optimal Resource Allocation for Efficient Test-Time Scaling
Xinglin Wang (Beijing Institute of Technology), Kan Li (Beijing Institute of Technology)
OptimizationReinforcement LearningTabular
🎯 What it does: A theoretical optimal resource allocation method for search during testing based on a Bayesian probability model is proposed, and a direction-oriented resource allocation method (DORA) is implemented to achieve higher accuracy under different computational powers.
EverybodyDance: Bipartite Graph–Based Identity Correspondence for Multi-Character Animation
Haotian Ling (University of Science and Technology of China), Xun Yang (University of Science and Technology of China)
SegmentationGenerationDiffusion modelVideo
🎯 What it does: A framework called EverybodyDance is proposed to specifically address the identity correspondence problem in multi-character animation.
EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks
Rikuto Kotoge (SANKEN, University of Osaka), Yasushi Sakurai (SANKEN, University of Osaka)
Anomaly DetectionRecurrent Neural NetworkGraph Neural NetworkSupervised Fine-TuningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: EvoBrain has been developed, a time-then-graph neural network based on explicit dynamic multi-channel EEG graph modeling for epilepsy seizure detection and early prediction.
EVODiff: Entropy-aware Variance Optimized Diffusion Inference
Shigui Li (South China University of Technology), Delu Zeng (South China University of Technology)
GenerationOptimizationDiffusion modelImage
🎯 What it does: This paper proposes EVODiff, which optimizes the inference process of diffusion models using an information-theoretic framework, achieving reference-free conditional variance optimization by reducing conditional entropy.
EvoLM: In Search of Lost Language Model Training Dynamics
Zhenting Qi (Harvard University), Hanlin Zhang (Harvard University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: A systematic and transparent analysis of the entire process of training language models from pre-training to reinforcement learning is conducted, training and publicly releasing over 100 1B/4B decoder models, and evaluating their language modeling and reasoning capabilities.
Evolution of Information in Interactive Decision Making: A Case Study for Multi-Armed Bandits
Yuzhou Gu, Jian Qian
🎯 What it does: In the multi-armed bandit problem under a uniform prior, the evolution of information and the probability of learning success over time during the interactive decision-making process is studied, and both the complete optimal upper and lower bounds for the two are provided;
Evolutionary Multi-View Classification via Eliminating Individual Fitness Bias
Xinyan Liang (Shanxi University), Liang Du (Shanxi University)
ClassificationOptimizationKnowledge DistillationMultimodality
🎯 What it does: This study investigates the 'Fitness Evaluation Bias' (FEB) present in Evolutionary Multi-View Classification (EMVC) and proposes the introduction of an Evolutionary Navigator (EN) and a knowledge distillation loss based on Wasserstein distance (ENL) to eliminate this bias, thereby enhancing the performance of EMVC.
Evolutionary Prediction Games
Eden Saig (Technion - Israel Institute of Technology), Nir Rosenfeld (Technion - Israel Institute of Technology)
ClassificationOptimizationConvolutional Neural NetworkReinforcement LearningImageTabular
🎯 What it does: A framework for evolutionary prediction games is proposed, characterizing the co-evolution dynamics of learning algorithms and user groups in a self-selection feedback loop, and analyzing the long-term population evolution outcomes under ideal and actual conditions.
Evolutionary Reasoning Does Not Arise in Standard Usage of Protein Language Models
Yasha Ektefaie (Broad Institute of Harvard and MIT), Marinka Zitnik (Harvard Medical School)
Protein Structure PredictionTransformerSupervised Fine-TuningBiomedical Data
🎯 What it does: This paper demonstrates that existing protein language models cannot perform evolutionary inference under standard usage, and subsequently proposes and trains the PHYLA model based on tree loss to achieve evolutionary inference on multiple sequences.
EvolvedGRPO: Unlocking Reasoning in LVLMs via Progressive Instruction Evolution
Zhebei Shen (Zhejiang University), Yueting Zhuang (Zhejiang University)
OptimizationTransformerReinforcement LearningVision Language ModelDiffusion modelMultimodality
🎯 What it does: The EvolvedGRPO framework is proposed, which enhances the reasoning ability of large visual language models through progressive instruction evolution.
Evolving and Regularizing Meta-Environment Learner for Fine-Grained Few-Shot Class-Incremental Learning
Li-Jun Zhao (Shandong University), Xin-Shun Xu (Shandong University)
ClassificationMeta LearningConvolutional Neural NetworkImage
🎯 What it does: This paper proposes a Meta-Environment Learner (MEL) for fine-grained few-sample incremental learning, achieving the evolution of feature extraction and regularization of hypothesis space complexity.
EVOREFUSE: Evolutionary Prompt Optimization for Evaluation and Mitigation of LLM Over-Refusal to Pseudo-Malicious Instructions
Xiaorui Wu (Wuhan University), Zhuang Li (Royal Melbourne Institute of Technology)
OptimizationTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes an evolutionary search-based prompt optimization framework called EVOREFUSE, which is used to automatically generate pseudo-malicious instructions and construct corresponding testing and alignment datasets to help evaluate and mitigate the excessive refusal problem of LLMs.
Exact and Linear Convergence for Federated Learning under Arbitrary Client Participation is Attainable
Bicheng Ying (Google), Haibo Yang (Rochester Institute of Technology)
OptimizationFederated LearningTabular
🎯 What it does: A Federated Learning algorithm named FOCUS is proposed, which can achieve precise linear convergence under the participation of any clients;
Exact Expressive Power of Transformers with Padding
William Merrill (Allen Institute for Artificial Intelligence), Ashish Sabharwal (Allen Institute for Artificial Intelligence)
Transformer
🎯 What it does: This paper theoretically proves that after inserting padding tokens and cyclic layers, the expressive power of the Transformer can be precisely described as FO-uniform TC0 (fixed-depth padding) and FO-uniform TCd (with logarithmic-depth cycles), thereby achieving a complete representation of NC class problems.
Execution Guided Line-by-Line Code Generation
Boaz Lavon (Tel Aviv University), Lior Wolf (Tel Aviv University)
GenerationAI Code AssistantTransformerLarge Language ModelReinforcement LearningAgentic AIText
🎯 What it does: This paper proposes a real-time reasoning method that uses execution signals during code generation—Execution-Guided Classifier-Free Guidance (EG-CFG). It dynamically guides the language model to generate executable code by generating and executing candidate code line by line.
ExGra-Med: Extended Context Graph Alignment for Medical Vision-Language Models
Duy Minh Ho Nguyen, Mathias Niepert (Max Planck Research School for Intelligent Systems)
Graph Neural NetworkTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: A multi-image alignment framework (EXGRA-MED) is proposed, which enhances the semantic alignment and cross-modal consistency of medical vision-language models by jointly aligning images, instruction responses, and extended descriptions in the latent space.
Explainable Reinforcement Learning from Human Feedback to Improve Alignment
Shicheng Liu (Pennsylvania State University), Minghui Zhu (Pennsylvania State University)
Explainability and InterpretabilityReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a post-training method after RLHF, which uses interpretability techniques to identify training samples that lead to unsatisfactory outputs, and applies 'forgetting' and model fine-tuning to further enhance the alignment effect of the language model.
Explainably Safe Reinforcement Learning
Sabine Rieder (Masaryk University), Bettina Könighofer
Safty and PrivacyExplainability and InterpretabilityReinforcement LearningTabularBenchmark
🎯 What it does: An interpretable safe reinforcement learning framework is proposed, providing human-readable explanations for the safety shield through a hierarchical decision tree.
Explaining and Mitigating Crosslingual Tokenizer Inequities
Catherine Arnett (EleutherAI), Ben Bergen
CompressionSupervised Fine-TuningText
🎯 What it does: This study investigates the token premium that appears in monolingual tokenizers across different languages and explores its causes and mitigation methods through large-scale experiments.
Explaining Similarity in Vision-Language Encoders with Weighted Banzhaf Interactions
Hubert Baniecki (University of Warsaw), Przemyslaw Biecek (University of Warsaw)
Explainability and InterpretabilityVision Language ModelImageMultimodality
🎯 What it does: A second-order interaction interpretation method based on weighted Banzhaf interaction (FIXLIP) is proposed to reveal the similarity prediction of visual-language encoders.
Explaining the Law of Supply and Demand via Online Learning
Stratis Skoulakis (Aarhus University)
OptimizationReinforcement LearningTabular
🎯 What it does: This paper constructs a pricing game model to analyze how prices converge to the market clearing price when sellers adopt online learning algorithms.
Explicitly Modeling Subcortical Vision with a Neuro-Inspired Front-End Improves CNN Robustness
Lucas Piper (Instituto Superior Técnico, Universidade de Lisboa), Tiago Marques
ClassificationRecognitionAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A novel neural-inspired CNN called Early Vision Networks (EVNets) is proposed, which combines subcortical processing and VOneBlock, demonstrating its ability to enhance robustness against visual perturbations and alignment with primate vision.
Exploiting Dynamic Sparsity in Einsum
Christoph Staudt (Friedrich Schiller University Jena), Joachim Giesen (Friedrich Schiller University Jena)
OptimizationComputational EfficiencyTabularBenchmark
🎯 What it does: This study investigates the dynamic sparsity in Einsum expressions and proposes a hybrid algorithm that dynamically switches between dense and sparse representations based on the average density of the remaining tensors during execution.
Exploiting LLMs for Automatic Hypothesis Assessment via a Logit-Based Calibrated Prior
Yue Gong (Amazon Web Services), Raul Castro Fernandez (University of Chicago)
TransformerLarge Language ModelTabular
🎯 What it does: Utilize large language models (LLM) to construct prior distributions for the correlation coefficients of variable pairs to achieve automatic hypothesis evaluation.
Exploiting Task Relationships in Continual Learning via Transferability-Aware Task Embeddings
Yanru Wu (Shenzhen International Graduate School Tsinghua University), Yang Li (Shenzhen International Graduate School Tsinghua University)
Convolutional Neural NetworkTransformerSupervised Fine-TuningImage
🎯 What it does: This paper proposes a task embedding H-embedding based on the information-theoretic transferability metric H-score, and uses it to construct a hypernetwork framework for achieving continual learning without replay.
Exploiting the Asymmetric Uncertainty Structure of Pre-trained VLMs on the Unit Hypersphere
Li Ju (Uppsala University), Prashant Singh (Uppsala University)
RetrievalRepresentation LearningVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes AsymVLM, a posterior probability adaptation framework that utilizes a pre-trained visual language model (VLM) to probabilistically embed text on the unit hypersphere, enabling cross-modal retrieval and uncertainty quantification.
Exploiting Vocabulary Frequency Imbalance in Language Model Pre-training
Woojin Chung (Korea Advanced Institute of Science and Technology), Jeonghoon Kim (NAVER Cloud)
TransformerLarge Language ModelText
🎯 What it does: The study investigates the impact of expanding the vocabulary size from 24K to 196K on pre-trained language models, systematically comparing the effects of vocabulary size on text Kolmogorov complexity, word frequency imbalance, cross-entropy, and downstream task accuracy.
Exploration from a Primal-Dual Lens: Value-Incentivized Actor-Critic Methods for Sample-Efficient Online RL
Tong Yang (Carnegie Mellon University), Yuejie Chi (Yale University)
Reinforcement Learning
🎯 What it does: This paper proposes a new value incentive-based Actor-Critic (VAC) algorithm, which utilizes the primal-dual Lagrangian framework to design a single, easily optimizable objective function, achieving a unified optimization of exploration and exploitation in online reinforcement learning without the need for a complex two-layer structure.
Exploration via Feature Perturbation in Contextual Bandits
Seouh-won Yi (Seoul National University), Min-hwan Oh (Seoul National University)
Recommendation SystemReinforcement Learning from Human FeedbackGraph Neural NetworkReinforcement LearningTabular
🎯 What it does: Feature Perturbation (FP) is proposed as a random exploration strategy for contextual bandits, achieving exploration by adding noise to the feature vector during decision-making;
Explore In-Context Message Passing Operator for Graph Neural Networks in A Mean Field Game
Tingting Dan (University of North Carolina), Guorong Wu (University of North Carolina)
ClassificationOptimizationGraph Neural NetworkReinforcement LearningGraphBiomedical DataAlzheimer's Disease
🎯 What it does: Modeling graph neural networks as Mean-Field Games (MFG) and deriving learnable 'context-selective message passing operators' through a variational framework to achieve an adaptive GNN mechanism; simultaneously presenting an end-to-end deep model Game-of-GNN based on Hamiltonian flow.
Exploring and Exploiting Model Uncertainty in Bayesian Optimization
Zishi Zhang (Peking University), Yijie Peng (Peking University)
OptimizationHyperparameter SearchTabularTime SeriesFinance Related
🎯 What it does: A new Bayesian optimization surrogate model, ∞-GP, is proposed, which utilizes a spatial Dirichlet process mixed Gaussian process to simultaneously quantify value uncertainty and model uncertainty, and implements Thompson Sampling to achieve exploration and exploitation in the distribution space.
Exploring and Leveraging Class Vectors for Classifier Editing
Jaeik Kim (Seoul National University), Jaeyoung Do (Seoul National University)
ClassificationDomain AdaptationAdversarial AttackTransformerSupervised Fine-TuningImage
🎯 What it does: Class Vectors are proposed for category-level editing of image classifiers in the absence of training or with limited data, supporting feature space adjustments or weight space mappings.
Exploring Data Scaling Trends and Effects in Reinforcement Learning from Human Feedback
Wei Shen (ByteDance), Lin Yan (ByteDance)
Reinforcement Learning from Human FeedbackReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper discusses the bottleneck of data scale expansion in RLHF and proposes a hybrid reward system (RTV + GenRM) and a Pre-PPO prompt selection strategy to enhance model performance and response diversity.
Exploring Diffusion Transformer Designs via Grafting
Keshigeyan Chandrasegaran (Stanford University), Li Fei-Fei (Stanford University)
GenerationData SynthesisComputational EfficiencyKnowledge DistillationTransformerDiffusion modelImage
🎯 What it does: Proposed and implemented the grafting method, which quickly constructs new model architectures using a pre-trained Diffusion Transformer (DiT) by replacing or rearranging internal operators, and evaluates its effectiveness on various tasks.
Exploring Landscapes for Better Minima along Valleys
Tong Zhao (Chinese Academy of Sciences), Weile Jia (Chinese Academy of Sciences)
OptimizationConvolutional Neural NetworkTransformerReinforcement LearningImage
🎯 What it does: A gradient optimizer adapter named 'E' is proposed, which allows the optimizer to continue exploring along the loss valley after reaching a local minimum, thereby seeking lower and flatter minima, particularly suitable for large-batch training.
Exploring Neural Granger Causality with xLSTMs: Unveiling Temporal Dependencies in Complex Data
Harsh Poonia (Carnegie Mellon University), Devendra Singh Dhami (TU Eindhoven)
Recurrent Neural NetworkTime SeriesSequentialFinance Related
🎯 What it does: This paper studies a neural Granger causality model based on xLSTM, called GC-xLSTM, to discover Granger causality in complex time series data.
Exploring Polyglot Harmony: On Multilingual Data Allocation for Large Language Models Pretraining
Ping Guo (ByteDance), Yin Zheng (ByteDance)
OptimizationTransformerLarge Language ModelText
🎯 What it does: The CLIMB framework is proposed, which models the cross-lingual interaction-aware language ratio to predict the validation loss of multilingual pre-trained models and automatically find the optimal language allocation.
Exploring Semantic-constrained Adversarial Example with Instruction Uncertainty Reduction
Jin Hu (Beihang University), Xianglong Liu (Beihang University)
GenerationAdversarial AttackDiffusion modelGaussian SplattingImagePoint Cloud
🎯 What it does: A multi-dimensional instruction uncertainty reduction framework (InSUR) is proposed to directly generate adversarial examples (SemanticAE) that comply with semantic constraints from natural language instructions, achieving reference-free 3D SemanticAE generation for the first time.
Exploring Structural Degradation in Dense Representations for Self-supervised Learning
Siran Dai (Institute of Information Engineering, Chinese Academy of Sciences), Qingming Huang (University of Chinese Academy of Sciences)
SegmentationDepth EstimationContrastive LearningImage
🎯 What it does: A systematic study of the phenomenon of performance degradation over training time in self-supervised learning for dense tasks is conducted, and a label-free evaluation metric DSE based on inter-class margin and effective dimension is proposed to predict and mitigate this degradation.
Exploring the Design Space of Diffusion Bridge Models
Shaorong Zhang (University of California Riverside), Greg Ver Steeg (University of California Riverside)
Image TranslationGenerationData SynthesisDiffusion modelImageStochastic Differential Equation
🎯 What it does: Proposed and implemented the Endpoint Conditional Stochastic Interpolation (ECSI) model, which unifies and expands the design space of the diffusion bridge model, enhances sampling efficiency and image quality, and addresses the issue of insufficient conditional diversity in one-to-many translations.
Exploring the limits of strong membership inference attacks on large language models
Jamie Hayes (Google DeepMind), A. Feder Cooper (Stanford University)
Adversarial AttackTransformerLarge Language ModelText
🎯 What it does: This study investigates the feasibility and limits of high-intensity membership inference attacks (LiRA) on large-scale pre-trained language models.
Exploring the Limits of Vision-Language-Action Manipulation in Cross-task Generalization
Jiaming Zhou (Hong Kong University of Science and Technology), Junwei Liang (Hong Kong University of Science and Technology)
Robotic IntelligenceTransformerLarge Language ModelReinforcement LearningVision-Language-Action ModelDiffusion modelMultimodalityBenchmark
🎯 What it does: This paper proposes the AGNOSTOS benchmark for rigorously evaluating the zero-shot cross-task generalization ability of visual-language-action (VLA) models on unseen tasks, and based on this benchmark, introduces the X-ICM method, which enhances cross-task zero-shot performance through context learning of large language models and dynamically guided sample selection.
Exploring the Noise Robustness of Online Conformal Prediction
HuaJun Xi, Hongxin Wei (Southern University of Science and Technology)
ClassificationSupervised Fine-TuningImage
🎯 What it does: This study investigates the robustness of online synthetic prediction in the presence of uniform label noise.
Exploring the Translation Mechanism of Large Language Models
Hongbin Zhang (Harbin Institute of Technology), Min Zhang (Harbin Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A contrastive vocabulary translation dataset was constructed, and causal analysis of the LLM translation mechanism was conducted using the subspace intervention path patch method. It was found that a small number of attention heads and MLP drive translation, and targeted fine-tuning was validated to improve performance.
Exploring Tradeoffs through Mode Connectivity for Multi-Task Learning
Zhipeng Zhou (Nanyang Technological University), Chunyan Miao (Nanyang Technological University)
SegmentationAutonomous DrivingOptimizationImage
🎯 What it does: Exploring the optimal trade-off of multi-task learning through mode connectivity, the EXTRA method is proposed, optimizing the curve rather than the endpoints, using NURBS curves and sequentially aware objectives;
ExPO: Unlocking Hard Reasoning with Self-Explanation-Guided Reinforcement Learning
Ruiyang Zhou (University of Texas at Austin), Liu Leqi (University of Texas at Austin)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: Proposes the ExPO framework, which utilizes self-explanation based on correct answers to generate positive samples, combined with DPO and GRPO for RL fine-tuning, to enhance the performance of large language models on difficult reasoning tasks.
Exponential Convergence Guarantees for Iterative Markovian Fitting
Marta Gentiloni Silveri (Ecole Polytechnique), Alain Oliviero Durmus (Ecole Polytechnique)
Physics Related
🎯 What it does: This paper provides a non-asymptotic exponential convergence guarantee for the Iterative Markov Fitting (IMF) algorithm in the Schrödinger bridge problem, covering both strong and weak log-concave cases.
Exponential Dynamic Energy Network for High Capacity Sequence Memory
Arjun Karuvally (Salk Institute for Biological Studies), Hava T Siegelmann
Sequential
🎯 What it does: Proposed and analyzed EDEN (Exponential Dynamic Energy Network), achieving high-capacity sequential memory under the energy paradigm;
Extracting task-relevant preserved dynamics from contrastive aligned neural recordings
Yiqi Jiang (Stanford University), Mark Schnitzer (Stanford University)
Representation LearningContrastive LearningMultimodalityTime Series
🎯 What it does: Proposes the CANDY framework, which uses contrastive learning and linear dynamical systems for joint training to align neural embeddings across sessions and extract task-relevant conservative dynamics;
Extragradient Method for $(L_0, L_1)$-Lipschitz Root-finding Problems
Sayantan Choudhury (Johns Hopkins University), Nicolas Loizou (Johns Hopkins University)
OptimizationReinforcement Learning
🎯 What it does: This paper studies the Extragradient (EG) method under the α-symmetric (L, L₀, L₁)-Lipschitz condition, proposes an adaptive step size strategy, and provides convergence analysis for strongly monotone, monotone, and weak Minty problems, achieving linear or sublinear convergence rates.
Extrapolation by Association: Length Generalization Transfer In Transformers
Ziyang Cai (University of Wisconsin Madison), Dimitris Papailiopoulos (University of Wisconsin Madison)
TransformerLarge Language ModelSupervised Fine-TuningTextSequential
🎯 What it does: This study investigates the cross-task transfer effects of Transformer in length generalization, demonstrating that training alongside longer auxiliary tasks enables the model to infer longer inputs in the main task, and verifies that natural language pre-training can also provide similar auxiliary tasks.
Extremely Simple Multimodal Outlier Synthesis for Out-of-Distribution Detection and Segmentation
Moru Liu (Technical University of Munich), Mario Trapp (Technical University of Munich)
SegmentationAnomaly DetectionAutonomous DrivingMultimodalityPoint Cloud
🎯 What it does: A very simple and fast multimodal anomaly synthesis method called Feature Mixing is proposed for OOD detection and segmentation.
Eyes Wide Open: Ego Proactive Video-LLM for Streaming Video
Yulin Zhang (ShanghaiTech University), Sibei Yang (Sun Yat-sen University)
GenerationRetrievalCompressionComputational EfficiencyTransformerLarge Language ModelVision Language ModelVideoBenchmarkRetrieval-Augmented Generation
🎯 What it does: A system has been constructed that can actively and synchronously answer multi-turn questions in first-person perspective video streams, and a corresponding evaluation benchmark ESTP-Bench and evaluation metric ESTP-F1 have been proposed.
F-Adapter: Frequency-Adaptive Parameter-Efficient Fine-Tuning in Scientific Machine Learning
Hangwei Zhang (University of Hong Kong), Difan Zou (University of Hong Kong)
TransformerSupervised Fine-TuningTime SeriesPhysics Related
🎯 What it does: This paper systematically studies the parameter-efficient fine-tuning (PEFT) of pre-trained large operator models and proposes a frequency-adaptive F-Adapter module.
FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens
Chao Wang (University of Science and Technology of China), Yanyong Zhang (University of Science and Technology of China)
Recommendation SystemTransformerAuto EncoderContrastive LearningTabular
🎯 What it does: The FACE framework is proposed, which maps the continuous embeddings of collaborative filtering models to discrete semantic tokens in the vocabulary of pre-trained LLMs, achieving semantic alignment through contrastive learning.
FACE: Faithful Automatic Concept Extraction
Dipkamal Bhusal (Rochester Institute of Technology), Nidhi Rastogi (Rochester Institute of Technology)
OptimizationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: A new conceptual explanation framework called FACE is proposed, which utilizes NMF and incorporates KL divergence regularization to ensure that the extracted concepts are consistent with the model predictions, thereby improving the reliability of the explanations.
Fact-R1: Towards Explainable Video Misinformation Detection with Deep Reasoning
Fanrui Zhang (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
Anomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelReinforcement LearningVideoTextMultimodalityChain-of-Thought
🎯 What it does: A large-scale video misinformation detection framework, Fact-R1, is proposed, and the first FakeVV dataset containing 100k news video-text pairs is constructed.
FaCT: Faithful Concept Traces for Explaining Neural Network Decisions
Amin Parchami-Araghi (Max Planck Institute for Informatics), Bernt Schiele (Max Planck Institute for Informatics)
Explainability and InterpretabilityConvolutional Neural NetworkTransformerAuto EncoderContrastive LearningImage
🎯 What it does: The FaCT model is proposed, utilizing B-cos networks and unbiased sparse autoencoders to achieve concept-level trustworthy explanations across all layers, with concept contributions directly applied to predictions.
FACT: Mitigating Inconsistent Hallucinations in LLMs via Fact-Driven Alternating Code-Text Training
Xinxin You (Tsinghua University), Xien Liu (Tsinghua University)
GenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Designed and implemented the FACT framework to reduce the inconsistency hallucination of LLMs through alternating training of text and code.
Factor Decorrelation Enhanced Data Removal from Deep Predictive Models
Wenhao Yang (Wuhan University of Technology), Kaize Shi (University of Southern Queensland)
ClassificationComputational EfficiencyData-Centric LearningImageText
🎯 What it does: An efficient data deletion framework named DecoRemoval is proposed, which can maintain model prediction performance in out-of-distribution (OOD) scenarios;
Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation
Guoqing Hu (University of Science and Technology of China), Xiang Wang (University of Science and Technology of China)
Recommendation SystemRecurrent Neural NetworkDiffusion modelTabularSequential
🎯 What it does: A recommendation framework called PreferGrow based on discrete diffusion is proposed, which directly models the relative preference ratios between items by 'diluting' user preferences during the forward process (replacing liked items with negative samples) and gradually 'restoring' user preferences during the backward process.
FADRM: Fast and Accurate Data Residual Matching for Dataset Distillation
Jiacheng Cui (MBZUAI), Zhiqiang Shen (MBZUAI)
Computational EfficiencyKnowledge DistillationImage
🎯 What it does: This paper proposes the FADRM framework, which achieves efficient data distillation through data-level residual matching.
Failure by Interference: Language Models Make Balanced Parentheses Errors When Faulty Mechanisms Overshadow Sound Ones
Daking Rai (George Mason University), Ziyu Yao (George Mason University)
GenerationTransformerLarge Language ModelText
🎯 What it does: Analyzes the error mechanisms of large language models in generating balanced parentheses and proposes the RASTEER method, which enhances performance by ranking and weighting the model's internal attention heads.
Failure Prediction at Runtime for Generative Robot Policies
Ralf Römer (Technical University of Munich), Angela P. Schoellig (Technical University of Munich)
Robotic IntelligenceReinforcement LearningDiffusion modelMultimodality
🎯 What it does: A framework named FIPER is proposed, which can predict the failure of generative imitation learning strategies during robot operation without requiring any failure data.
Fair Continuous Resource Allocation with Equality of Impact
Blossom Metevier (University of Massachusetts), Philip S. Thomas (University of Massachusetts)
OptimizationTabular
🎯 What it does: A framework for fair resource allocation based on group influence (Equality of Impact) is proposed, along with noise-free and noisy online learning algorithms for continuous resource allocation in a diminishing marginal returns environment.
Fair Cooperation in Mixed-Motive Games via Conflict-Aware Gradient Adjustment
Woojun Kim (Carnegie Mellon University), Katia P. Sycara
Reinforcement LearningSequential
🎯 What it does: The FCGrad method is proposed, which utilizes conflict detection and gradient projection to dynamically balance individual and collective rewards in multi-agent reinforcement learning with mixed motivations, enhancing cooperation while ensuring fairness.
Fair Deepfake Detectors Can Generalize
Harry Cheng (National University of Singapore), Mohan Kankanhalli (National University of Singapore)
ClassificationDomain AdaptationConvolutional Neural NetworkGenerative Adversarial NetworkImageVideo
🎯 What it does: A framework has been constructed that simultaneously enhances the fairness and generalization performance of deepfake detection using causal relationships.
Fair Matroid Selection
Kiarash Banihashem (University of Maryland), Danny Mittal (University of Maryland)
Optimization
🎯 What it does: This paper studies an online fair matroid selection problem, aiming to maximize the minimum probability of selecting all elements.
Fair Minimum Labeling: Efficient Temporal Network Activations for Reachability and Equity
Lutz Oettershagen (University of Liverpool), Othon Michail (University of Liverpool)
Graph
🎯 What it does: This paper proposes the Fair Minimum Labeling (FML) problem, provides its definition, proves its NP-hardness, and designs an approximate algorithm based on tree embedding to minimize time labels while ensuring group fairness.
Fair Representation Learning with Controllable High Confidence Guarantees via Adversarial Inference
Yuhong Luo (Rutgers University), Przemyslaw A. Grabowicz (University of Massachusetts)
Representation LearningAuto EncoderGenerative Adversarial NetworkTabular
🎯 What it does: A framework (FRG) is proposed to achieve controllable high-confidence fairness in representation learning, ensuring that the disparity in downstream tasks does not exceed a user-specified threshold ε, with a confidence level of 1-δ.
FairDD: Fair Dataset Distillation
Qihang Zhou (Zhejiang University), Jiming Chen (Zhejiang University)
Knowledge DistillationData-Centric LearningImage
🎯 What it does: A fair dataset distillation framework called FairDD is proposed, which utilizes synchronous matching to maintain fairness on protected attributes in the distilled small dataset while preserving the accuracy of the target attributes.
FairDICE: Fairness-Driven Offline Multi-Objective Reinforcement Learning
Woosung Kim (Korea University), Byung-Jun Lee (Gauss Labs Inc.)
OptimizationReinforcement Learning
🎯 What it does: The paper presents FairDICE, a multi-objective reinforcement learning framework that can directly optimize nonlinear welfare objectives (such as Nash social welfare) on offline datasets.
FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models
Zihao Fu (Chinese University of Hong Kong), Chris Russell (University of Oxford)
GenerationDiffusion modelImageText
🎯 What it does: This paper proposes FairImagen, a post-processing debiasing method that eliminates social biases such as gender and race in text-to-image generation directly in the CLIP embedding space of Stable Diffusion through FairPCA projection, noise injection, and cross-attribute joint projection, without the need to retrain the model or modify prompts.
Fairness under Competition
Ronen Gradwohl (University of Haifa), Moshe Tennenholtz (Technion - Israel Institute of Technology)
ClassificationRecommendation SystemOptimizationTabularFinance Related
🎯 What it does: This paper studies the impact of multiple institutions using fair algorithms on the overall fairness of the ecosystem in a competitive environment. It demonstrates that even if a single algorithm meets fairness standards, the overall ecosystem may still be unfair, and explores mechanisms by which fairness adjustments may lead to a decrease in overall fairness.
Fairness-aware Anomaly Detection via Fair Projection
Feng Xiao (Chinese University of Hong Kong), Jicong Fan (Chinese University of Hong Kong)
Anomaly DetectionTabular
🎯 What it does: An unsupervised anomaly detection method called FairAD is proposed, which achieves a synergistic optimization of detection accuracy and group fairness by learning to project onto a unified, concise, and compact target distribution.
Fairness-aware Bayes Optimal Functional Classification
Xiaoyu Hu (Xi'an Jiaotong University), Yi Yu (Warwick University)
ClassificationTabular
🎯 What it does: Construct a fair Bayes optimal classifier on functional data and propose a Fair-FLDA post-processing algorithm based on functional linear discrimination.
Fairness-Regularized Online Optimization with Switching Costs
Pengfei Li (Rochester Institute of Technology), Shaolei Ren (University of California)
OptimizationTime Series
🎯 What it does: This paper proposes a smooth online convex optimization framework with fair regularization, introducing the FairOBD algorithm to address scenarios that need to consider hit costs, switching costs, and long-term fairness costs simultaneously, achieving a dynamic balance of the three costs at each step and ensuring long-term fairness.
FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA
Songqi Zhou (Tsinghua University), Benben Jiang (Tsinghua University)
ClassificationRepresentation LearningContrastive LearningImageText
🎯 What it does: The FairNet framework is proposed, which utilizes an internal bias detector to dynamically trigger the LoRA module, making representation-level fairness corrections only for instances predicted to have potential bias, thereby improving the performance of disadvantaged groups while maintaining overall accuracy.
Fairshare Data Pricing via Data Valuation for Large Language Models
Luyang Zhang (Carnegie Mellon University), Chenyan Xiong (Carnegie Mellon University)
Large Language ModelTextPhysics Related
🎯 What it does: A fair sharing pricing mechanism based on data value is proposed and validated to achieve fair, transparent, and sustainable transactions in the LLM training data market.
Faithful Dynamic Imitation Learning from Human Intervention with Dynamic Regret Minimization
Bo Ling (Southeast University), Yan Lyu (Southeast University)
Autonomous DrivingOptimizationReinforcement LearningMultimodalityBenchmark
🎯 What it does: Proposes Faithful Dynamic Imitation Learning (FaithDaIL), which learns agents through dynamic regression minimization and unbiased expert imitation targets under real-time human intervention.
Faithful Group Shapley Value
Kiljae Lee (Ohio State University), Yuan Zhang (Ohio State University)
Explainability and InterpretabilityAdversarial AttackTabular
🎯 What it does: This paper proposes the Faithful Group Shapley Value (FGSV) and provides an algorithm that can directly and quickly approximate FGSV; it defends against group-level data valuation attacks using FGSV, addressing the vulnerability of traditional Group Shapley Value to shell company attacks.
FALCON: An ML Framework for Fully Automated Layout-Constrained Analog Circuit Design
Asal Mehradfar (University of Southern California), Salman Avestimehr (University of Southern California)
OptimizationGraph Neural NetworkTabular
🎯 What it does: The FALCON framework is proposed to achieve full automation of analog circuit design from performance specifications to complete layout constraints.
FALCON: Fine-grained Activation Manipulation by Contrastive Orthogonal Unalignment for Large Language Model
Jinwei Hu (University of Liverpool), Xiaowei Huang (University of Liverpool)
OptimizationSafty and PrivacyTransformerLarge Language ModelContrastive LearningText
🎯 What it does: A fine-grained activation manipulation method for large language models, FALCON, is proposed, which utilizes information-theoretic metrics to select parameters and combines contrastive learning with gradient orthogonal projection to achieve secure knowledge forgetting.
FALQON: Accelerating LoRA Fine-tuning with Low-Bit Floating-Point Arithmetic
Kanghyun Choi (Seoul National University), Jinho Lee (Seoul National University)
OptimizationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes an FP8 quantization-based LoRA fine-tuning framework called FALQON, which directly integrates the LoRA adapter into the FP8 quantized model backbone, redesigning the forward/backward computation paths to eliminate the quantization overhead caused by small-sized tensors in traditional LoRA fine-tuning, thereby significantly accelerating the fine-tuning of large language models.
FAME: Adaptive Functional Attention with Expert Routing for Function-on-Function Regression
Yifei Gao (Tsinghua University), Chen Zhang (Tsinghua University)
Mixture of ExpertsTime SeriesStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: An end-to-end function-to-function regression model called FAME is proposed, which learns mappings directly in infinite-dimensional function spaces using continuous attention and mixture of experts techniques.
FANS: A Flatness-Aware Network Structure for Generalization in Offline Reinforcement Learning
Da Wang (Shanxi University), Jiye Liang (Shanxi University)
Reinforcement LearningTabularBenchmark
🎯 What it does: This paper proposes the FANS (Flatness-Aware Network Structure) framework, which achieves a preference for flat minimization through structured network design in offline reinforcement learning, thereby enhancing policy generalization and robustness.
Fantastic Features and Where to Find Them: A Probing Method to combine Features from Multiple Foundation Models
Benjamin Ramtoula (University of Oxford), Daniele De Martini (University of Oxford)
ClassificationComputational EfficiencyRepresentation LearningTransformerImage
🎯 What it does: A detection adapter named ComBo is proposed, which can utilize intermediate feature layers from various frozen base models, compress them into low-dimensional representations, and perform downstream task predictions through a lightweight Transformer.
FAPEX: Fractional Amplitude-Phase Expressor for Robust Cross-Subject Seizure Prediction
Ruizhe Zheng (Fudan University), Yuguo Yu (Fudan University)
Anomaly DetectionMultimodalityBiomedical DataElectrocardiogram
🎯 What it does: A cross-subject and cross-device epilepsy seizure prediction framework named FAPEX is proposed, which can extract pre-seizure signals from multimodal EEG data and issue warnings.
Far from the Shallow: Brain-Predictive Reasoning Embedding through Residual Disentanglement
Linyang He (Columbia University), Nima Mesgarani (Columbia University)
TransformerLarge Language ModelTextElectrocardiogramAudio
🎯 What it does: This paper proposes a residual decoupling method that splits the hidden layer representations of large language models into four almost orthogonal embeddings: lexical, syntactic, semantic, and reasoning, and uses these embeddings for linear encoding of neural electrode recordings.
Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning
Yeongbin Seo (Yonsei University), Jinyoung Yeo (Yonsei University)
GenerationOptimizationComputational EfficiencyLarge Language ModelSupervised Fine-TuningDiffusion modelText
🎯 What it does: This paper addresses the long decoding window (LDW) problem of diffusion language models in open text generation, proposing two methods: Convolutional Decoding (Conv) and Rule-based Negative Sample Fine-tuning (R2FT), which significantly enhance the model's fluency and consistency while maintaining high speed and supporting bidirectional generation.
Fast attention mechanisms: a tale of parallelism
Jingwen Liu (Columbia University), Daniel Hsu (Columbia University)
Computational EfficiencyKnowledge DistillationTransformerTabular
🎯 What it does: An efficient attention mechanism called ANNA-transformer based on approximate nearest neighbor search is proposed, and it is proven to maintain the same expressive power as the standard Transformer while completing computations in sub-quadratic time.
Fast Computation and Optimization for Opinion-Based Quantities of Friedkin-Johnsen Model
Haoxin Sun (Fudan University), Zhongzhi Zhang (Fudan University)
OptimizationComputational EfficiencyGraph
🎯 What it does: This paper proposes a fast algorithm based on partial root forest sampling to estimate and optimize public opinion-related quantities under the Friedkin-Johnsen model;
Fast constrained sampling in pre-trained diffusion models
Alexandros Graikos (Stony Brook University), Dimitris Samaras (Stony Brook University)
RestorationGenerationSuper ResolutionDiffusion modelImage
🎯 What it does: A fast constrained sampling algorithm is proposed on large-scale pre-trained diffusion models, achieving high-quality generation of linear and nonlinear constraints using approximate Newton steps without the need for backpropagation through the denoising network.