NeurIPS 2025 Papers — Page 49
Conference on Neural Information Processing Systems · 5275 papers
Training-Free Bayesianization for Low-Rank Adapters of Large Language Models
Haizhou Shi (Rutgers University), Hao Wang (Rutgers University)
TransformerLarge Language ModelText
🎯 What it does: A training-independent Bayesian framework TFB is proposed, which transforms the pre-trained low-rank adapter LoRA into a Bayesian model, allowing for uncertainty estimation without further training.
Training-Free Constrained Generation With Stable Diffusion Models
Stefano Zampini (Polytechnic of Turin), Ferdinando Fioretto (University of Virginia)
GenerationOptimizationDiffusion modelImage
🎯 What it does: A training-free, constraint generation method based on robust diffusion models is proposed, achieving real-time satisfaction of strict constraints such as physical, functional, or copyright constraints through the use of proximal mapping and gradient projection in the latent space.
Training-free Detection of AI-generated images via Cropping Robustness
Sungik Choi (LG AI Research), Moontae Lee (Sungkyunkwan University)
Anomaly DetectionTransformerContrastive LearningImage
🎯 What it does: A training-free AI-generated image detection method called WaRPAD is proposed, which utilizes the robustness of self-supervised models to random cropping (RandomResizedCrop) and distinguishes between real and synthetic images through high-frequency wavelet perturbation sensitivity.
Training-Free Efficient Video Generation via Dynamic Token Carving
Yuechen Zhang (Chinese University of Hong Kong), Jiaya Jia (Hong Kong University of Science and Technology)
GenerationData SynthesisComputational EfficiencyVideo
🎯 What it does: A training-agnostic Jenga reasoning pipeline is proposed, achieving efficient video generation through dynamic attention pruning and stage-wise resolution generation;
Training-Free Guidance Beyond Differentiability: Scalable Path Steering with Tree Search in Diffusion and Flow Models
Yingqing Guo (Princeton University), Mengdi Wang (Princeton University)
GenerationData SynthesisDiffusion modelFlow-based ModelSequentialAudio
🎯 What it does: The TreeG framework is proposed, achieving untrained guided generation through tree search, suitable for both continuous and discrete diffusion and flow models, and addressing non-differentiable objectives.
Training-free Online Video Step Grounding
Luca Zanella (University of Trento), Elisa Ricci (Fondazione Bruno Kessler)
Large Language ModelVideoTextMultimodality
🎯 What it does: This paper proposes a training-free online video step localization method called BAGLM, which utilizes Bayesian filtering to integrate predictions from multimodal models and language models.
Training-Free Safe Denoisers for Safe Use of Diffusion Models
Mingyu Kim (University of British Columbia), Mijung Park (University of British Columbia)
GenerationSafty and PrivacyDiffusion modelImage
🎯 What it does: A training-free safe denoiser (Safe Denoiser) is proposed, which automatically removes unsafe content (such as NSFW, copyrighted, or private data) by adding a penalty term based on unsafe samples to the sampling trajectory of the diffusion model.
Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models
Byeonghu Na (Korea Advanced Institute of Science and Technology), Il-chul Moon
GenerationSafty and PrivacyDiffusion modelImageText
🎯 What it does: A training-free safe text-to-image diffusion model method is proposed—Safe Text Embedding Guidance (STG), which achieves safe output by dynamically adjusting text embeddings during the sampling process.
Training-Free Test-Time Adaptation via Shape and Style Guidance for Vision-Language Models
Shenglong Zhou (Hikvision Research Institute), Jiang Zhu (Hikvision Research Institute)
ClassificationDomain AdaptationTransformerVision Language ModelContrastive LearningImage
🎯 What it does: A training-free adaptive method called SSG is proposed, which utilizes shape-sensitive and style-insensitive factors to enhance the zero-shot classification performance of CLIP in out-of-domain and cross-domain scenarios.
TrajAgent: An LLM-Agent Framework for Trajectory Modeling via Large-and-Small Model Collaboration
Yuwei Du (Tsinghua University), Yong Li (Tsinghua University)
Anomaly DetectionOptimizationTransformerLarge Language ModelAgentic AIPrompt EngineeringTime SeriesSequential
🎯 What it does: This paper presents TrajAgent, an agent framework based on large language models that can automate the entire process of trajectory modeling (task identification, planning, execution, and summarization), supporting various trajectory tasks and diverse data.
Trajectory Balance with Asynchrony: Decoupling Exploration and Learning for Fast, Scalable LLM Post-Training
Brian R. Bartoldson (Lawrence Livermore National Laboratory), Bhavya Kailkhura (Lawrence Livermore National Laboratory)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes Trajectory Balance with Asynchrony (TBA), a post-training framework that combines the offline Trajectory Balance objective with distributed asynchronous search, decoupling exploration from learning.
Trajectory Bellman Residual Minimization: A Simple Value-Based Method for LLM Reasoning
Yurun Yuan (University of Wisconsin-Madison), Tengyang Xie (University of Wisconsin-Madison)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Proposed and implemented the Trajectory Bellman Residual Minimization (TBRM) algorithm, using the logits of the LLM itself as Q-values, with single-trajectory offline training, eliminating the need for complex components such as critics and importance sampling;
Trajectory Graph Learning: Aligning with Long Trajectories in Reinforcement Learning Without Reward Design
Yunfan Li (University of California), Lin Yang
Graph Neural NetworkReinforcement LearningSequential
🎯 What it does: A framework called Trajectory Graph Learning (TGL) is proposed to directly align expert trajectories, achieving precise replication of long-term behaviors by constructing a trajectory conflict graph.
TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model
Yichen Liu (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
CompressionAutonomous DrivingKnowledge DistillationRepresentation LearningContrastive LearningTime Series
🎯 What it does: This paper proposes an efficient and semantically rich vehicle trajectory pre-training model called TrajMamba, which can simultaneously capture motion patterns and learn travel semantics from both GPS and road perspectives.
Transcending Cost-Quality Tradeoff in Agent Serving via Session-Awareness
Yanyu Ren (Tsinghua University), Yu Bai (Zhongguancun Laboratory)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: Designed and implemented AGSERVE, a conversation-aware server for LLM Agents, addressing the trade-off between cost and quality.
Transductive Conformal Inference for Full Ranking
Jean-Baptiste Fermanian (University of Montpellier), Gilles Blanchard (University of Paris Saclay)
Recommendation SystemOptimizationTabular
🎯 What it does: A conduction-based non-distributional hypothesis confidence interval method is proposed to quantify the errors in fully ranking m new items among n existing items with relative rankings.
Transfer Faster, Price Smarter: Minimax Dynamic Pricing under Cross-Market Preference Shift
Yi Zhang (Columbia University), Yujun Yan (Dartmouth College)
Recommendation SystemOptimizationMeta LearningTabular
🎯 What it does: A cross-market transfer dynamic pricing framework CM-TDP is proposed, compatible with offline to online and online to online transfers, applicable to linear and RKHS nonlinear utility models.
Transfer Learning for Benign Overfitting in High-Dimensional Linear Regression
Yeichan Kim, Seyoung Park
OptimizationTabular
🎯 What it does: Proposes a two-step Transfer MNI and its weighted integration, utilizing the minimum L2 norm interpolation of the source task to enhance the Bayesian overfitting generalization performance of high-dimensional linear regression.
Transfer Learning on Edge Connecting Probability Estimation Under Graphon Model
Yuyao Wang (Boston University), Huimin Cheng (Boston University)
Domain AdaptationOptimizationGraph Neural NetworkGraph
🎯 What it does: In small-scale target graphs, structural information is extracted from large-scale source graphs through transfer learning to improve the accuracy of graph connectivity probability (graph node) estimation.
Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models
Tomas Soucek, Alexandre Mourachko (Meta)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: A black-box attack method has been developed that requires only a single watermark image for watermark forgery and removal.
Transferring Causal Effects using Proxies
Manuel Iglesias-Alonso (ETH Zürich), Jonas Peters (ETH Zürich)
🎯 What it does: The study estimates causal effects under unobserved confounding variables using observable proxy variables in a multi-domain setting.
Transferring Linear Features Across Language Models With Model Stitching
Alan Chen (Brown University), Ellie Pavlick (Brown University)
Large Language ModelAuto EncoderText
🎯 What it does: Proposes and validates a method for transferring Sparse Autoencoders (SAE), detectors, and guiding vectors between language models of different scales using linear mapping (model stitching);
TransferTraj: A Vehicle Trajectory Learning Model for Region and Task Transferability
Tonglong Wei (Beijing Jiaotong University), Huaiyu Wan (Beijing Jiaotong University)
GenerationAutonomous DrivingTransformerMixture of ExpertsTime Series
🎯 What it does: This paper proposes a vehicle trajectory learning model called TransferTraj, which can transfer across different regions and tasks, addressing the issue that traditional models require separate training for each region and task.
Transformer brain encoders explain human high-level visual responses
Hossein Adeli (Columbia University), Nikolaus Kriegeskorte (Columbia University)
TransformerImageMultimodalityMagnetic Resonance Imaging
🎯 What it does: A brain encoding model based on the Transformer attention mechanism is proposed, dynamically routing retinal spatial features to higher-order visual areas to predict fMRI signals during natural scene viewing.
Transformer Copilot: Learning from The Mistake Log in LLM Fine-tuning
Jiaru Zou (University of Illinois Urbana-Champaign), Jingrui He (University of Illinois Urbana-Champaign)
Recommendation SystemTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the Transformer Copilot framework, which records and utilizes the model's own error logs (Mistake Log) during the fine-tuning process of LLMs to enhance inference performance.
Transformer Key-Value Memories Are Nearly as Interpretable as Sparse Autoencoders
Mengyu Ye (Tohoku University), Tatsuki Kuribayashi (MBZUAI)
Explainability and InterpretabilityTransformerLarge Language ModelAuto EncoderText
🎯 What it does: This paper compares the performance of the feed-forward (FF) layer in Transformer as key-value memory (FF-KV) with that of sparse autoencoders (SAE) and other proxy modules like Transcoder in terms of interpretability, exploring whether FF-KV can serve as a strong baseline.
Transformers are almost optimal metalearners for linear classification
Roey Magen (Weizmann Institute of Science), Gal Vardi (Weizmann Institute of Science)
ClassificationMeta LearningTransformerTabular
🎯 What it does: This paper theoretically proves that the trained linear Transformer can achieve near-optimal classification of new tasks with only a small number of contextual samples, demonstrating excellent metalearning capabilities.
Transformers for Mixed-type Event Sequences
Felix Draxler (University of California), Stephan Mandt (Chan Zuckerberg Initiative)
TransformerFlow-based ModelTime SeriesBiomedical DataElectronic Health Records
🎯 What it does: A unified Transformer-based intensity-free point process (FLEXTPP) is proposed, capable of handling variable-length event sequences with both discrete and continuous labels, and achieving structured prediction through conditional input.
Transformers Learn Faster with Semantic Focus
Parikshit Ram (IBM Research), Alexander G. Gray
TransformerText
🎯 What it does: This study investigates the learning convergence and generalization performance of sparse attention (especially input-dependent high-frequency vertex attention) in Transformers, providing a theoretical analysis.
Transformers Provably Learn Chain-of-Thought Reasoning with Length Generalization
Yu Huang (University of Pennsylvania), Yuxin Chen (University of Pennsylvania)
OptimizationTransformerSequentialChain-of-Thought
🎯 What it does: It is proven that a one-layer Transformer without positional encoding can learn chain-of-thought reasoning through gradient descent and achieve length generalization under limited training duration; at the same time, a recursive self-training curriculum is introduced to extend the inferable length of the model in more challenging NC1 level tasks.
Transforming Gaps into Gains: Bridging Model and Data Heterogeneity in Federated Learning via Knowledge Weak-Aware Zones
Ke Li (Chongqing University of Posts and Telecommunications), Shenhai Zheng (Chongqing University of Posts and Telecommunications)
Federated LearningSafty and PrivacyImage
🎯 What it does: Proposes the FedKWAZ framework, which bridges model and data heterogeneity in heterogeneous federated learning through the KWAZ mechanism.
Transforming Generic Coder LLMs to Effective Binary Code Embedding Models for Similarity Detection
Litao Li (Queen's University), Philippe Charland (Defence Research and Development Canada)
RetrievalAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningContrastive LearningText
🎯 What it does: A multi-stage fine-tuning framework (data augmentation, translation-based autoregressive training, LLM2Vec, cumulative GTE loss) is proposed and implemented for binary code similarity retrieval, transforming a general LLM into an efficient binary embedding model.
Transition Matching: Scalable and Flexible Generative Modeling
Neta Shaul (Weizmann Institute of Science), Yaron Lipman (Meta)
GenerationData SynthesisDiffusion modelImageText
🎯 What it does: This paper proposes the Transition Matching (TM) framework, unifying diffusion, flow matching, and continuous autoregressive models into a discrete-time, continuous-state generative paradigm, and implements three variants (DTM, ARTM, FHTM).
TransMLA: Migrating GQA Models to MLA with Full DeepSeek Compatibility and Speedup
Fanxu Meng (Institute for Artificial Intelligence Peking University), Muhan Zhang (Institute for Artificial Intelligence Peking University)
CompressionComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: Transfer the existing GQA-based pre-trained large models (such as LLaMA, Qwen, etc.) to the MLA structure, and achieve significant inference acceleration by compressing the KV cache while maintaining or only slightly losing performance.
Transstratal Adversarial Attack: Compromising Multi-Layered Defenses in Text-to-Image Models
Chunlong Xie (Chongqing University), Tao Xiang (Chongqing University)
GenerationAdversarial AttackLarge Language ModelImageText
🎯 What it does: A black-box attack framework based on LLM-generated candidate words and genetic optimization is proposed, capable of simultaneously breaking through the multi-layer security defenses of text-to-image models, generating implicit NSFW prompts while evading image filters while maintaining subjective undesirability.
TranSUN: A Preemptive Paradigm to Eradicate Retransformation Bias Intrinsically from Regression Models in Recommender Systems
Jiahao Yu (Taobao and Tmall Group of Alibaba), Bo Zheng (Taobao and Tmall Group of Alibaba)
Recommendation SystemTabular
🎯 What it does: This paper proposes TranSUN and its generalized framework GTS, designing a regression model that preemptively removes transformation bias to achieve unbiased predictions in recommendation systems.
TRAP: Targeted Redirecting of Agentic Preferences
Hangoo Kang (University of Illinois Urbana-Champaign), Gagandeep Singh (University of Illinois Urbana-Champaign)
GenerationRetrievalAdversarial AttackDiffusion modelContrastive LearningImageTextMultimodality
🎯 What it does: A black-box adversarial method called TRAP is proposed, which injects semantic information into the CLIP latent space using diffusion models, thereby inducing multimodal proxy systems to prefer altered target images.
Traversal Verification for Speculative Tree Decoding
Yepeng Weng (Lenovo AI Technology Center), zhongchao shi
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: The Traversal Verification algorithm is proposed, improving the verification method in existing tree-structured speculative decoding.
Treasure Hunt: Real-time Targeting of the Long Tail using Training-Time Markers
Daniel D'souza (Cohere Labs), Sara Hooker (Adaption Labs)
GenerationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: Introduce 'Treasure markers' during training, labeling samples with multi-dimensional attributes, allowing the model to automatically infer and controllably output during inference, enhancing performance on long-tail tasks.
Treatment Effect Estimation for Optimal Decision-Making
Dennis Frauen (Ludwig Maximilian University of Munich), Stefan Feuerriegel (Ludwig Maximilian University of Munich)
OptimizationMeta LearningTabular
🎯 What it does: This study investigates the optimality of a two-stage CATE meta-learner under constrained function classes and proposes a method for redirecting CATE to enhance the performance of threshold decisions.
Tree Ensemble Explainability through the Hoeffding Functional Decomposition and TreeHFD Algorithm
Clement Benard
Explainability and InterpretabilityTabular
🎯 What it does: This paper proposes the TreeHFD algorithm, which estimates the Hoeffding Functional Decomposition (HFD) of tree ensembles using data samples, achieving an interpretable breakdown of tree models.
Tree of Preferences for Diversified Recommendation
Hanyang Yuan (Zhejiang University), Mingli Song (Zhejiang University)
Recommendation SystemGraph Neural NetworkTransformerLarge Language ModelText
🎯 What it does: This paper proposes a diversified recommendation framework based on large language models (LLM) called ToP-Rec. It constructs a Tree of Preferences using user attributes and interaction records to infer unexplored user interests through LLM systems, generating composite interactions based on these preferences to enhance training data, ultimately training a recommendation model that balances diversity and relevance.
Tree-Based Premise Selection for Lean4
Zichen Wang (Peking University), Zaiwen Wen (Peking University)
RetrievalOptimizationGraph
🎯 What it does: A Lean4 premise retrieval framework based on expression tree structure is proposed, utilizing CSE for tree simplification, WL kernel for coarse screening, TED for fine-tuning, and multi-metric fusion to achieve efficient and accurate retrieval.
Tree-Guided Diffusion Planner
Hyeonseong Jeon (Seoul National University), Jaesik Park (Seoul National University)
OptimizationRobotic IntelligenceDiffusion model
🎯 What it does: This paper proposes a zero-shot testing planning framework based on a pre-trained diffusion model called Tree-Guided Diffusion Planner (TDP). It generates diverse trajectories in the search tree through dual-layer sampling and performs gradient-guided refinement to achieve efficient planning for non-convex, multi-objective, and non-differentiable constraint tasks.
Tree-Sliced Entropy Partial Transport
Viet-Hoang Tran (National University of Singapore), Tan Minh Nguyen
GenerationData SynthesisOptimizationImagePoint Cloud
🎯 What it does: This paper proposes Tree-Sliced Entropy Partial Transport (PartialTSW), a distance that transforms partial transport into a balanced optimal transport in tree metric spaces and achieves efficient computation through tree slicing.
TreeGen: A Bayesian Generative Model for Hierarchies
Marcel Kollovieh (Technical University of Munich), Stephan Günnemann (Technical University of Munich)
GenerationData SynthesisGraph Neural NetworkTransformerFlow-based ModelGraph
🎯 What it does: Proposes the TreeGen generative model, which derives tree structures using a Bayesian Flow Network.
TreeSplat: Mergeable Tree for Deformable Gaussian Splatting
Qiuhong Shen (National University of Singapore), Xinchao Wang (National University of Singapore)
GenerationData SynthesisComputational EfficiencyNeural Radiance FieldGaussian SplattingVideo
🎯 What it does: This paper proposes TreeSplat, which utilizes a mergeable hierarchical tree structure to collaboratively model the motion of dynamic Gaussian Splatting.
TreeSynth: Synthesizing Diverse Data from Scratch via Tree-Guided Subspace Partitioning
Sheng Wang (University of Hong Kong), Chuan Wu (University of Hong Kong)
GenerationData SynthesisLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper proposes a tree-guided subspace partitioning method (TREESYNTH) that recursively divides the task data space into mutually exclusive and complementary atomic subspaces from a global perspective. Samples are then synthesized using LLM within each subspace, and finally aggregated into a high-diversity, fully covered dataset.
TREND: Unsupervised 3D Representation Learning via Temporal Forecasting for LiDAR Perception
Runjian Chen (University of Hong Kong), Alex Wong (Yale University)
Object DetectionSegmentationAutonomous DrivingRepresentation LearningNeural Radiance FieldPoint Cloud
🎯 What it does: This paper proposes an unsupervised 3D representation learning framework called TREND, which utilizes temporal prediction of LiDAR sequences to pre-train a point cloud encoder.
Tri-MARF: A Tri-Modal Multi-Agent Responsive Framework for Comprehensive 3D Object Annotation
Jusheng Zhang (Sun Yat-sen University), Keze Wang (Sun Yat-sen University)
Object DetectionTransformerReinforcement LearningVision Language ModelMultimodalityPoint Cloud
🎯 What it does: Proposes the Tri-MARF tri-modal multi-agent framework for 3D object annotation.
TRiCo: Triadic Game-Theoretic Co-Training for Robust Semi-Supervised Learning
Hongyang He (University of Warwick), Wenqiao Zhang (Zhejiang University)
🎯 What it does: What was done
TRIDENT: Tri-Modal Molecular Representation Learning with Taxonomic Annotations and Local Correspondence
Feng Jiang (University of Texas at Arlington), Junzhou Huang (University of Texas at Arlington)
Representation LearningDrug DiscoveryTransformerContrastive LearningTextMultimodality
🎯 What it does: Proposes the TRIDENT tri-modal molecular representation learning framework, integrating SMILES, textual descriptions, and hierarchical classification annotations for global and local alignment.
TRIM: Scalable 3D Gaussian Diffusion Inference with Temporal and Spatial Trimming
Zeyuan Yin (Michigan State University), Xiaoming Liu (Michigan State University)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelImageText
🎯 What it does: This paper proposes the TRIM framework, which accelerates the inference of 3D Gaussian diffusion models using two post-training strategies: trajectory pruning and instance masking.
Triplets Better Than Pairs: Towards Stable and Effective Self-Play Fine-Tuning for LLMs
Yibo Wang (Nanjing University), Lijun Zhang (Nanjing University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: In the absence of massive expert-annotated data, a triplet self-play fine-tuning method (T SPIN‑) is proposed, which iteratively allows the model to compete on samples generated by itself to improve performance.
Tropical Attention: Neural Algorithmic Reasoning for Combinatorial Algorithms
Baran Hashemi (Data Science Lab Technical University of Munich), Ruriko Yoshida (Naval Postgraduate School)
OptimizationTransformerSequentialBenchmark
🎯 What it does: This paper proposes the Tropical Attention mechanism, which replaces the traditional softmax pointwise attention, reasoning in the tropical projective space, and proves its approximation to maximum sum dynamic programming and closure.
TRoVe: Discovering Error-Inducing Static Feature Biases in Temporal Vision-Language Models
Maya Varma (Stanford University), Curtis Langlotz (Stanford University)
TransformerPrompt EngineeringVision Language ModelVideoMultimodality
🎯 What it does: An automated method called TROVE has been developed to discover static feature biases that lead to errors in temporal visual-language models, and to improve model performance on real tasks through prompt learning.
Tru-POMDP: Task Planning Under Uncertainty via Tree of Hypotheses and Open-Ended POMDPs
Wenjing Tang (Shanghai Jiao Tong University), Panpan Cai (Shanghai Jiao Tong University)
OptimizationRobotic IntelligenceTransformerLarge Language ModelReinforcement LearningAgentic AIMultimodality
🎯 What it does: A framework named Tru-POMDP is proposed, which uses large language models (LLM) to generate a Tree of Hypotheses to construct particle beliefs for kitchen object rearrangement tasks that are partially observable and have uncertain goals, and combines Bayesian filtering with online POMDP tree search for task planning.
True Impact of Cascade Length in Contextual Cascading Bandits
Hyunjun Choi, Min-hwan Oh (Seoul National University)
Recommendation SystemOptimizationReinforcement LearningTabular
🎯 What it does: This paper proposes an online mirror descent-based UCB algorithm called UCB-CLB, redefines the scheduling strategy of contextual cascading bandits, and proves precise bounds on the impact of cascade length K on cumulative rewards.
True Zero-Shot Inference of Dynamical Systems Preserving Long-Term Statistics
Christoph Jürgen Hemmer (Heidelberg University), Daniel Durstewitz (Heidelberg University)
Recurrent Neural NetworkMixture of ExpertsTime SeriesSequentialMagnetic Resonance Imaging
🎯 What it does: Developed the DynaMix model, achieving the reconstruction of long-term statistical properties of any dynamic system under zero-shot conditions.
Trust Region Constrained Measure Transport in Path Space for Stochastic Optimal Control and Inference
Denis Blessing (Karlsruhe Institute of Technology), Gerhard Neumann (Karlsruhe Institute of Technology)
GenerationOptimizationReinforcement LearningDiffusion modelMultimodality
🎯 What it does: This paper proposes a trust-region-based Stochastic Optimal Control (SOC) framework to approximate target measures in path space, enabling high-dimensional sampling, transition path sampling, and reward fine-tuning for text-to-image diffusion models.
Trust Region Reward Optimization and Proximal Inverse Reward Optimization Algorithm
Yang Chen (Shanghai Artificial Intelligence Laboratory), Michael J. Witbrock (University of Auckland)
OptimizationRobotic IntelligenceReinforcement LearningAgentic AISequential
🎯 What it does: This paper proposes a non-adversarial inverse reinforcement learning framework TRRO and implements a practical algorithm PIRO to stably learn the expert reward function and generate high-quality imitation policies.
Trust, But Verify: A Self-Verification Approach to Reinforcement Learning with Verifiable Rewards
Xiaoyuan Liu (Chinese University of Hong Kong), Dong Yu (Tencent)
TransformerReinforcement LearningPrompt EngineeringText
🎯 What it does: Designed and implemented an online reinforcement learning framework called RISE, which can simultaneously train large language models to solve problems and self-verify their generated answers.
TRUST: Test-Time Refinement using Uncertainty-Guided SSM Traverses
Sahar Dastani (École de technologie supérieure), Christian Desrosiers (École de technologie supérieure)
ClassificationDomain AdaptationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A testing-time adaptive method called TRUST for Visual State Space Models (VMamba) is proposed, which generates multiple causal perspectives by traversing different directional scanning arrangements and updates model parameters using pseudo-labels, ultimately performing weight averaging to enhance cross-domain robustness.
Truth over Tricks: Measuring and Mitigating Shortcut Learning in Misinformation Detection
Herun Wan (Xi'an Jiaotong University), Zhixiong Su (Xi'an Jiaotong University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes the TRUTHOVERTRICKS evaluation framework, which systematically assesses the dependence of rumor detection models on internal and external shortcuts; constructs two new factual rumor datasets, NQ-Misinfo and Streaming-Misinfo; and introduces the LLM-assisted data augmentation framework SMF to mitigate the model's reliance on shortcuts.
Truthful Aggregation of LLMs with an Application to Online Advertising
Ermis Soumalias, Sven Seuken
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: A novel auction mechanism named MOSAIC is proposed to aggregate advertisers' preferences for responses generated by large language models (LLMs) in online advertising scenarios, producing responses that align with the interests of both users and advertisers without altering the LLM weights.
TS-MOF: Two-Stage Multi-Objective Fine-tuning for Long-Tailed Recognition
Zhe Zhao (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
ClassificationRecognitionConvolutional Neural NetworkSupervised Fine-TuningImage
🎯 What it does: A two-stage multi-objective fine-tuning framework, TS-MOF, is proposed to address the class imbalance problem in long-tail recognition. First, high-quality general features are obtained through pre-training; then, the feature network is frozen, and only the multi-task classification head undergoes multi-objective fine-tuning.
TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster
Kanghui Ning (University of Connecticut), Dongjin Song (University of Connecticut)
GenerationRetrievalAnomaly DetectionTime SeriesRetrieval-Augmented Generation
🎯 What it does: This paper proposes TS-RAG, a retrieval-augmented generation (RAG) framework that utilizes a pre-trained time series foundation model (TSFM) and retrieved similar time series segments to achieve high-accuracy predictions in a zero-shot setting.
TSENOR: Highly-Efficient Algorithm for Finding Transposable N:M Sparse Masks
Xiang Meng (Massachusetts Institute of Technology), Rahul Mazumder (Massachusetts Institute of Technology)
OptimizationComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: An efficient algorithm TSENOR is proposed for generating binary masks that satisfy the transposed N:M sparse constraints, enabling forward and backward acceleration on large-scale language models.
TTRL: Test-Time Reinforcement Learning
Yuxin Zuo (Tsinghua University), Bowen Zhou (Tsinghua University)
Large Language ModelReinforcement LearningText
🎯 What it does: The research uses reinforcement learning for self-evolution of large language models (TTRL) on unlabeled test data.
TTS-VAR: A Test-Time Scaling Framework for Visual Auto-Regressive Generation
Zhekai Chen (Hong Kong University), Xihui Liu (Hong Kong University)
GenerationData SynthesisTransformerReinforcement LearningImageText
🎯 What it does: A testing time scaling framework for visual autoregressive (VAR) models, TTS-VAR, is proposed to enhance the quality of generated images through adaptive batching, coarse-scale clustering diversity search, and fine-scale potential resampling.
Turbocharging Gaussian Process Inference with Approximate Sketch-and-Project
Pratik Rathore (Stanford University), Madeleine Udell (Stanford University)
OptimizationGaussian SplattingTabular
🎯 What it does: An approximate, distributed, accelerated sketch-and-project algorithm ADASAP is proposed for large-scale Gaussian process inference.
Turning Sand to Gold: Recycling Data to Bridge On-Policy and Off-Policy Learning via Causal Bound
Tal Fiskus (Bar Ilan University), Uri Shaham (Bar Ilan University)
Reinforcement LearningSequential
🎯 What it does: Proposes the SUFT method, which utilizes the Neyman-Rubin potential outcomes framework in deep reinforcement learning to construct an upper bound on the true (on-policy) loss, and enhances sample efficiency and learning stability by calculating the SUFT OPE constraint through saving the value network outputs in experience replay.
Turning the Tables: Enabling Backward Transfer via Causal-Aware LoRA in Continual Learning
Chaoyang Li (Harbin Institute of Technology), Qing Liao (Harbin Institute of Technology)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The CaLoRA framework is proposed, studying backward knowledge transfer in continual learning through PEFT, achieved via causal attribution and gradient adaptation.
TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation
Yehjin Shin (Korea Advanced Institute of Science and Technology), Noseong Park (Korea Advanced Institute of Science and Technology)
Recommendation SystemGraph Neural NetworkSequential
🎯 What it does: A sequence recommendation model based on time-varying convolutional filters, TV-Rec, is proposed, which captures the temporal changes in user behavior sequences using node transformation filters from graph signal processing, completely omitting the self-attention module.
Twilight: Adaptive Attention Sparsity with Hierarchical Top-$p$ Pruning
Chaofan Lin (Tsinghua University), Mingyu Gao (Tsinghua University)
TransformerLarge Language ModelText
🎯 What it does: Designed and implemented the Twilight framework, which incorporates an adaptive budget mechanism into the existing top-k sparse attention models, using topp sampling to dynamically prune key tokens in the KV cache.
TwinMarket: A Scalable Behavioral and Social Simulation for Financial Markets
Yuzhe YANG, Benyou Wang (Chinese University of Hong Kong)
Recommendation SystemAnomaly DetectionOptimizationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkTransformerLarge Language ModelAgentic AIPrompt EngineeringTextTabularTime SeriesFinance Related
🎯 What it does: A multi-agent framework called TwinMarket, based on large language models, has been constructed to simulate stock market investor behavior and study how micro behaviors lead to macro socio-economic phenomena.
Two Experts Are All You Need for Steering Thinking: Reinforcing Cognitive Effort in MoE Reasoning Models Without Additional Training
Mengru Wang (Zhejiang University), Dong Yu (Tencent)
Mixture of ExpertsTextBiomedical DataPhysics Related
🎯 What it does: This paper proposes to identify experts related to reasoning using nPMI in the Mixture-of-Experts structure of large reasoning models, and to enhance only a few experts during inference, thereby improving reasoning accuracy and efficiency.
Two Heads are Better than One: Simulating Large Transformers with Small Ones
Hantao Yu (Columbia University), Josh Alman (Columbia University)
TransformerLarge Language Model
🎯 What it does: It is proven that the computation of large-scale Transformers can be simulated using Transformers that only process short sequences (small Transformers), and the optimal number of oracle calls required for the simulation is provided.
Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising
Mateo Clémente, Yinchuan Li (Huawei)
OptimizationRobotic IntelligenceReinforcement LearningDiffusion modelSequential
🎯 What it does: The researchers propose a Genetic Diffusion Policy (GDP) based on genetic algorithms, which significantly reduces sampling steps (only 2 steps) and improves control performance by selecting low-discrepancy (OoD) trajectories in a low-dimensional robot action space.
Two‑Stage Learning of Stabilizing Neural Controllers via Zubov Sampling and Iterative Domain Expansion
Haoyu Li (University of Illinois), Huan Zhang (University of Illinois)
OptimizationReinforcement LearningTime Series
🎯 What it does: A two-stage training framework is proposed, which jointly synthesizes a neural network controller for continuous-time systems with a Lyapunov function. It significantly reduces training conservativeness through a physics-inspired loss based on the Zubov theorem and dynamic training domain expansion, ultimately achieving a larger region of attraction (ROA).
Týr-the-Pruner: Structural Pruning LLMs via Global Sparsity Distribution Optimization
Guanchen Li (Advanced Micro Devices Inc), Emad Barsoum (Advanced Micro Devices Inc)
OptimizationTransformerLarge Language ModelText
🎯 What it does: Perform structural pruning on large language models, construct a supernet, and use evolutionary search to find the optimal sparse distribution, achieving end-to-end global pruning.
U-CAN: Unsupervised Point Cloud Denoising with Consistency-Aware Noise2Noise Matching
Junsheng Zhou (Tsinghua University), Zhizhong Han (Wayne State University)
RestorationGraph Neural NetworkImagePoint Cloud
🎯 What it does: An unsupervised point cloud denoising framework U-CAN is proposed, achieving multi-step denoising through noise-to-noise matching and consistency constraints.
U-REPA: Aligning Diffusion U-Nets to ViTs
Yuchuan Tian (Peking University), Yunhe Wang (Huawei)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelImage
🎯 What it does: This paper proposes the U-REPA framework, which aligns the hidden states of U-Net with the Vision Transformer encoder to accelerate the training of diffusion models.
UEPI: Universal Energy-Behavior-Preserving Integrators for Energy Conservative/Dissipative Differential Equations
Elena Celledoni (Norwegian University of Science and Technology), Takaharu Yaguchi (Kobe University)
Time SeriesPhysics RelatedOrdinary Differential Equation
🎯 What it does: This paper studies a neural network-based energy-preserving numerical integrator, proposing a general discrete gradient framework and achieving high-precision energy conservation or dissipation schemes through learning.
UFM: A Simple Path towards Unified Dense Correspondence with Flow
Yuchen Zhang (Carnegie Mellon University), Wenshan Wang (Carnegie Mellon University)
Image TranslationTransformerOptical FlowImageBenchmark
🎯 What it does: Proposes the Unified Flow & Matching (UFM) model, which unifies the learning of optical flow and wide baseline matching tasks, achieving high-resolution pixel-level correspondences and visibility predictions.
UFO-RL: Uncertainty-Focused Optimization for Efficient Reinforcement Learning Data Selection
Yang Zhao (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
OptimizationReinforcement LearningText
🎯 What it does: This paper proposes the UFO-RL framework, which utilizes a confidence assessment method based on single forward inference to efficiently filter RL training data for LLMs, thereby focusing learning on the model's Zone of Proximal Development (ZPD).
UFO: A Unified Approach to Fine-grained Visual Perception via Open-ended Language Interface
Hao Tang (Peking University), Liwei Wang (Peking University)
Object DetectionSegmentationTransformerVision Language ModelImageTextMultimodality
🎯 What it does: Proposes the UFO framework, which unifies the mapping of fine-grained visual tasks such as detection and segmentation to an open language interface, achieving a unified model without a task-specific decoder;
UFT: Unifying Supervised and Reinforcement Fine-Tuning
Mingyang Liu (Massachusetts Institute of Technology), Asuman E. Ozdaglar (Massachusetts Institute of Technology)
OptimizationTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringText
🎯 What it does: A unified post-training framework UFT is proposed, which integrates Supervised Fine-Tuning (SFT) and Reinforcement Learning Fine-Tuning (RFT) into a continuous process, guiding the model to smoothly transition between exploration and memory using hints.
UGG-ReID: Uncertainty-Guided Graph Model for Multi-Modal Object Re-Identification
Xixi Wan (Anhui University), Jin Tang (Anhui University)
RecognitionRetrievalConvolutional Neural NetworkGraph Neural NetworkMixture of ExpertsImageMultimodality
🎯 What it does: Proposes the UGG-ReID framework for multi-modal target ReID;
UGM2N: An Unsupervised and Generalizable Mesh Movement Network via M-Uniform Loss
Zhichao Wang (National University of Defense Technology), Jie Liu (National University of Defense Technology)
Graph Neural NetworkTransformerMeshGraph
🎯 What it does: An unsupervised and generalizable grid-moving network UGM2N is proposed, which utilizes local node patches and graph Transformers to achieve adaptive grid generation without the need for pre-adapted grids.
UGoDIT: Unsupervised Group Deep Image Prior Via Transferable Weights
Shijun Liang (University of Michigan), Saiprasad Ravishankar (Michigan State University)
RestorationSuper ResolutionConvolutional Neural NetworkAuto EncoderImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: A completely unsupervised inverse problem solving framework called UGoDIT is proposed, which achieves fast and high-quality reconstruction of unknown measurements by training a shared encoder and multiple decoders of a deep image prior model on only a small number of degraded images (such as under-sampled images).
UI-Genie: A Self-Improving Approach for Iteratively Boosting MLLM-based Mobile GUI Agents
Han Xiao (CUHK MMLab), Hongsheng Li (CUHK MMLab)
Robotic IntelligenceReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningVision Language ModelImageTextMultimodality
🎯 What it does: A self-improvement framework called UI-Genie is proposed for training and enhancing mobile GUI agents based on multimodal large language models (MLLMs), focusing on addressing the challenges of trajectory result validation and the lack of high-quality training data.
Ultra-high Resolution Watermarking Framework Resistant to Extreme Cropping and Scaling
Nan Sun (Huazhong University of Science and Technology), Chengxin Zhao (Huazhong University of Science and Technology)
GenerationData SynthesisCompressionNeural Radiance FieldImage
🎯 What it does: A framework for embedding high-resolution image watermarks based on Implicit Neural Representation (INR) is proposed, capable of generating pixel-level watermarks at any resolution while maintaining strong robustness.
UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset
Chen Zhao (Nanjing University), Ying Tai (Nanjing University)
GenerationData SynthesisSuper ResolutionDiffusion modelImageText
🎯 What it does: A high-quality ultra-high-resolution image dataset UltraHR-100K, consisting of 100,000 images with a resolution exceeding 3K, was constructed, and a frequency-aware post-training method (DOTS+SWFR) for UHR detail generation was proposed on this dataset, significantly enhancing the detail performance of text-to-image models at ultra-high resolutions.
UltraLED: Learning to See Everything in Ultra-High Dynamic Range Scenes
Yuang Meng (Nankai University), Chongyi Li (Nankai University)
RestorationConvolutional Neural NetworkImage
🎯 What it does: A two-stage network called UltraLED is proposed to reconstruct ultra-high dynamic range (UHDR) scenes using only a single frame RAW image.
Ultrametric Cluster Hierarchies: I Want ‘em All!
Andrew Draganov (Aarhus University), Ira Assent (Aarhus University)
OptimizationTabular
🎯 What it does: It is proposed that center-based clustering (k-means, k-median, k-center) can obtain optimal solutions for all k in O(n) time on any ultrametric (represented by LCA-tree), and these solutions form a hierarchy; based on this, the SHiP framework is constructed, which can quickly generate various clustering hierarchies and partitions after a single ultrametric fitting.
UMA: A Family of Universal Models for Atoms
Brandon M Wood, C. Lawrence Zitnick (FAIR at Meta)
Graph Neural NetworkMixture of ExpertsGraphBenchmarkPhysics Related
🎯 What it does: This study proposes and trains a series of universal atomic machine learning potential models called UMA (Universal Models for Atoms), which can directly replace DFT for energy, force, and stress calculations without fine-tuning, applicable to various fields such as materials, molecules, and catalysts.
UMAMI: Unifying Masked Autoregressive Models and Deterministic Rendering for View Synthesis
Tung Le, Stephan Mandt (University of California, Irvine)
GenerationData SynthesisComputational EfficiencyTransformerDiffusion modelVideo
🎯 What it does: A novel perspective synthesis (NVS) hybrid framework called UMAMI is proposed, which combines deterministic rendering with autoregressive diffusion models, enabling rapid generation of visible areas while reasonably inferring unobserved regions.
UMoE: Unifying Attention and FFN with Shared Experts
Yuanhang Yang (Institute of Science Tokyo), Jing Li (Hong Kong Polytechnic University)
TransformerMixture of ExpertsText
🎯 What it does: This paper proposes a unified sparse expert mixture model UMoE, which integrates the attention layer and feedforward network layer of the Transformer into a shareable FFN-style expert through a pre-mixing attention reformulation, achieving improved model performance while maintaining the same parameter count.
un$^2$CLIP: Improving CLIP's Visual Detail Capturing Ability via Inverting unCLIP
Yinqi Li (Institute of Computing Technology, Chinese Academy of Sciences), Xilin Chen (Institute of Computing Technology, Chinese Academy of Sciences)
SegmentationGenerationTransformerVision Language ModelDiffusion modelContrastive LearningImageMultimodality
🎯 What it does: Based on the pre-trained CLIP image encoder, reverse fine-tuning is achieved using the unCLIP generative model to enhance its ability to capture visual details.