ICLR 2025 Papers — Page 35
International Conference on Learning Representations · 3704 papers
Towards Robust Multimodal Open-set Test-time Adaptation via Adaptive Entropy-aware Optimization
Hao Dong (ETH Zurich), Olga Fink (EPFL)
Domain AdaptationOptimizationVideoMultimodalityAudio
🎯 What it does: A framework for adaptive multi-modal open set testing (MM-OSTTA) called AEO is proposed and validated, which aims to enhance the entropy difference between known and unknown samples online, thereby improving unknown class detection and overall adaptation performance.
Towards Scalable Exact Machine Unlearning Using Parameter-Efficient Fine-Tuning
Somnath Basu Roy Chowdhury (University of North Carolina Chapel Hill), Snigdha Chaturvedi (Google Research)
TransformerLarge Language ModelSupervised Fine-TuningImageText
🎯 What it does: This paper proposes an Exact Machine Unlearning framework called S3T, based on sharding, slicing, and Parameter-Efficient Fine-Tuning (PEFT), which enables a quick and training-free response to data deletion requests.
Towards Scalable Topological Regularizers
Wong Hiu Tung, Hong Yan (City University of Hong Kong)
ClassificationGenerationGenerative Adversarial NetworkImagePoint Cloud
🎯 What it does: A scalable topological regularization method based on Principal Persistence Measures (PPM) is proposed and applied to tasks such as Generative Adversarial Networks (GAN) and Semi-Supervised Learning (SSL) to improve the model's generation quality and classification performance.
Towards Self-Supervised Covariance Estimation in Deep Heteroscedastic Regression
Megh Shukla (École Polytechnique Fédérale de Lausanne), Alexandre Alahi (École Polytechnique Fédérale de Lausanne)
Pose EstimationOptimizationTransformerContrastive LearningTabularSequential
🎯 What it does: This paper studies covariance self-supervised estimation in deep heteroscedastic regression, proposing the use of an upper bound of the 2-Wasserstein distance and neighborhood-based pseudo-labels to achieve more stable and cost-effective covariance learning.
Towards Semantic Equivalence of Tokenization in Multimodal LLM
Shengqiong Wu (National University of Singapore), Shuicheng YAN (National University of Singapore)
TransformerLarge Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: Proposes a dynamic semantic equivalent visual tokenizer SeTok and integrates it into MLLM SETOKIM, achieving a variable number of semantically complete visual tokens.
Towards Synergistic Path-based Explanations for Knowledge Graph Completion: Exploration and Evaluation
Tengfei Ma (Hunan University), xiangxiang Zeng
Explainability and InterpretabilityKnowledge DistillationGraph Neural NetworkGraph
🎯 What it does: KGExplainer provides interpretable path explanations for knowledge graph completion models, identifying collaborative paths and assessing their credibility.
Towards Unbiased Learning in Semi-Supervised Semantic Segmentation
Rui Sun (University of Science and Technology of China), Tianzhu Zhang (University of Science and Technology of China)
SegmentationDiffusion modelImage
🎯 What it does: Proposes DiffMatch, modeling the semi-supervised semantic segmentation task as a conditional discrete data generation problem, and generates pseudo-labels through a diffusion model;
Towards Understanding Text Hallucination of Diffusion Models via Local Generation Bias
Rui Lu (Tsinghua University), Mengdi Wang (Princeton University)
GenerationExplainability and InterpretabilityTransformerDiffusion modelScore-based ModelText
🎯 What it does: This study investigates the phenomenon of hallucination in text generation using diffusion models, proposing and quantifying the 'Local Generation Bias' and its measurement index LDR, along with theoretical analysis and experimental validation.
Towards Understanding the Robustness of Diffusion-Based Purification: A Stochastic Perspective
Yiming Liu (Sun Yat-Sen University), Liang Lin (Sun Yat-Sen University)
OptimizationAdversarial AttackDiffusion modelImage
🎯 What it does: This paper re-examines the robustness of diffusion-based purification (DBP) from the perspective of randomness, proposes a Deterministic White-box Attack (DW-box) evaluation method, and experimentally verifies that randomness is the main source of robustness for DBP. It then introduces two techniques, Adversarial Denoising Diffusion Training (ADDT) and Rank-Based Gaussian Mapping (RBGM), to further enhance the adversarial robustness of the DBP model.
Towards Understanding the Universality of Transformers for Next-Token Prediction
Michael Eli Sander, Gabriel Peyré (Ecole Normale Supérieure)
TransformerSequential
🎯 What it does: This paper proposes a causal Transformer explanation framework based on kernel methods and provides an explicit construction that allows the Transformer to approximate the next token through a 'causal kernel gradient descent' method given a context.
Towards Understanding Why FixMatch Generalizes Better Than Supervised Learning
Jingyang Li (National University of Singapore), Pan Zhou (Singapore Management University)
ClassificationSegmentationConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: Theoretical analysis of the generalization advantages of FixMatch on CNNs, and the proposal of the SA-FixMatch improvement algorithm based on semantic-aware CutOut.
Towards Understanding Why Label Smoothing Degrades Selective Classification and How to Fix It
Guoxuan Xia (Imperial College London), Christos-Savvas Bouganis (Imperial College London)
ClassificationSegmentationImage
🎯 What it does: The paper studies the negative impact of Label Smoothing on Selective Classification and provides a gradient-level explanation.
Towards Unified Human Motion-Language Understanding via Sparse Interpretable Characterization
Guangtao Lyu (Xidian University), Cheng Deng (Institute for Infocomm Research)
RetrievalExplainability and InterpretabilityTransformerContrastive LearningVideoTextMultimodality
🎯 What it does: A multi-stage sparse interpretable vocabulary space pre-training framework is proposed, which maps human actions and text into a shared sparse vocabulary space to enhance cross-modal semantic alignment and interpretability.
Towards Universality: Studying Mechanistic Similarity Across Language Model Architectures
Junxuan Wang (Fudan University), Xipeng Qiu (Fudan University)
Explainability and InterpretabilityTransformerAuto EncoderText
🎯 What it does: A mechanism interpretability study of two language model architectures, Transformer and Mamba, verifying their universality at the feature and circuit levels.
TPO: Aligning Large Language Models with Multi-branch & Multi-step Preference Trees
Weibin Liao (Peking University), Yasha Wang (Peking University)
Recommendation SystemOptimizationTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: This paper proposes TPO, a method for directly learning multi-branch and multi-step preferences based on tree-shaped preference trees, as a replacement for traditional DPO.
TRACE: Temporal Grounding Video LLM via Causal Event Modeling
Yongxin Guo (Chinese University of Hong Kong), Xiaoying Tang (Chinese University of Hong Kong)
GenerationRetrievalTransformerLarge Language ModelVision Language ModelVideoText
🎯 What it does: A causal event modeling framework is proposed, constructing the TRACE video LLM, which predicts the next event in an autoregressive manner based on preceding events, video, and text instructions, achieving zero-shot video temporal localization tasks.
TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies
Ruijie Zheng (University of Maryland), Jianwei Yang (Microsoft Research)
Robotic IntelligenceTransformerPrompt EngineeringVision-Language-Action ModelImageVideo
🎯 What it does: Proposed and implemented a visual trace prompting technology that enhances the spatiotemporal perception of the VLA (Vision-Language-Action) model using visual traces, and based on this, fine-tuned OpenVLA and Phi-3-Vision to obtain TraceVLA and its 4B version.
Tracing Representation Progression: Analyzing and Enhancing Layer-Wise Similarity
Jiachen Jiang (Ohio State University), Zhihui Zhu (Ohio State University)
ClassificationRepresentation LearningTransformerLarge Language ModelImageText
🎯 What it does: This paper studies the representation similarity within the layers of Transformer, proposing a simple metric based on sample cosine similarity, and uses it to explain the improvement in inter-layer prediction probabilities and saturation events; subsequently, it designs aligned training, allowing all layers to share a single classifier, enhancing shallow layer performance and achieving a multi-exit model.
Track-On: Transformer-based Online Point Tracking with Memory
Görkay Aydemir (Koc University), Fatma Guney
Object TrackingTransformerVideo
🎯 What it does: Track-On is proposed, an online long-term point tracking model based on Transformer, which utilizes two memory modules to achieve inter-frame continuity and employs patch classification and refined offset methods for correspondence point localization.
Tracking objects that change in appearance with phase synchrony
Sabine Muzellec (CerCo CNRS Universite de Toulouse), Thomas Serre (Brown University)
Object TrackingRecurrent Neural NetworkVideo
🎯 What it does: A neural synchronous attention mechanism based on complex-valued recurrent neural networks (CV-RNN) has been developed to address tracking tasks when the appearance of objects changes over time.
Tracking the Copyright of Large Vision-Language Models through Parameter Learning Adversarial Images
Yubo Wang (University of Science and Technology of China), Linli Xu (University of Science and Technology of China)
Adversarial AttackVision Language ModelImageText
🎯 What it does: A visual language model copyright tracking method based on Parameter Learning Attack (PLA) is proposed, which can construct trigger images without changing the original model parameters.
Tractable Multi-Agent Reinforcement Learning through Behavioral Economics
Eric Mazumdar (California Institute of Technology), Laixi Shi (California Institute of Technology)
Reinforcement LearningTabular
🎯 What it does: This paper studies the introduction of risk aversion and bounded rationality in multi-agent reinforcement learning and proposes a new computable equilibrium—Risk Quantified Response Equilibrium (RQE), proving its solvability in all finite-horizon Markov games.
Train Small, Infer Large: Memory-Efficient LoRA Training for Large Language Models
Jun Zhang (Zhejiang University), Kunlong Zhou (Guangdong OPPO Mobile Telecommunications Corporation)
CompressionComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes LORAM, a memory-efficient fine-tuning framework that trains LoRA on a compressed model and infers on the full model.
Trained Transformer Classifiers Generalize and Exhibit Benign Overfitting In-Context
Spencer Frei (University of California Davis), Gal Vardi (Weizmann Institute of Science)
ClassificationTransformerTabular
🎯 What it does: This study investigates the behavior of linear transformers in random linear classification tasks, analyzes the implicit regularization of gradient descent, and explores the impact of the number of pre-training tasks and context examples on the model's generalization ability during testing.
Training Free Exponential Context Extension via Cascading KV Cache
Jeffrey Willette (KAIST AI), Sung Ju Hwang (KAIST AI)
TransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A training-agnostic cascading KV cache is proposed, which dynamically retains important tokens and extends context length with linear complexity.
Training Free Guided Flow-Matching with Optimal Control
Luran Wang (University of Cambridge), Ge Liu (University of Illinois Urbana-Champaign)
GenerationOptimizationDrug DiscoveryFlow-based ModelImageTabularOrdinary Differential Equation
🎯 What it does: This paper proposes a training-free, optimal control-based flow matching guidance framework called OC-Flow, which can guide pre-trained ODE generative models in Euclidean space and the SO(3) rotation group.
Training Language Models on Synthetic Edit Sequences Improves Code Synthesis
Ulyana Piterbarg (New York University), Rob Fergus (New York University)
GenerationComputational EfficiencyAI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The paper designs the LintSeq algorithm, which decomposes the original program into insertion edit sequences guided by a static linter, thereby reformulating the code generation task as a step-by-step sequence generation problem.
Training Language Models to Self-Correct via Reinforcement Learning
Aviral Kumar (Carnegie Mellon University), Aleksandra Faust (Google DeepMind)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: Train LLM through multi-round reinforcement learning to achieve self-correction without relying on external supervision or teacher models;
Training Large Language Models for Retrieval-Augmented Question Answering through Backtracking Correction
Huawen Feng (South China University of Technology), Qianli Ma (South China University of Technology)
RetrievalTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextRetrieval-Augmented Generation
🎯 What it does: Training large language models to enhance their discrimination and reasoning abilities regarding retrieved documents in the retrieval-augmented generation (RAG) task through self-correction learning.
Training Neural Networks as Recognizers of Formal Languages
Alexandra Butoi (ETH Zurich), Brian DuSell (ETH Zurich)
RecognitionRecurrent Neural NetworkTransformerTextBenchmark
🎯 What it does: This paper proposes a general method for training neural networks as recognizers of formal languages and provides an efficient length-constrained sampling algorithm for regular languages.
Training Nonlinear Transformers for Chain-of-Thought Inference: A Theoretical Generalization Analysis
Hongkang Li (Rensselaer Polytechnic Institute), Meng Wang (Rensselaer Polytechnic Institute)
OptimizationExplainability and InterpretabilityTransformerTabularChain-of-Thought
🎯 What it does: This paper theoretically proves that training a single-layer nonlinear attention Transformer can achieve chain-of-thought (CoT) capabilities and provides an upper bound on the number of training samples and iterations.
Training on the Test Task Confounds Evaluation and Emergence
Ricardo Dominguez-Olmedo (Max Planck Institute for Intelligent Systems), Moritz Hardt (Max Planck Institute for Intelligent Systems)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper studies the confounding effect of 'using test task knowledge during training' on the evaluation of large language models and proposes a correction method by uniformly fine-tuning all models with sufficient task-specific data before evaluation; experiments show that after correction, the performance gap between models released at different times disappears, and the previously observed 'emergence' phenomenon decreases as the training amount of the test task increases.
Training One-Dimensional Graph Neural Networks is NP-Hard
Robert Ganian (TU Wien), Simon Wietheger (TU Wien)
Graph Neural NetworkGraph
🎯 What it does: The paper studies the computational complexity of training Graph Neural Networks (GNNs) in one-dimensional (1-dimensional) scenarios, proving that training 1-dimensional GNNs with ReLU activation functions and SUM/MEAN/SPECTRAL aggregation functions is NP-hard; it also provides algorithmic upper bounds and polynomially solvable cases under ReLU, linear activation, and special graph structures.
Training Robust Ensembles Requires Rethinking Lipschitz Continuity
Ali Ebrahimpour-Boroojeny, Varun Chandrasekaran (University of Illinois Urbana-Champaign)
OptimizationAdversarial AttackConvolutional Neural NetworkImage
🎯 What it does: This study investigates the impact of Lipschitz continuity on the transferability of attacks and proposes a new ensemble training method called LOTOS, which enhances the robustness of model ensembles through hierarchical orthogonalization.
Training-Free Activation Sparsity in Large Language Models
James Liu (Massachusetts Institute of Technology), Ben Athiwaratkun (Together AI)
TransformerLarge Language ModelText
🎯 What it does: We propose TEAL, a training-independent magnitude pruning activation sparsity method that can achieve 40-50% full model sparsity on modern LLMs;
Training-free Camera Control for Video Generation
Chen Hou (University of Science and Technology of China), Zhibo Chen (University of Science and Technology of China)
GenerationData SynthesisDepth EstimationDiffusion modelVideoPoint Cloud
🎯 What it does: A training-free camera control framework, CamTrol, based on 3D point clouds and noise layout priors, is proposed, which can achieve camera movement control for offline video diffusion models without any fine-tuning.
Training-Free Dataset Pruning for Instance Segmentation
Yalun Dai (Agency for Science Technology and Research), Yang He (National University of Singapore)
Object DetectionSegmentationImage
🎯 What it does: A training-independent dataset pruning framework (TFDP) is proposed, specifically designed for instance segmentation tasks;
Training-Free Diffusion Model Alignment with Sampling Demons
Po-Hung Yeh (Academia Sinica), Jun-cheng Chen
GenerationOptimizationVision Language ModelDiffusion modelImageTextStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A training-free, non-backpropagation sampling method called Demon is proposed, which optimizes the noise in the reverse-time SDE randomly to align the diffusion model with user preferences.
Training-free LLM-generated Text Detection by Mining Token Probability Sequences
Yihuai Xu (Zhejiang University), Fei Wu (Zhejiang University)
ClassificationAnomaly DetectionTransformerLarge Language ModelTextTime Series
🎯 What it does: A training-independent LLM text generation detection method called Lastde/Lastde++ is proposed, which identifies the differences between human writing and model-generated text through time series analysis of token probability sequences.
Training-Free Message Passing for Learning on Hypergraphs
Bohan Tang (University of Oxford), Xiaowen Dong (Shanghai Jiao Tong University)
ClassificationGraph Neural NetworkGraph
🎯 What it does: A training-free hypergraph message passing module, TF-MP-Module, is proposed, which constructs TF-HNN and can aggregate hypergraph structural information during the preprocessing stage for direct use in subsequent tasks.
Trajectory attention for fine-grained video motion control
Zeqi Xiao (Peking University), Xingang Pan (Nanyang Technological University)
GenerationData SynthesisTransformerDiffusion modelVideo
🎯 What it does: Proposes trajectory attention as an auxiliary branch for achieving fine control of camera motion in video generation, enhancing motion consistency and generation quality through its synergistic effect with the original temporal attention.
Trajectory-Class-Aware Multi-Agent Reinforcement Learning
Hyungho Na (Korea Advanced Institute of Science and Technology), Il-chul Moon
Reinforcement LearningSequential
🎯 What it does: A new multi-agent reinforcement learning framework called TRAMA is proposed, aimed at assisting agents in decision-making in multi-task environments by identifying trajectory categories.
Trajectory-LLM: A Language-based Data Generator for Trajectory Prediction in Autonomous Driving
Kairui Yang (Tianjin University), Di Lin (Shanghai Jiaotong University)
GenerationData SynthesisAutonomous DrivingTransformerLarge Language ModelTextMultimodality
🎯 What it does: A vehicle trajectory generator called Traj-LLM based on large language models has been designed, which can automatically generate realistic, controllable, and diverse vehicle trajectories based on brief interactive descriptions, and has created the L2T dataset containing 240k pieces of interactive text, behaviors, and trajectories.
Transformer Block Coupling and its Correlation with Generalization in LLMs
Murdock Aubry (University of Toronto), Vardan Papyan (University of Toronto)
TransformerLarge Language ModelText
🎯 What it does: A linearization analysis of the Jacobian matrix of large language model (LLM) Transformer blocks is conducted to study the relationship between the coupling of singular vectors between blocks and the model's generalization performance.
Transformer Encoder Satisfiability: Complexity and Impact on Formal Reasoning
Marco Sälzer (University of Kaiserslautern Landau), Martin Lange (University of Kassel)
Transformer
🎯 What it does: This paper studies the decidability and complexity of the Transformer Encoder Satisfiability Problem (TRSAT), proving that TRSAT is undecidable in commonly expressive classes of TE, and providing cases where it is decidable with a complexity of NEXPTIME under finite input lengths or fixed-width arithmetic.
Transformer Learns Optimal Variable Selection in Group-Sparse Classification
Chenyang Zhang (University of Hong Kong), Yuan Cao (University of Michigan)
ClassificationOptimizationTransformerTabular
🎯 What it does: This paper proves and demonstrates that a single layer Transformer can achieve optimal variable selection in group-sparse binary classification tasks through gradient descent, and can achieve high-accuracy transfer learning in downstream tasks with the same sparse structure.
Transformer Meets Twicing: Harnessing Unattended Residual Information
Laziz Abdullaev, Tan Minh Nguyen
RecognitionSegmentationTransformerImageText
🎯 What it does: A new Twicing Attention mechanism is proposed to replace traditional self-attention, enhancing the representational diversity of Transformers and mitigating the over-smoothing problem through residual self-correction.
Transformer-Squared: Self-adaptive LLMs
Qi Sun (Sakana AI), Yujin Tang (Sakana AI)
TransformerLarge Language ModelReinforcement LearningPrompt EngineeringMixture of ExpertsText
🎯 What it does: Proposes the Transformer 2 framework, utilizing Singular Value Fine-Tuning (SVF) to dynamically adapt large language models through two passes during inference;
Transformers are Universal In-context Learners
Takashi Furuya (Doshisha University), Gabriel Peyré (Centre National de la Recherche Scientifique)
Transformer
🎯 What it does: The research proves the universal approximation property of deep Transformers when handling context of arbitrary length.
Transformers Can Learn Temporal Difference Methods for In-Context Reinforcement Learning
Jiuqi Wang (University of Virginia), Shangtong Zhang (University of Virginia)
TransformerReinforcement LearningSequential
🎯 What it does: This paper studies whether the Transformer can achieve policy evaluation reinforcement learning (ICRL) through forward inference without updating parameters, and provides a formal white-box proof.
Transformers Handle Endogeneity in In-Context Linear Regression
Haodong Liang (University of California Davis), Lifeng Lai (University of California Davis)
TransformerTabular
🎯 What it does: This paper studies the capabilities of transformers in addressing endogeneity linear regression tasks, proving that it can implement the two-stage least squares (2SLS) algorithm through a looped transformer, achieving robust predictions and coefficient estimates under various endogeneity scenarios after pre-training.
Transformers Learn Low Sensitivity Functions: Investigations and Implications
Bhavya Vasudeva (University of Southern California), Vatsal Sharan (University of Southern California)
TransformerImageTextMultimodality
🎯 What it does: This paper studies the low sensitivity function learned by Transformers in multimodal tasks and extends the sensitivity metric to non-Bool data, exploring its impact on robustness, loss landscape flatness, and training dynamics (grokking).
Transformers Learn to Implement Multi-step Gradient Descent with Chain of Thought
Jianhao Huang (Shanghai Jiaotong University), Jason D. Lee (Princeton University)
OptimizationTransformerTabularChain-of-Thought
🎯 What it does: This paper studies the training dynamics of transformers under Chain of Thought (CoT) prompting, particularly their performance in context weight prediction tasks, demonstrating that transformers can achieve multi-step gradient descent (GD) through CoT.
Transformers Provably Learn Two-Mixture of Linear Classification via Gradient Flow
Hongru Yang (Princeton University), Yingbin Liang (Ohio State University)
ClassificationTransformerFlow-based ModelTabular
🎯 What it does: This study investigates the training dynamics of two-headed Transformers in binary mixed linear classification tasks and presents a three-stage training algorithm along with theoretical analysis.
Transformers Provably Solve Parity Efficiently with Chain of Thought
Juno Kim (University of Tokyo), Taiji Suzuki (University of Tokyo)
TransformerChain-of-Thought
🎯 What it does: This paper studies how to efficiently solve the k-parity problem by training a Transformer in a Chain-of-Thought (CoT) manner, providing theoretical analysis and experimental validation.
Transformers Struggle to Learn to Search
Abulhair Saparov (Purdue University), He He (New York University)
TransformerLarge Language ModelGraph
🎯 What it does: Investigate whether transformers can learn graph search tasks and train a small GPT-2 structured model by generating infinite DAG search samples.
Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal Model
Chunting Zhou (Meta), Omer Levy (Meta)
GenerationData SynthesisTransformerDiffusion modelImageTextMultimodality
🎯 What it does: A multi-modal model training method called Transfusion is proposed, which can simultaneously handle discrete and continuous data by combining the loss functions of language modeling and diffusion models.
Transition Path Sampling with Improved Off-Policy Training of Diffusion Path Samplers
Kiyoung Seong (Korea Advanced Institute of Science and Technology), Sungsoo Ahn (Korea Advanced Institute of Science and Technology)
Drug DiscoveryProtein Structure PredictionDiffusion modelSequentialBiomedical DataStochastic Differential Equation
🎯 What it does: A diffusion path sampler (TPS-DPS) is proposed that does not require collective variables to automatically sample transition paths of molecular systems from one ground state to another.
Tree of Attributes Prompt Learning for Vision-Language Models
Tong Ding (Harvard University), Hanspeter Pfister (Harvard University)
ClassificationRecognitionTransformerLarge Language ModelPrompt EngineeringVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: A Tree of Attributes Prompt Learning (TAP) is constructed, embedding structured knowledge into the prompt learning of VLM through a hierarchical attribute tree generated by LLM;
Tree-Wasserstein Distance for High Dimensional Data with a Latent Feature Hierarchy
Ya-Wei Eileen Lin (Technion), Ronen Talmon (Technion)
ClassificationDiffusion modelTextBiomedical Data
🎯 What it does: A new Tree-Wasserstein distance (TWD) is proposed to compute meaningful distances between samples in high-dimensional data with unknown feature hierarchies.
TRENDy: Temporal Regression of Effective Nonlinear Dynamics
Matt Ricci, Mor Nitzan (Hebrew University of Jerusalem)
VideoPhysics RelatedOrdinary Differential Equation
🎯 What it does: By mapping the spatial state of PDEs to low-dimensional features extracted from scattering transforms and fitting with parameterized neural ODEs, predictable parameterized effective dynamics were learned, achieving bifurcation localization in noisy environments.
Triples as the Key: Structuring Makes Decomposition and Verification Easier in LLM-based TableQA
Zhen Yang (Anhui University), Shu Zhao (Anhui University)
TransformerLarge Language ModelAgentic AIPrompt EngineeringTabular
🎯 What it does: A triplet-based splitting and verification strategy TIDE is proposed to improve the problem splitting, reasoning, and answer verification processes of LLM in the TableQA task.
Trivialized Momentum Facilitates Diffusion Generative Modeling on Lie Groups
Yuchen Zhu (Georgia Institute of Technology), Molei Tao (Georgia Institute of Technology)
GenerationData SynthesisDiffusion modelScore-based ModelBiomedical Data
🎯 What it does: A Lie group-based non-approximating diffusion generative model TDM is proposed, which transforms the manifold problem into score learning in Euclidean space using trivialized momentum, achieving efficient generation of manifold data.
Truncated Consistency Models
Sangyun Lee (Carnegie Mellon University), Weili Nie (NVIDIA)
GenerationData SynthesisDiffusion modelImageOrdinary Differential Equation
🎯 What it does: This paper proposes a Truncated Consistency Model (TCM) that significantly improves single-step generation quality by training the consistency model over shorter time intervals, focusing solely on the generation task from noise to data.
Trust or Escalate: LLM Judges with Provable Guarantees for Human Agreement
Jaehun Jung (University of Washington), Yejin Choi (University of Washington)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a 'Selective Evaluation' framework that uses confidence thresholds to determine whether to allow LLMs to evaluate samples, providing a provable guarantee of human consistency (human agreement rate); based on this, it introduces 'Simulated Annotators' to enhance confidence estimation and constructs a cascaded evaluation system to reduce evaluation costs.
Trusted Multi-View Classification via Evolutionary Multi-View Fusion
Xinyan Liang (Shanxi University), Guoqing Liu (Taiyuan University of Science and Technology)
ClassificationNeural Architecture SearchMultimodality
🎯 What it does: This paper proposes a method to enhance trustworthy multi-view classification (TEF) through evolutionary multi-view fusion, which automatically searches for high-quality pseudo-views and addresses the information imbalance problem through view enhancement.
TS-LIF: A Temporal Segment Spiking Neuron Network for Time Series Forecasting
FENG SHIBO, Zhiqi Shen (University of Chinese Academy of Sciences)
Spiking Neural NetworkTime Series
🎯 What it does: A dual-chamber pulse neuron model for time series prediction, TS-LIF, is proposed, utilizing two frequency processing mechanisms in dendrites and the soma to achieve multi-scale information extraction.
TSC-Net: Prediction of Pedestrian Trajectories by Trajectory-Scene-Cell Classification
BO HU, Tat-Jen Cham (Nanyang Technological University)
Object TrackingGenerationConvolutional Neural NetworkTransformerGenerative Adversarial NetworkVideoMultimodality
🎯 What it does: This paper proposes a trajectory-scene-unit (TSC) feature representation and constructs TSC-Net based on this feature, using unit classification and offset regression to achieve pedestrian future trajectory prediction.
TTVD: Towards a Geometric Framework for Test-Time Adaptation Based on Voronoi Diagram
Mingxi Lei (University at Buffalo), Jinhui Xu (University at Buffalo)
Domain AdaptationImage
🎯 What it does: A test-time adaptive framework TTVD based on Voronoi diagrams is proposed, achieving more robust feature alignment and noise filtering through Cluster-induced Voronoi Diagram and Power Diagram.
TULIP: Token-length Upgraded CLIP
Ivona Najdenkoska (University of Amsterdam), Cees G. M. Snoek (University of Amsterdam)
GenerationRetrievalKnowledge DistillationTransformerContrastive LearningImageTextMultimodality
🎯 What it does: By migrating the CLIP text encoder from absolute position encoding to relative position encoding (RoPE), support for long texts (over 77 words) is achieved, forming the TULIP model.
Tuning Frequency Bias of State Space Models
Annan Yu (Cornell University), N. Benjamin Erichson (Lawrence Berkeley National Laboratory)
RestorationData SynthesisRecurrent Neural NetworkSupervised Fine-TuningImageVideoBenchmark
🎯 What it does: This paper analyzes the frequency bias in the initialization and training processes of state space models (SSM) from the frequency domain perspective and proposes two mechanisms to adjust this bias: changing the learnable frequency range by scaling the eigenvalues (α) in the HiPPO initialization; and re-weighting frequencies during training using Sobolev-norm based filtering (β) to alter the gradient's sensitivity to high frequencies. These methods were experimentally validated on tasks such as image denoising, sCIFAR-10, and Long-Range Arena.
Tuning Timestep-Distilled Diffusion Model Using Pairwise Sample Optimization
Zichen Miao (Purdue University), Qiang Qiu (Purdue University)
GenerationOptimizationKnowledge DistillationDiffusion modelImage
🎯 What it does: A method based on Pairwise Sample Optimization (PSO) is proposed to directly fine-tune the time-slot distillation diffusion model for achieving human preferences, style transfer, and concept customization.
Tuning-Free Bilevel Optimization: New Algorithms and Convergence Analysis
Yifan Yang (University at Buffalo), Kaiyi Ji (Rice University)
OptimizationTabular
🎯 What it does: Two completely step-size parameter-free bilevel optimization algorithms, D-TFBO (double loop) and S-TFBO (single loop), are proposed, achieving adaptive step-size updates.
Turning Up the Heat: Min-p Sampling for Creative and Coherent LLM Outputs
Nguyen Nhat Minh (Apart Research), Ravid Shwartz-Ziv (Wand.ai)
GenerationTransformerLarge Language ModelTextChain-of-Thought
🎯 What it does: The minp sampling method is proposed, which dynamically adjusts the sampling threshold using model confidence, thereby balancing diversity and coherence in LLM text generation.
TVNet: A Novel Time Series Analysis Method Based on Dynamic Convolution and 3D-Variation
Chenghan Li (Tsinghua University), Ruisheng Diao (Zhejiang University)
ClassificationAnomaly DetectionOptimizationConvolutional Neural NetworkTransformerTime Series
🎯 What it does: A unified temporal analysis framework named TVNet based on 3D deformation and dynamic convolution is proposed, capable of achieving unified modeling across various tasks such as prediction, interpolation, classification, and anomaly detection.
TweedieMix: Improving Multi-Concept Fusion for Diffusion-based Image/Video Generation
Gihyun Kwon (Krafton), Jong Chul Ye (Kim Jaechul Graduate School of AI KAIST)
GenerationData SynthesisDiffusion modelImageVideo
🎯 What it does: This paper proposes a training-free two-stage sampling framework (TweedieMix) for synthesizing multi-concept images and videos during the inference phase of diffusion models.
Two Effects, One Trigger: On the Modality Gap, Object Bias, and Information Imbalance in Contrastive Vision-Language Models
Simon Schrodi (University of Freiburg), Thomas Brox (University of Freiburg)
ClassificationRetrievalVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: Analyze and compare the relationship between modality gap, target bias, and information imbalance in contrastive vision models, and verify that information imbalance is the fundamental cause of both;
Two Sparse Matrices are Better than One: Sparsifying Neural Networks with Double Sparse Factorization
Vladimír Boža (Comenius University), Vladimír Macko (Comenius University)
CompressionOptimizationConvolutional Neural NetworkTransformerLarge Language ModelImageText
🎯 What it does: The Double Sparse Factorization (DSF) algorithm is proposed, which decomposes the weight matrix into two sparse matrices for a round of hierarchical sparsification;
TypedThinker: Diversify Large Language Model Reasoning with Typed Thinking
Danqing Wang (Carnegie Mellon University), Lei Li (Qwen Team)
TransformerLarge Language ModelTextRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: Designed and implemented the TypedThinker framework, which utilizes a meta-thinker to predict suitable reasoning types and diversifies the reasoning process of LLM through explicit demonstration retrieval.
u-$\mu$P: The Unit-Scaled Maximal Update Parametrization
Charlie Blake (Graphcore), Douglas Orr (Cohere)
OptimizationHyperparameter SearchTransformerLarge Language ModelText
🎯 What it does: A new scheme called u-µP is proposed, which combines Maximum Update Parameterization (µP) with Unit Scaling, and based on this, interpretable and mutually independent hyperparameters are designed to support simpler independent searches.
U-Nets as Belief Propagation: Efficient Classification, Denoising, and Diffusion in Generative Hierarchical Models
Song Mei (University of California, Berkeley)
ClassificationRestorationConvolutional Neural NetworkDiffusion modelImage
🎯 What it does: This paper explains the structure of U-Net, including its encoding-decoding, skip connections, pooling, and upsampling, by viewing it as a Bayesian inference algorithm (belief propagation) in a Generative Hierarchical Model (GHM), and provides an upper bound on the sample complexity for classification and denoising tasks under this model.
U-shaped and Inverted-U Scaling behind Emergent Abilities of Large Language Models
Tung-Yu Wu (National Taiwan University), Melody Lo (National Taiwan University)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the phenomenon of generative capabilities of large language models in multiple-choice tasks, finding that hard questions exhibit a U-shaped curve while easy questions show an inverted U-shaped curve. It explains the reasons for the performance stagnation followed by a sudden increase and proposes the Slice-and-Sandwich prediction pipeline.
UGMathBench: A Diverse and Dynamic Benchmark for Undergraduate-Level Mathematical Reasoning with Large Language Models
Xin Xu (Hong Kong University of Science and Technology), Can Yang (Hong Kong University of Science and Technology)
TransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: This paper presents UGMathBench, a dynamic benchmark containing 5,062 undergraduate-level math problems, covering 16 subjects, 111 topics, and 10 types of answers, with three random versions for each problem.
UIFace: Unleashing Inherent Model Capabilities to Enhance Intra-Class Diversity in Synthetic Face Recognition
Xiao Lin (Tencent Youtu Lab), Shouhong Ding (Tencent Youtu Lab)
RecognitionGenerationData SynthesisDiffusion modelImage
🎯 What it does: Using a diffusion model for two-stage sampling and attention injection, we generate synthetic faces that retain identity information while exhibiting high intra-class diversity, aimed at training face recognition models.
Ultra-Sparse Memory Network
Zihao Huang (ByteDance), zhou Xun
TransformerMixture of ExpertsTextBenchmark
🎯 What it does: Introducing UltraMem—a large-scale ultra-sparse memory layer designed to replace MoE, reducing memory access overhead and latency during inference.
Unbounded: A Generative Infinite Game of Character Life Simulation
Jialu Li (Google), Nataniel Ruiz (Google)
GenerationData SynthesisKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringDiffusion modelImageText
🎯 What it does: Developed a real-time interactive infinite generation game UNBOUNDED, where players can customize characters and engage in open-ended interactions with characters and the environment using natural language. All game actions, plots, and visual content are generated instantaneously by generative models.
Uncertainty and Influence aware Reward Model Refinement for Reinforcement Learning from Human Feedback
Zexu Sun (Renmin University of China), Ji-Rong Wen
Reinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a data augmentation method based on uncertainty gradients, called UGDA, to adaptively correct the reward model in RLHF and alleviate the issue of discrete distribution shift.
Uncertainty Herding: One Active Learning Method for All Label Budgets
Wonho Bae (University of British Columbia), Gabriel L. Oliveira (Borealis AI)
ClassificationDomain AdaptationData-Centric LearningConvolutional Neural NetworkContrastive LearningImage
🎯 What it does: A unified active learning method called Uncertainty Herding (UHerding) is proposed, which integrates low-budget and high-budget approaches by utilizing Uncertainty Coverage and achieving smooth interpolation through parameter adaptation.
Uncertainty modeling for fine-tuned implicit functions
Anna Susmelj (ETH AI Center), Ender Konukoglu (ETH Zurich)
Supervised Fine-TuningPoint CloudMeshBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper proposes a framework for sparse 3D shape reconstruction and uncertainty estimation based on Dropsembles, which uses high-quality synthetic data as prior information and fine-tunes implicit functions under sparse noisy input.
Uncertainty Modeling in Graph Neural Networks via Stochastic Differential Equations
Richard Bergna (University of Cambridge), José Miguel Hernández-Lobato (University of Oxford)
Graph Neural NetworkGraphStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: Introducing stochastic differential equations (SDE) in graph neural networks, the Latent Graph Neural Stochastic Differential Equations (LGNSDE) model is proposed for learning graph embeddings with uncertainty.
Uncertainty-Aware Decoding with Minimum Bayes Risk
Nico Daheim (Technical University of Darmstadt), Iryna Gurevych (Technical University of Darmstadt)
GenerationOptimizationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a language model decoding method that considers weight uncertainty by integrating parameter posteriors into minimum Bayes risk (MBR) decoding, reducing hallucinations and errors in generation, and achieving performance improvements without additional inference overhead.
Uncovering Gaps in How Humans and LLMs Interpret Subjective Language
Erik Jones (University of California Berkeley), Jacob Steinhardt (University of California Berkeley)
TransformerLarge Language ModelText
🎯 What it does: This study investigates the discrepancies between the operational semantics of LLMs when processing subjective natural language and human expectations, and proposes a dictionary comparison-based error detection method.
Uncovering Latent Memories in Large Language Models
Sunny Duan (Massachusetts Institute of Technology), Ila R Fiete
TransformerLarge Language ModelText
🎯 What it does: The study investigates the memory behavior of large language models during the pre-training process for sequences that appear only once and have high complexity, discovering the existence of 'latent memory' that can be recovered through noise perturbation.
Uncovering Overfitting in Large Language Model Editing
Mengqi Zhang (Shandong University), Zhumin Chen (Shandong University)
TransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This study investigates the overfitting phenomenon that occurs in knowledge editing of large language models and proposes a plugin strategy of multi-stage reasoning constraints to alleviate this issue.
Underdamped Diffusion Bridges with Applications to Sampling
Denis Blessing (Karlsruhe Institute of Technology), Gerhard Neumann (FZI Research Center for Information Technology)
Diffusion modelTabularStochastic Differential Equation
🎯 What it does: A framework based on underdamped diffusion bridges is proposed, achieving finite-time convergence from prior to target distribution.
Understanding and Enhancing Safety Mechanisms of LLMs via Safety-Specific Neuron
Yiran Zhao (National University of Singapore), Michael Shieh
Safty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes a method for the detection and tuning of safe neurons, identifying safe neurons that account for less than 1% of the parameters and enhancing the safety of LLMs by updating only these neurons.
Understanding and Enhancing the Transferability of Jailbreaking Attacks
Runqi Lin (Sydney AI Centre, University of Sydney), Tongliang Liu (Sydney AI Centre, University of Sydney)
Adversarial AttackTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies the transferability of jailbreak attacks from the perspective of intent perception in large language models and proposes an attack method based on Intent Importance Flattening (PiF), significantly enhancing the destructive effect on proprietary models.
Understanding and Mitigating Bottlenecks of State Space Models through the Lens of Recency and Over-smoothing
Peihao Wang (University of Texas at Austin), Pan Li (Georgia Tech)
RetrievalOptimizationTransformerImageTextBenchmark
🎯 What it does: The bottlenecks of the Structured State Space Model (SSM) were studied, revealing the presence of asymptotic bias and over-smoothing issues, and a polarization technique was proposed to address them.
Understanding and Mitigating Hallucination in Large Vision-Language Models via Modular Attribution and Intervention
Tianyun Yang (Institute of Computing Technology, Chinese Academy of Sciences), Chang Xu (University of Sydney)
GenerationExplainability and InterpretabilityTransformerVision Language ModelImageMultimodality
🎯 What it does: A causal attribution and intervention study on the hallucination problem of large visual language models, identifying specific heads in multi-head attention that cause hallucinations, and proposing two methods—deactivation and fine-tuning during decoding—to mitigate hallucinations.
Understanding Constraint Inference in Safety-Critical Inverse Reinforcement Learning
Bo Yue (Chinese University of Hong Kong), Guiliang Liu (Chinese University of Hong Kong)
OptimizationSafty and PrivacyReinforcement LearningSequential
🎯 What it does: This paper compares two methods in safety-critical inverse reinforcement learning: Inverse Reward Correction (IRC) with implicit constraint embedding and Inverse Constraint Reinforcement Learning (ICRL) with explicit constraint inference. It provides an analysis of theoretical sample complexity and cross-environment transfer, and validates the results through grid world experiments.