NeurIPS 2025 Papers — Page 46
Conference on Neural Information Processing Systems · 5275 papers
Systematic Reward Gap Optimization for Mitigating VLM Hallucinations
Lehan He (Beihang University), Jing Shao (Shanghai AI Laboratory)
OptimizationTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Proposes the Topic-level Preference Rewriting (TPR) framework, which systematically optimizes the reward gap of visual language models (VLM) during the data collection phase, significantly reducing visual hallucinations;
T-norm Selection for Object Detection in Autonomous Driving with Logical Constraints
Thomas Eiter (Vienna University of Technology), Sota Moriyama (National Institute of Informatics)
Object DetectionAutonomous DrivingImage
🎯 What it does: This paper proposes a neural-symbolic framework MOD-ECL, which integrates logical constraints into multi-label object detection in autonomous driving using t-norms, and adaptively selects the optimal t-norm and dynamically adjusts the regularization coefficient λ of the constraint loss during the training process.
T-REGS: Minimum Spanning Tree Regularization for Self-Supervised Learning
Julie Mordacq (Inria), Steve Oudot (Inria)
RetrievalRepresentation LearningContrastive LearningImageMultimodality
🎯 What it does: A regularization method T-REG, based on the combination of maximizing the length of the Minimum Spanning Tree (MST) and spherical constraints, is proposed. It can simultaneously suppress dimensional collapse and enhance the uniformity of embeddings. An extension for self-supervised learning, T-REGS, is also introduced.
T-SHIRT: Token-Selective Hierarchical Data Selection for Instruction Tuning
Yanjun Fu (University of Maryland), Sanghamitra Dutta (University of Maryland)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A data selection framework for instruction fine-tuning called T-SHIRT is proposed, which combines token-level information with sample neighborhood robustness assessment.
T2I-R1: Reinforcing Image Generation with Collaborative Semantic-level and Token-level CoT
Dongzhi Jiang (Chinese University of Hong Kong), Hongsheng Li (Chinese University of Hong Kong)
GenerationOptimizationReinforcement LearningImageTextBenchmarkChain-of-Thought
🎯 What it does: This paper proposes a text-to-image generation model T2I-R1 based on dual-layer chain-of-thought (semantic-level CoT and token-level CoT), and achieves joint optimization of the two levels of thinking through reinforcement learning.
T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models
Jindong Yang (University of Science and Technology of China), Kejiang Chen (University of Science and Technology of China)
GenerationData SynthesisDiffusion modelImageOrdinary Differential Equation
🎯 What it does: A two-stage watermarking scheme T2SMark based on Tail-Truncated Sampling is proposed, specifically designed for diffusion model image generation, balancing watermark robustness and generation diversity.
T2V-OptJail: Discrete Prompt Optimization for Text-to-Video Jailbreak Attacks
Jiayang Liu (Nanyang Technological University), Siew Kei Lam
OptimizationAdversarial AttackTransformerLarge Language ModelPrompt EngineeringVideoText
🎯 What it does: A text-to-video (T2V) jailbreak attack framework called T2V-OptJail is designed based on discrete optimization, capable of bypassing security filters of T2V models through prompt rewriting and iterative search while maintaining semantic consistency.
TabDPT: Scaling Tabular Foundation Models on Real Data
Junwei Ma (Layer 6 AI), Maksims Volkovs (Layer 6 AI)
ClassificationTransformerTabular
🎯 What it does: This paper presents TabDPT, a table-based model based on row-level Transformers, which utilizes retrieval-based context and self-supervised learning for pre-training, enabling direct inference on unseen classification and regression tasks.
Table as a Modality for Large Language Models
Liyao Li (Zhejiang University), Junbo Zhao (Zhejiang University)
Graph Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringMultimodalityTabularBenchmark
🎯 What it does: The TAMO framework is proposed, which integrates tables as an independent modality with large language models and retains the structural information of tables through a hypergraph encoder; simultaneously, a StructQA benchmark dataset focused on table structure understanding is constructed.
Table2LaTeX-RL: High-Fidelity LaTeX Code Generation from Table Images via Reinforced Multimodal Language Models
Jun Ling (University of Electronic Science and Technology of China), Peng Wang (University of Electronic Science and Technology of China)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTabular
🎯 What it does: This study addresses the task of generating LaTeX code from table images, aiming to automatically reconstruct high-quality, publication-ready tables.
TabSTAR: A Tabular Foundation Model for Tabular Data with Text Fields
Alan Arazi (Technion - Israel Institute of Technology), Roi Reichart
ClassificationTransformerLarge Language ModelSupervised Fine-TuningTabular
🎯 What it does: A foundational model called TabSTAR is proposed and implemented, which is oriented towards table text features, supports cross-dataset transfer learning, and can quickly adapt to downstream tasks after multi-task pre-training.
Tabula: A Tabular Self-Supervised Foundation Model for Single-Cell Transcriptomics
Jiayuan Ding (Stanford University), Xiaojie Qiu (Stanford University)
Federated LearningSafty and PrivacyTransformerContrastive LearningTabularBiomedical Data
🎯 What it does: TABULA, a self-supervised tabular pre-training model for single-cell transcriptomics, has been developed, achieving privacy-preserving pre-training and fine-tuning for downstream tasks such as cell type annotation, gene imputation, and gene perturbation prediction within a federated learning framework.
Taccel: Scaling Up Vision-based Tactile Robotics via High-performance GPU Simulation
Yuyang Li (Peking University), Siyuan Huang (University of California, Los Angeles)
Robotic IntelligenceReinforcement LearningPoint Cloud
🎯 What it does: A high-performance visual-tactile simulation platform named Taccel has been developed, capable of real-time and accurate simulation of soft and rigid body interactions between robots, visual-tactile sensors (VBTS), and objects, generating high-quality tactile signals.
Tackling Biased Evaluators in Dueling Bandits
Ming Tang (Southern University of Science and Technology), Chao Huang (Montclair State University)
Recommendation SystemOptimization
🎯 What it does: This paper proposes a dueling bandits framework with biased evaluators and presents two UCB-based algorithms that achieve optimal returns under both known and unknown evaluator bias conditions.
Tackling Continual Offline RL through Selective Weights Activation on Aligned Spaces
Jifeng Hu (Jilin University), Dacheng Tao (Nanyang Technological University)
Reinforcement LearningDiffusion modelSequential
🎯 What it does: A continuous offline reinforcement learning framework VQ-CD is proposed, which is suitable for arbitrary state and action spaces, capable of continuous learning in task sequences while retaining old knowledge.
Tackling Feature-Classifier Mismatch in Federated Learning via Prompt-Driven Feature Transformation
Xinghao Wu (Beihang University), Jing Yuan (University of North Texas)
Federated LearningPrompt EngineeringContrastive LearningImage
🎯 What it does: This paper proposes FedPFT, a method that addresses the feature-classifier mismatch problem during the federated learning training process through personalized prompts and a global self-attention feature transformation module, thereby achieving a better personalized model.
TADA: Improved Diffusion Sampling with Training-free Augmented DynAmics
Tianrong Chen (Apple), Shuangfei Zhai (Apple)
GenerationData SynthesisComputational EfficiencyDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: TADA is proposed, a training-independent diffusion sampling method that achieves fast sampling using higher-dimensional initial noise and momentum dynamics.
TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling
Yuancheng Wang (Chinese University of Hong Kong), Zhizheng Wu (Chinese University of Hong Kong)
GenerationCompressionTransformerDiffusion modelAudio
🎯 What it does: Proposed TaDiCodec and implemented an end-to-end text-aware diffusion speech tokenizer.
TAI3: Testing Agent Integrity in Interpreting User Intent
Shiwei Feng (Purdue University), Xiangyu Zhang (Purdue University)
TransformerLarge Language ModelAgentic AITextFinance Related
🎯 What it does: A framework TAI3 has been developed for testing the integrity of LLM agent intentions, utilizing semantic partitioning of API parameters, intention-preserving mutations, and strategy memory to identify misunderstandings of user intentions by agents when executing natural language instructions.
Tail-Optimized Caching for LLM Inference
Wenxin Zhang (Columbia Business School), Tianyi Peng (Columbia Business School)
OptimizationTransformerLarge Language ModelText
🎯 What it does: A tail latency optimization caching strategy for large language model inference (Tail-Optimized LRU) is proposed and implemented, achieving intelligent pruning of KV cache by adding two lines of code on top of LRU.
Talk2Event: Grounded Understanding of Dynamic Scenes from Event Cameras
Lingdong Kong (National University of Singapore), Benoit R Cottereau
Object DetectionAutonomous DrivingTransformerMixture of ExpertsImageText
🎯 What it does: This paper presents the Talk2Event dataset and the EventRefer framework, achieving natural language-driven object localization based on event cameras.
TAMI: Taming Heterogeneity in Temporal Interactions for Temporal Graph Link Prediction
Zhongyi Yu (Beijing Normal-Hong Kong Baptist University), Weipeng Zhuo (Beijing Normal-Hong Kong Baptist University)
Graph Neural NetworkGraphTime SeriesFinance Related
🎯 What it does: This study proposes the TAMI framework, aimed at addressing the negative impact of node interaction heterogeneity in continuous time graphs on link prediction.
Taming Adversarial Constraints in CMDPs
Francesco Emanuele Stradi (Politecnico di Milano), Nicola Gatti (Politecnico di Milano)
OptimizationReinforcement Learning
🎯 What it does: This paper studies constrained Markov decision processes (CMDPs) under adversarial rewards and constraints, proposing new algorithms to mitigate known negative outcomes, allowing for lower regret and constraint violations in adversarial environments.
Taming generative video models for zero-shot optical flow extraction
Seungwoo Kim (Stanford University), Daniel LK Yamins
Object TrackingGenerationData SynthesisGenerative Adversarial NetworkOptical FlowVideo
🎯 What it does: A zero-shot optical flow extraction method called KL-tracing has been developed, which utilizes a large-scale self-supervised video generation model (LRAS) to extract optical flow from the generative distribution without fine-tuning;
Taming Hyperparameter Sensitivity in Data Attribution: Practical Selection Without Costly Retraining
Weiyi Wang (University of Michigan), Jiaqi W. Ma (University of Illinois Urbana-Champaign)
Hyperparameter SearchData-Centric LearningConvolutional Neural NetworkRecurrent Neural NetworkImageText
🎯 What it does: This paper evaluates the hyperparameter sensitivity of various data attribution methods (such as influence functions, TRAK, TracIn, LoGra, etc.) through large-scale experiments, and based on this, proposes a theoretical method for selecting the regularization parameter λ for non-retraining.
TANDEM: Bi-Level Data Mixture Optimization with Twin Networks
Jiaxing Wang (JD.com), Qixia Jiang (JD.com)
OptimizationSupervised Fine-TuningMultimodality
🎯 What it does: The TANDEM framework is proposed, utilizing dual network dynamic synchronization to solve the dual-layer optimization problem of data mixing ratios.
Tapered Off-Policy REINFORCE - Stable and efficient reinforcement learning for large language models
Nicolas Le Roux (Mila), Sam Work (Reliant AI)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A new offline reinforcement learning algorithm TOPR is proposed to fine-tune large language models and improve reasoning and search performance.
TAPIP3D: Tracking Any Point in Persistent 3D Geometry
Bowei Zhang (Peking University), Katerina Fragkiadaki (Carnegie Mellon University)
Object TrackingDepth EstimationTransformerSimultaneous Localization and MappingVideoPoint Cloud
🎯 What it does: This paper presents TAPIP3D, a method for long-term 3D point tracking using RGB or RGB-D video.
TARFVAE: Efficient One-Step Generative Time Series Forecasting via TARFLOW based VAE
Jiawen Wei (Meituan), Guangrui Ma (Meituan)
GenerationComputational EfficiencyTransformerFlow-based ModelAuto EncoderTime SeriesSequential
🎯 What it does: A one-shot generative time series forecasting framework named TARFVAE is proposed, which enhances posterior estimation capabilities by combining Transformer-based Autoregressive Flow (TARFLOW) with VAE, achieving high-quality deterministic and uncertainty predictions while maintaining one-shot inference speed.
Target Speaker Extraction through Comparing Noisy Positive and Negative Audio Enrollments
Shitong Xu (University of Oxford), Andrew Markham (University of Oxford)
RecognitionKnowledge DistillationConvolutional Neural NetworkAudio
🎯 What it does: This paper proposes a method for target speaker voice extraction under noisy enrollment (both positive and negative segments);
Targeted Maximum Likelihood Learning: An Optimization Perspective
Diyang Li (Cornell University), Kyra Gan (Cornell University)
Optimization
🎯 What it does: This paper systematically studies the iterative process of Targeted Maximum Likelihood Estimation (TMLE) from the perspective of optimization theory and provides a proof of its global convergence.
Task-Optimized Convolutional Recurrent Networks Align with Tactile Processing in the Rodent Brain
Trinity Chung (Carnegie Mellon University), Aran Nayebi (Carnegie Mellon University)
Convolutional Neural NetworkRecurrent Neural NetworkTransformerAuto EncoderContrastive LearningTime SeriesSequentialPhysics Related
🎯 What it does: Modeling the task optimization of tactile perception in mice using time neural networks, exploring the performance of convolutional recurrent networks and self-supervised learning in matching cortical neural responses.
Task-Specific Data Selection for Instruction Tuning via Monosemantic Neuronal Activations
Da Ma (Shanghai Jiao Tong University), Lu Chen (Shanghai Jiao Tong University)
TransformerSupervised Fine-TuningAuto EncoderText
🎯 What it does: A method based on sparse autoencoders called Monosemantic Neural Activation (MONA) is proposed for task-specific data selection in instruction tuning.
Taught Well Learned Ill: Towards Distillation-conditional Backdoor Attack
Yukun Chen (Zhejiang University), Kui Ren (Zhejiang University)
OptimizationKnowledge DistillationAdversarial AttackConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a conditional backdoor attack method for knowledge distillation, which activates the backdoor of the teacher model in the student model.
Taxonomy of reduction matrices for Graph Coarsening
Antonin Joly (National Centre for Scientific Research), Aline Roumy (Inria)
OptimizationGraph Neural NetworkGraph
🎯 What it does: This paper studies the asymmetry of projection matrices (reduction matrices) and lifting matrices in graph coarsening, and proposes to improve the projection matrix while keeping the lifting matrix unchanged to reduce the Restricted Spectral Approximation (RSA) error, thereby enhancing the performance of Graph Neural Networks (GNNs) on sparse graphs.
TC-Light: Temporally Coherent Generative Rendering for Realistic World Transfer
Yang Liu (Chinese Academy of Sciences), Zhaoxiang Zhang (Chinese Academy of Sciences)
GenerationOptimizationComputational EfficiencyDiffusion modelOptical FlowVideoBenchmark
🎯 What it does: This paper proposes TC-Light, a two-stage post-optimization framework for temporally consistent and physically feasible relighting and retexturing of long videos in high dynamic scenes.
Teaching Language Models to Evolve with Users: Dynamic Profile Modeling for Personalized Alignment
Weixiang Zhao (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposes the RLPA framework, which utilizes reinforcement learning to dynamically construct and update user profiles in multi-turn dialogues, achieving personalized alignment.
Teaching Language Models to Reason with Tools
Chengpeng Li (University of Science and Technology of China), Dayiheng Liu
AI Code AssistantReinforcement Learning from Human FeedbackLarge Language ModelSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: This paper proposes the CoRT framework, which efficiently utilizes a code interpreter in large reasoning models (LRM) for mathematical reasoning, and achieves the collaboration of code and internal reasoning through prompt engineering, rejection fine-tuning, and reinforcement learning.
Teaching Transformers to Solve Combinatorial Problems through Efficient Trial & Error
Panagiotis Giannoulis (National Technical University of Athens), Christos Tzamos (National and Kapodistrian University of Athens)
OptimizationTransformerReinforcement LearningSequential
🎯 What it does: By allowing the Transformer to perform trial-and-error searches in solving NP-class combinatorial problems like Sudoku, it learns rule application, guessing, and backtracking, achieving nearly perfect problem-solving capabilities.
Technical Debt in In-Context Learning: Diminishing Efficiency in Long Context
Taejong Joo (Northwestern University), Diego Klabjan (Northwestern University)
Computational EfficiencyMeta LearningTransformerLarge Language ModelText
🎯 What it does: This paper proposes an evaluation method based on the Meta ICL framework, systematically comparing the non-parametric learning capabilities of large language models with the sample complexity of Bayesian optimal and other benchmark algorithms. It finds that although the performance is close to optimal with few samples, the efficiency of ICL significantly decreases as the number of examples increases.
Temperature is All You Need for Generalization in Langevin Dynamics and other Markov Processes
Itamar Harel (Technion), Daniel Soudry (Technion)
Stochastic Differential Equation
🎯 What it does: This paper provides a time-independent bound on the generalization error for any data-related Markov process (such as Langevin dynamics), proving that the difference between training and testing errors can be bounded by \(\sqrt{\beta\mathbb{E}_{\theta_0}L(\theta)+\ln(1/\delta)}/N\).
Template-Guided 3D Molecular Pose Generation via Flow Matching and Differentiable Optimization
Noémie Bergues (Iktos SA), Hamza Tajmouati (Iktos SA)
GenerationOptimizationDrug DiscoveryGraph Neural NetworkFlow-based ModelGraphBenchmark
🎯 What it does: A 3D molecular pose generation method based on template molecules is proposed. First, an initial conformation aligned with the template is generated using the Flow-Matching model, and then the pose is finely adjusted through differentiable coordinate optimization (shape, ligand functional group, protein pocket matching, and internal energy).
TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles
Yaoyao Xu (Chinese University of Hong Kong), Mingchen Chen (Changping Laboratory)
GenerationProtein Structure PredictionRecurrent Neural NetworkTime SeriesSequentialBiomedical DataStochastic Differential Equation
🎯 What it does: The TEMPO framework is proposed, utilizing multi-scale autoregressive SDEs to generate protein conformation trajectories that comply with physical laws.
Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames
Anurag Arnab (Google DeepMind), Cordelia Schmid (Google DeepMind)
RecognitionRetrievalTransformerPrompt EngineeringVision Language ModelVideo
🎯 What it does: Proposes the Temporal Chain of Thought (TCoT) reasoning strategy, which allows the VLM to select relevant video frames to answer questions.
Temporal In‑Context Fine‑Tuning for Versatile Control of Video Diffusion Models
Kinam Kim (Korea Advanced Institute of Science and Technology), Jaegul Choo (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisSupervised Fine-TuningDiffusion modelVideo
🎯 What it does: This paper proposes the Temporal In-Context Fine-Tuning (TIC-FT) method, which utilizes temporal axis stitching and noise buffer frames to efficiently fine-tune pre-trained video diffusion models for conditional generation with a small number of samples.
Temporal Logic-Based Multi-Vehicle Backdoor Attacks against Offline RL Agents in End-to-end Autonomous Driving
Xuan Chen (Purdue University), Xiangyu Zhang (Purdue University)
Autonomous DrivingAdversarial AttackReinforcement LearningSequential
🎯 What it does: This paper proposes a backdoor attack using multi-vehicle trajectory triggers, targeting end-to-end offline reinforcement learning driving models. During the training phase, realistic triggering trajectories are automatically generated by imposing temporal logic (TL) constraints on the behavior of the attacking vehicle, and the attack's stealthiness is enhanced through negative sample training.
Temporal Representation Alignment: Successor Features Enable Emergent Compositionality in Robot Instruction Following
Vivek Myers (University of California), Sergey Levine (University of California)
Robotic IntelligenceReinforcement LearningContrastive LearningTextMultimodality
🎯 What it does: This paper proposes a method to enhance the zero-shot generalization ability of robots in new composite tasks (multi-step, language, or image targets) by learning a unified representation of tasks and states through time-contrastive loss.
Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting
Hyeongmin Lee (Twelve Labs), Kyungjune Baek (Sejong University)
CompressionOptimizationGaussian SplattingVideo
🎯 What it does: An end-to-end rate-distortion (RD) optimization compression framework oriented towards 4D Gaussian expansion (4DGS) is proposed, enabling dynamic scene rendering with flexible control over compression quality and bit rate.
Temporal-Difference Variational Continual Learning
Luckeciano Carvalho Melo, Yarin Gal (University of Oxford)
Reinforcement LearningImage
🎯 What it does: A new variational continual learning objective is proposed - Temporal-Difference VCL (TD-VCL), which introduces multi-step (n-step) KL regularization and the TD(λ) mechanism into the learning objective, utilizing past multiple posterior estimates to alleviate catastrophic forgetting.
TempSamp-R1: Effective Temporal Sampling with Reinforcement Fine-Tuning for Video LLMs
Yunheng Li (Nankai University), Ming-Ming Cheng (Nankai University)
Large Language ModelReinforcement LearningVideoMultimodalityChain-of-Thought
🎯 What it does: The TempSamp-R1 framework is proposed, which enhances the performance of multimodal large models in video temporal localization tasks through reinforcement learning combined with offline supervision.
Tensor Decomposition Networks for Accelerating Machine Learning Force Field Computations
Yuchao Lin (Lambda, Inc.), Shuiwang Ji (Texas A&M University)
Computational EfficiencyDrug DiscoveryTransformerTabular
🎯 What it does: The paper proposes Tensor Decomposition Networks (TDN), which accelerates molecular potential energy calculations by replacing SO(3) equivariant tensor products with low-rank CANDECOMP/PARAFAC (CP) decomposition.
Tensor Product Attention Is All You Need
Yifan Zhang (Princeton University), Andrew C Yao
TransformerText
🎯 What it does: Proposes the Tensor Product Attention (TPA) mechanism, which significantly compresses the KV cache by performing context-aware low-rank tensor decomposition on queries, keys, and values; builds the T6 Transformer based on TPA and implements the FlashTPA decoder;
Tensor-Parallelism with Partially Synchronized Activations
Itay Lamprecht (Intel), Daniel Soudry (Amazon)
TransformerLarge Language ModelText
🎯 What it does: This paper studies the feasibility of using partially synchronized activations in tensor parallel training, proposing the CAAT-Net architecture, which significantly reduces communication overhead while maintaining the pre-training accuracy of large language models.
TensorRL-QAS: Reinforcement learning with tensor networks for improved quantum architecture search
Akash Kundu (University of Helsinki), Stefano Mangini (University of Helsinki)
OptimizationNeural Architecture SearchReinforcement LearningTabularPhysics Related
🎯 What it does: Combining the ground state (MPS) pre-trained with tensor networks and reinforcement learning to automatically search for and optimize quantum circuit architectures suitable for NISQ hardware.
Test Time Scaling for Neural Processes
Hyungi Lee (Kookmin University), Juho Lee (KAIST)
OptimizationComputational EfficiencyImage
🎯 What it does: This paper proposes a method called TTSNP that uses Sequential Monte Carlo Sampler (SMCS) for adaptive stretching during the testing of Neural Process (NP) models. It calibrates the global latent variable posterior without modifying the pre-trained model, improving prediction accuracy and uncertainty estimation.
Test-Time Adaptation by Causal Trimming
Yingnan Liu (National University of Singapore), Wynne Hsu (National University of Singapore)
Domain AdaptationImage
🎯 What it does: A gradient-free test-time adaptation method called TACT is proposed, which enhances the model's robustness under distribution shifts by trimming non-causal features.
Test-Time Adaptation of Vision-Language Models for Open-Vocabulary Semantic Segmentation
Mehrdad Noori (École de Technologie Supérieure), Christian Desrosiers (École de Technologie Supérieure)
SegmentationDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: This paper proposes a testing-time adaptation framework called MLMP for open-vocabulary semantic segmentation (OVSS), which dynamically adjusts the model during inference to cope with domain shifts.
Test-Time Adaptive Object Detection with Foundation Model
Yingjie Gao (Beihang University), Di Huang (Beihang University)
Object DetectionDomain AdaptationTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: The paper proposes a test-time adaptive object detection method based on a visual-language foundation model (Grounding DINO), which can adapt in real-time without accessing source data, across domains, and even across categories.
Test-Time Scaling of Diffusion Models via Noise Trajectory Search
Vignav Ramesh (Harvard University), Morteza Mardani (NVIDIA)
GenerationOptimizationReinforcement LearningDiffusion modelImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes optimizing the noise trajectory during inference of diffusion models by balancing global exploration and local exploitation at each step to improve generation quality.
Test-Time Spectrum-Aware Latent Steering for Zero-Shot Generalization in Vision-Language Models
Konstantinos M. Dafnis (Rutgers University), Dimitris N. Metaxas (Rutgers University)
Domain AdaptationOptimizationComputational EfficiencyTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A lightweight framework called STS is proposed for adaptive testing on visual-language models (such as CLIP). This framework learns a small coefficient vector in a low-dimensional spectral subspace obtained from the singular value decomposition (SVD) of text prototypes, which offsets the original text prototypes, thereby enhancing the robustness of zero-shot inference without modifying the frozen encoder.
Test3R: Learning to Reconstruct 3D at Test Time
Yuheng Yuan (National University of Singapore), Xinchao Wang (National University of Singapore)
Depth EstimationImage
🎯 What it does: A Test3R method is proposed for learning 3D reconstruction by maximizing the point cloud consistency of different image pairs during the testing phase.
Text to Sketch Generation with Multi-Styles
Tengjie Li (Shanghai Jiao Tong University), Lei Xu (Guangdong Laboratory of Artificial Intelligence and Digital Economy)
GenerationDiffusion modelImage
🎯 What it does: A training-free multi-style sketch generation framework M3S is proposed, utilizing a diffusion model to achieve zero-shot style control by combining reference style sketches and text prompts.
Text-Aware Real-World Image Super-Resolution via Diffusion Model with Joint Segmentation Decoders
Qiming Hu (Tianjin University), Qingnan Fan (vivo Mobile Communication Co. Ltd)
RestorationSegmentationSuper ResolutionDiffusion modelImageText
🎯 What it does: A text-aware diffusion model TADiSR is proposed for image super-resolution in real scenarios while simultaneously performing text segmentation.
Text-to-Code Generation for Modular Building Layouts in Building Information Modeling
Yinyi WEI, Xiao LI
GenerationData SynthesisAI Code AssistantLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A framework named Text2MBL is proposed for generating BIM code from text, which can directly convert natural language descriptions of modular building layouts into executable BIM code, automatically constructing the hierarchical structure of modules, units, and rooms.
Text-to-Decision Agent: Offline Meta-Reinforcement Learning from Natural Language Supervision
Shilin Zhang (Nanjing University), Zhi Wang (Nanjing University)
Robotic IntelligenceMeta LearningTransformerReinforcement LearningDiffusion modelContrastive LearningWorld ModelText
🎯 What it does: This study explores how to utilize natural language supervision to train a general agent for offline meta reinforcement learning, achieving zero-shot text-to-decision generation.
TF-MAS: Training-free Mamba2 Architecture Search
Yi Fan (Nanjing University), Yu-Bin Yang (Nanjing University)
Neural Architecture SearchText
🎯 What it does: A training-free NAS method TF-MAS is proposed for efficient search of Mamba2 network architecture.
TGA: True-to-Geometry Avatar Dynamic Reconstruction
Bo Guo (Beihang University), Yifan Zhao (Beihang University)
GenerationPose EstimationGaussian SplattingMesh
🎯 What it does: This paper proposes TGA, a 4D high-precision portrait reconstruction framework aimed at pose awareness, which captures subtle expression changes through adaptive Gaussian transformation and homogeneous projection, and achieves rapid mesh extraction.
THD-BAR: Topology Hierarchical Derived Brain Autoregressive Modeling for EEG Generic Representations
Wenchao Yang (Beihang University), Yang Li (Beihang University)
RecognitionRepresentation LearningTransformerSupervised Fine-TuningTime SeriesBiomedical DataElectrocardiogram
🎯 What it does: A brain topology hierarchical autoregressive model THD-BAR is proposed for learning general EEG representations.
The $\varphi$ Curve: The Shape of Generalization through the Lens of Norm-based Capacity Control
Yichen Wang (University of Wisconsin Madison), Fanghui Liu (University of Warwick)
Image
🎯 What it does: This study investigates the impact of capacity measures based on parameter norms on the learning curves of Random Feature Models (RFM) and linear models, deriving a deterministic equivalence relationship between the norm and test error, and providing an analysis of the double descent phenomenon in the norm scale.
The Adaptive Complexity of Minimizing Relative Fisher Information
Huanjian Zhou (University of Tokyo), Masashi Sugiyama (University of Tokyo)
OptimizationStochastic Differential Equation
🎯 What it does: A parallel Picard iteration algorithm is proposed and analyzed, reducing the adaptive complexity from O(d/ε⁴) to O(d/ε²) in non-logarithmic convex sampling, and a corresponding lower bound is provided;
The Atlas of In-Context Learning: How Attention Heads Shape In-Context Retrieval Augmentation
Patrick Kahardipraja (Fraunhofer Heinrich Hertz Institute), Sebastian Lapuschkin (Fraunhofer Heinrich Hertz Institute)
RetrievalTransformerTextRetrieval-Augmented Generation
🎯 What it does: This study investigates the role of attention heads in retrieval-augmented language models, distinguishing and locating in-context heads responsible for instruction understanding and information retrieval in in-context learning (ICL) from parametric heads that store relational knowledge. It demonstrates their causal impact on answer generation through functional vectors and weight modifications, and further develops a linear detector to track the sources of retrieved information.
The Best Instruction-Tuning Data are Those That Fit
Dylan Zhang (University of Illinois Urbana-Champaign), Hao Peng (University of Illinois Urbana-Champaign)
Data-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A response selection method based on target model pre-training distribution alignment, called GRAPE, is proposed to customize supervised fine-tuning (SFT) data.
The Bias-Variance Tradeoff in Data-Driven Optimization: A Local Misspecification Perspective
Haixiang Lan (Columbia University), Haofeng Zhang (Morgan Stanley)
Optimization
🎯 What it does: The study compares the bias, variance, and scheduling performance of SAA, ETO, and IEO in data-driven optimization under local model error conditions.
The Boundaries of Fair AI in Medical Image Prognosis: A Causal Perspective
Thai-Hoang Pham (Ohio State University), Ping Zhang (Ohio State University)
ClassificationSegmentationAnomaly DetectionConvolutional Neural NetworkImageBiomedical DataMagnetic Resonance ImagingAlzheimer's Disease
🎯 What it does: This paper proposes the FairTTE framework for assessing and improving the fairness of time-to-event predictions in medical imaging.
The Burden of Interactive Alignment with Inconsistent Preferences
Ali Shirali (University of California Berkeley)
Recommendation SystemOptimizationReinforcement Learning
🎯 What it does: The study investigates how users can align algorithms with their true interests through strategies in the context of inconsistent preferences during multi-step interactions.
The Complexity of Correlated Equilibria in Generalized Games
Martino Bernasconi (Bocconi University), Gabriele Farina (Massachusetts Institute of Technology)
Optimization
🎯 What it does: This paper studies the computational complexity of calculating constrained Phi-equilibria (i.e., constrained correlated equilibria) in generalized games (where the strategy sets available to players are influenced by the actions of other players) and proves that their approximate computation is PPAD-complete.
The Complexity of Finding Local Optima in Contrastive Learning
Jingming Yan (University of California), Stelios Andrew Stavroulakis
OptimizationRepresentation LearningContrastive Learning
🎯 What it does: This paper studies the computational complexity of obtaining local optimal solutions for contrastive learning objectives (such as triplet maximization and triplet loss minimization) and proves that it is PLS-hard in discrete settings and CLS-hard in continuous settings, indicating that even finding local optima is, in the worst case, not polynomially solvable.
The Complexity of Symmetric Equilibria in Min-Max Optimization and Team Zero-Sum Games
Ioannis Anagnostides (Carnegie Mellon University), Jingming Yan (University of California Irvine)
Optimization
🎯 What it does: This paper conducts an in-depth study of the computational complexity of symmetric equilibrium problems in min-max optimization and team zero-sum games, proving that several related problems are complete or hard to solve in complexity classes such as CLS, PPAD, and FNP.
The Computational Advantage of Depth in Learning High-Dimensional Hierarchical Targets
Yatin Dandi (Ecole Polytechnique Federale de Lausanne), Florent Krzakala (Ecole Polytechnique Federale de Lausanne)
OptimizationComputational EfficiencyTabular
🎯 What it does: This paper characterizes the hierarchical feature learning process of deep networks by constructing single-index high-dimensional hierarchical objectives (SIGHT) and multi-index hierarchical objectives (MIGHT). It analyzes the learning dynamics of gradient descent training in the high-dimensional limit, proving that deep networks achieve lower sample complexity through layer-wise dimensionality reduction. It also provides a strict theorem for three-layer networks and proposes a general hierarchical information index (CIE) hypothesis.
The Computational Complexity of Counting Linear Regions in ReLU Neural Networks
Moritz Stargalla (University of Technology Nuremberg), Daniel Reichman (Worcester Polytechnic Institute)
🎯 What it does: A systematic analysis of the computational complexity of counting linear regions in ReLU networks is conducted, identifying six non-equivalent definitions, and proving that the counting problem is #P-hard and NP-hard in both shallow and deep networks, with polynomial space algorithms available for some definitions.
The Cost of Compression: Tight Quadratic Black-Box Attacks on Sketches for $\ell_2$ Norm Estimation
Sara Ahmadian (Google Research), Uri Stemmer (Tel Aviv University)
OptimizationAdversarial AttackTabular
🎯 What it does: This paper studies the robustness of ℓ₂ norm estimation under linear compression (sketch) against black-box attacks, proposing a general non-adaptive attack method that can render any linear compression matrix or estimator ineffective within O(k²) queries.
The Cost of Robustness: Tighter Bounds on Parameter Complexity for Robust Memorization in ReLU Nets
Yujun Kim (Korea Advanced Institute of Science and Technology), Chulhee Yun (Korea Advanced Institute of Science and Technology)
🎯 What it does: This paper studies the parameter complexity required for robust memory in ReLU networks under different robustness ratios ρ, providing upper and lower bounds.
The Curse of Depth in Large Language Models
Wenfang Sun (Westlake University), Shiwei Liu (University of Oxford)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper analyzes the deep inefficacy of the 'deep curse' in large language models and proposes a layer-based LayerNorm scaling method to alleviate this issue.
The Dual Nature of Plasticity Loss in Deep Continual Learning: Dissection and Mitigation
Haoyu Wang, Yuguo Yu (Fudan University)
ClassificationConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: This study investigates the mechanism of loss of plasticity (LoP) in neural networks during deep continual learning and proposes a universal Mixup method to alleviate two types of LoP.
The Effect of Optimal Self-Distillation in Noisy Gaussian Mixture Model
Kaito Takanami (University of Tokyo), Ayaka Sakata (Ochanomizu University)
ClassificationKnowledge DistillationImage
🎯 What it does: The study investigates the behavior and mechanisms of multi-stage self-distillation under a noisy Gaussian mixture model with optimal hyperparameters.
The Emergence of Abstract Thought in Large Language Models Beyond Any Language
Yuxin Chen (National University of Singapore), Wenxuan Zhang (Singapore University of Technology and Design)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper analyzes the neuron activations of large language models to identify and distinguish language-related shared and exclusive neurons, studying their evolution with model iterations, and proposes a neuron-directed training method based on language independence scoring to enhance multilingual capabilities.
The emergence of sparse attention: impact of data distribution and benefits of repetition
Nicolas Zucchet (ETH Zurich), Stephanie C.Y. Chan
TransformerLarge Language ModelSequential
🎯 What it does: This paper studies the emergence of sparse attention in large language models and explores the impact of data distribution and repetition on the timing of its emergence.
The First Few Tokens Are All You Need: An Efficient and Effective Unsupervised Prefix Fine-Tuning Method for Reasoning Models
Ke Ji (Chinese University of Hong Kong), Dong Yu (Tencent)
TransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper proposes an Unsupervised Prefix Fine-Tuning (UPFT) method that enhances reasoning capabilities by utilizing the Prefix Self-Consistency shared by large language models when generating different answers, requiring only the first few words of each question (e.g., 8 tokens) for fine-tuning.
The Flood Complex: Large-Scale Persistent Homology on Millions of Points
Florian Graf (University of Salzburg), Roland Kwitt (University of Applied Sciences)
ClassificationComputational EfficiencyPoint CloudBenchmark
🎯 What it does: This paper proposes Flood complex, a filtering-style simplicial complex constructed on subsampling points, which can efficiently compute persistent homology (PH) on millions of point clouds.
The Fluorescent Veil: A Stealthy and Effective Physical Adversarial Patch Against Traffic Sign Recognition
Shuai Yuan (University of Electronic Science and Technology of China), Yuguang Fang (City University of Hong Kong)
Object DetectionAutonomous DrivingAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A fluorescent printed adversarial patch (FIPatch) that can be activated by UV light was designed and implemented on a traffic sign recognition system, achieving misclassification or non-detection attacks on target vehicles in the physical world.
The Fragile Truth of Saliency: Improving LLM Input Attribution via Attention Bias Optimization
Yihua Zhang (Michigan State University), Sijia Liu (Michigan State University)
OptimizationExplainability and InterpretabilityTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A new input saliency method called ABO (Attention Bias Optimization) is proposed, along with a long text saliency assessment framework based on needle-in-a-haystack (NIAH).
The Future Unmarked: Watermark Removal in AI-Generated Images via Next-Frame Prediction
Huming Qiu (Fudan University), Min Yang (Fudan University)
RestorationGenerationAdversarial AttackDiffusion modelOptical FlowImageVideo
🎯 What it does: The first semantic-level image watermark removal attack is proposed - Next Frame Prediction Attack (NFPA), which transforms the watermark removal task into the next frame prediction of video generation, achieving efficient removal of eight existing SOTA watermark schemes.
The Gaussian Mixing Mechanism: Renyi Differential Privacy via Gaussian Sketches
Omri Lev (Massachusetts Institute of Technology), Ashia C. Wilson (Massachusetts Institute of Technology)
OptimizationSafty and PrivacyConvolutional Neural NetworkGaussian SplattingTabular
🎯 What it does: This paper studies the Gaussian Mixing Mechanism (GaussMix), conducts a fine privacy analysis using Rényi Differential Privacy (RDP), and applies this mechanism to differential privacy linear regression and logistic regression, proposing improved algorithms and demonstrating theoretical and experimental superiority.
The Generative Leap: Tight Sample Complexity for Efficiently Learning Gaussian Multi-Index Models
Alex Damian (Princeton University), Joan Bruna (New York University)
Gaussian Splatting
🎯 What it does: This study considers the Gaussian multi-index model and proposes an effective unbiased estimation procedure to learn the hidden low-dimensional subspace, introducing the generative transition index as a natural extension of the multi-index setting.
The Good, the Bad and the Ugly: Meta-Analysis of Watermarks, Transferable Attacks and Adversarial Defenses
Grzegorz Gluch, Sebastian Pokutta (Zuse Institute Berlin)
Adversarial Attack
🎯 What it does: Formal definitions of three mechanisms: backdoor watermarking, adversarial defense, and a new transferable attack are provided, and their game-theoretic relationships are explained through an interactive protocol, proving that at least one of these three mechanisms exists in any learning task.
The Graphon Limit Hypothesis: Understanding Neural Network Pruning via Infinite Width Analysis
Hoang Pham (University of Warwick), Long Tran-Thanh (University of Warwick)
Graph Neural NetworkGraphStochastic Differential Equation
🎯 What it does: This paper proposes viewing the sparse sub-networks generated by network pruning as graph limits, constructing the corresponding graph convolution function (graphon), and deriving the Graphon Neural Tangent Kernel (Graphon NTK) in the infinite width limit to analyze the training dynamics and spectral characteristics of different pruning methods.
The Hawthorne Effect in Reasoning Models: Evaluating and Steering Test Awareness
Sahar Abdelnabi (Microsoft), Ahmed Salem (Microsoft)
Safty and PrivacyTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates and quantifies the impact of 'test awareness' on reasoning behavior and safety in large language models during evaluation, and proposes a control method based on white-box detection and parameter editing.
The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
Parshin Shojaee (Apple), Mehrdad Farajtabar (Apple)
Large Language ModelReinforcement LearningChain-of-Thought
🎯 What it does: Systematically evaluate the reasoning ability of large-scale reasoning models (LRM) in controllable puzzle environments and conduct a detailed analysis of their thought processes.
The Implicit Bias of Structured State Space Models Can Be Poisoned With Clean Labels
Yonatan Slutzky (Tel Aviv University), Nadav Cohen (Tel Aviv University)
OptimizationAdversarial AttackRecurrent Neural NetworkSequential
🎯 What it does: This paper studies the implicit bias of Structured State Space Models (SSM) and proves that when the training data is limited and contains specific clean label samples, this bias can be disrupted, leading to generalization failure. It also formally proves for the first time that SSM is vulnerable to clean label poisoning attacks, and provides a dynamical characterization based on gradient flow and a non-resonant linearization theory.
The Indra Representation Hypothesis
Jianglin Lu (Northeastern University), Yun Fu (Northeastern University)
RetrievalRepresentation LearningTransformerContrastive LearningImageTextMultimodalityAudio
🎯 What it does: Proposed the 'Indra Representation' - a relational representation method based on category theory, which constructs a global relationship vector for each sample using the relative distances of features generated by the model, and validates its effectiveness in unimodal, vision-language, and speech-language cross-modal tasks.