ICML 2025 Papers — Page 29
International Conference on Machine Learning · 3257 papers
Sundial: A Family of Highly Capable Time Series Foundation Models
Yong Liu (Tsinghua University), Mingsheng Long (Tsinghua University)
TransformerFlow-based ModelTime SeriesFinance Related
🎯 What it does: A family of foundational models for time series, called Sundial, is proposed based on flow matching, which can perform unsupervised pre-training directly on a continuous value domain without discretization and supports diverse generative predictions.
Super Deep Contrastive Information Bottleneck for Multi-modal Clustering
Zhengzheng Lou (Zhengzhou University), Shizhe Hu (Zhengzhou University)
Representation LearningContrastive LearningImageMultimodality
🎯 What it does: This study investigates multimodal clustering and proposes a model based on Super Deep Contrastive Information Bottleneck (SDCIB) that can fully exploit the implicit information in multimodal hidden layers and enhance clustering performance.
Supercharging Graph Transformers with Advective Diffusion
Qitian Wu (Broad Institute of MIT and Harvard), Michael M. Bronstein
Graph Neural NetworkTransformerGraphPhysics RelatedStochastic Differential Equation
🎯 What it does: This paper proposes a graph Transformer based on the advective diffusion equation—Advective Diffusion Transformer (ADVDIFFORMER)—and theoretically proves its ability to control the generalization error caused by topological distribution shifts.
Supervised Contrastive Learning from Weakly-Labeled Audio Segments for Musical Version Matching
Joan Serrà (Sony Group Corporation), Yuki Mitsufuji (Sony Group Corporation)
RetrievalRepresentation LearningConvolutional Neural NetworkContrastive LearningAudio
🎯 What it does: This paper proposes a supervised contrastive learning method based on weakly labeled audio segments for music version matching.
Surrogate Prompt Learning: Towards Efficient and Diverse Prompt Learning for Vision-Language Models
Liangchen Liu (Xidian University), Tongliang Liu (University of Sydney)
ClassificationRecognitionComputational EfficiencyTransformerPrompt EngineeringVision Language ModelImageMultimodality
🎯 What it does: A Surrogate Prompt Learning (SurPL) framework is proposed, which achieves efficient and diverse prompt learning by generating surrogate text features.
Survival Analysis via Density Estimation
Hiroki Yanagisawa (CyberAgent), Shunta Akiyama (CyberAgent)
TabularTime Series
🎯 What it does: Reformulates survival analysis as a density estimation problem and proposes a two-step algorithm: first, estimate the cumulative hazard function using any density estimation model, and then obtain the individual survival function through post-processing with a known copula.
SWAN: SGD with Normalization and Whitening Enables Stateless LLM Training
Chao Ma (Microsoft Research), Edward Meeds (Microsoft Research)
OptimizationTransformerLarge Language ModelText
🎯 What it does: A stateless optimizer called SWAN is proposed, which utilizes gradient normalization and whitening preprocessing to replace the stateful adaptive updates of traditional Adam, significantly reducing memory usage during the training of large language models.
SWE-Lancer: Can Frontier LLMs Earn $1 Million from Real-World Freelance Software Engineering?
Samuel Miserendino (OpenAI), Johannes Heidecke (OpenAI)
AI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: The SWE-Lancer benchmark is proposed to evaluate the capabilities of large language models in real-world freelance software engineering tasks.
Symmetric Reinforcement Learning Loss for Robust Learning on Diverse Tasks and Model Scales
Ju-Seung Byun (Ohio State University), Andrew Perrault (Ohio State University)
Reinforcement Learning from Human FeedbackReinforcement LearningText
🎯 What it does: A symmetric RL loss (Symmetric RL Loss) for A2C and PPO is proposed to enhance learning robustness in scenarios with reward noise and uncertainty in advantage prediction.
Symmetry-Aware GFlowNets
Hohyun Kim (Seoul National University), Min-hwan Oh (Seoul National University)
GenerationData SynthesisOptimizationGraph Neural NetworkReinforcement LearningGraph
🎯 What it does: In the graph generation task, the researchers proposed Symmetry-Aware GFlowNets (SA-GFN), which corrects the bias of GFlowNet in handling equivalent actions by incorporating the number of graph self-isomorphisms into the reward;
Symmetry-Driven Discovery of Dynamical Variables in Molecular Simulations
Jeet Mohapatra (Massachusetts Institute of Technology), Tommi Jaakkola (Massachusetts Institute of Technology)
OptimizationBiomedical DataPhysics Related
🎯 What it does: An effective degree of freedom (DOF) discovery method based solely on force field energy information is proposed, utilizing energy landscape approximate symmetry.
Symmetry-Robust 3D Orientation Estimation
Christopher Scarvelis (Massachusetts Institute of Technology), Paul Zhang (Backflip AI)
Pose EstimationGraph Neural NetworkPoint Cloud
🎯 What it does: A two-stage 3D shape orientation estimation pipeline is proposed, where the shape is first rotated to octahedral symmetry through regression, and then the final complete orientation to the coordinate system is determined using classification.
SyncMind: Measuring Agent Out-of-Sync Recovery in Collaborative Software Engineering
Xuehang Guo (University of Illinois Urbana-Champaign), Heng Ji (University of Illinois Urbana-Champaign)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes the SyncMind framework and the SyncBench benchmark for system evaluation of large language models (LLMs) in handling recovery capabilities in 'asynchronous' scenarios within collaborative software engineering.
SynEVO: A neuro-inspired spatiotemporal evolutional framework for cross-domain adaptation
Jiayue Liu (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
Domain AdaptationGraph Neural NetworkContrastive LearningTime Series
🎯 What it does: This paper proposes a neuroscience-based evolvable spatiotemporal network called SynEVO, which utilizes curriculum learning-style sample rearrangement, elastic public containers, task-agnostic feature extractors, and adaptive dynamic couplers to achieve continuous learning and knowledge transfer for cross-domain spatiotemporal prediction.
Synonymous Variational Inference for Perceptual Image Compression
Zijian Liang (Beijing University of Posts and Telecommunications), Ping Zhang (Beijing University of Posts and Telecommunications)
CompressionOptimizationTransformerGenerative Adversarial NetworkImage
🎯 What it does: This paper proposes a Synonym Variational Inference (SVI) method based on the semantic information theory of synonymity, conducts a theoretical analysis of the optimization direction for perceptual image compression, and designs a Synonym Image Compression (SIC) framework based on this analysis, further implementing a multi-level evolutionary encoder.
Synthesizing Images on Perceptual Boundaries of ANNs for Uncovering and Manipulating Human Perceptual Variability
Chen Wei (Southern University of Science and Technology), Quanying Liu (Southern University of Science and Technology)
ClassificationGenerationData SynthesisSupervised Fine-TuningDiffusion modelImage
🎯 What it does: By sampling images on the ANN perception boundary and combining them with human experiments, the variMNIST dataset is constructed to study and implement the prediction and manipulation of individualized perceptual differences.
Synthesizing Privacy-Preserving Text Data via Finetuning *without* Finetuning Billion-Scale LLMs
Bowen Tan (Carnegie Mellon University), Shanshan Wu (Google Research)
GenerationData SynthesisSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: A framework named CTCL is proposed, which pre-trains a 140M parameter conditional generator and a general topic model on public corpora, and then fine-tunes with differential privacy (DP) on private data and constructs a topic histogram, thereby generating privacy-preserving synthetic text data without the need for fine-tuning on trillion-parameter LLMs or manual prompt engineering.
Synthesizing Software Engineering Data in a Test-Driven Manner
Lei Zhang (Shenzhen Institutes of Advanced Technology), Junyang Lin (Alibaba Group)
Data SynthesisAI Code AssistantLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Proposes the SWE-Flow framework, which automatically constructs runtime dependency graphs using unit tests, generates verifiable incremental software engineering training data, and builds the SWE-Flow-Bench benchmark;
Synthetic Face Datasets Generation via Latent Space Exploration from Brownian Identity Diffusion
David Geissbühler (Idiap Research Institute), Sébastien Marcel (Idiap Research Institute)
RecognitionGenerationData SynthesisGenerative Adversarial NetworkImageStochastic Differential Equation
🎯 What it does: This paper proposes a three-stage sampling framework based on Brownian motion and particle dynamics to generate a large-scale synthetic face dataset in the GAN latent space.
Synthetic Text Generation for Training Large Language Models via Gradient Matching
Dang Nguyen (University of California), Baharan Mirzasoleiman (University of California)
GenerationData SynthesisOptimizationSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: A theoretically rigorous algorithm is proposed to generate readable, privacy-preserving synthetic text for fine-tuning large language models.
System-Aware Unlearning Algorithms: Use Lesser, Forget Faster
Linda Lu (Cornell University), Karthik Sridharan (Cornell University)
🎯 What it does: A system-aware machine forgetting algorithm is proposed, along with the corresponding definitions.
T1: Advancing Language Model Reasoning through Reinforcement Learning and Inference Scaling
Zhenyu Hou (Tsinghua University), Yuxiao Dong (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: This paper proposes the T1 framework, which first conducts chain-of-thought (CoT) pre-training with trial-and-error and self-checking on a large model, and then further enhances reasoning capabilities using reinforcement learning (RL);
TabFlex: Scaling Tabular Learning to Millions with Linear Attention
Yuchen Zeng (University of Wisconsin Madison), Andreas C Mueller
ClassificationComputational EfficiencyTabular
🎯 What it does: Improved TABPFN to achieve efficient in-context learning on tabular data with millions of samples and thousands of features.
TabFSBench: Tabular Benchmark for Feature Shifts in Open Environments
Zi-Jian Cheng (Nanjing University), Lan-Zhe Guo (Nanjing University)
ClassificationDomain AdaptationData-Centric LearningLarge Language ModelTabularBenchmarkFinance Related
🎯 What it does: TabFSBench is proposed - the first benchmark for feature shift in tabular data, systematically studying the impact of feature shift on model performance and evaluating multiple models.
TabICL: A Tabular Foundation Model for In-Context Learning on Large Data
Jingang QU, Marine Le Morvan (INRIA Saclay)
ClassificationTransformerTabular
🎯 What it does: A scalable table-based model TabICL is proposed, utilizing a two-stage column-row attention mechanism to achieve single forward inference for large-scale tables;
TabNAT: A Continuous-Discrete Joint Generative Framework for Tabular Data
Hengrui Zhang (University of Illinois at Chicago), Philip S. Yu (University of Illinois at Chicago)
GenerationData SynthesisTransformerDiffusion modelTabular
🎯 What it does: This paper presents TabNAT, a continuous-discrete hybrid table data generation framework that combines conditional diffusion models with masked bidirectional Transformers.
TabPFN Unleashed: A Scalable and Effective Solution to Tabular Classification Problems
Siyang Liu, Han-Jia Ye (Nanjing University)
ClassificationOptimizationSupervised Fine-TuningTabular
🎯 What it does: The BETA method is proposed, which enhances the performance of TabPFN on high-dimensional, large-scale, multi-class tasks through lightweight encoder fine-tuning and bootstrapped sampling.
TabSDS: a Lightweight, Fully Non-Parametric, and Model Free Approach for Generating Synthetic Tabular Data
Elias Chaibub Neto (Sage Bionetworks)
Data SynthesisSafty and PrivacyTabular
🎯 What it does: A lightweight, model-free, and non-parametric synthetic tabular data generation method called TabSDS is proposed, which approximates the joint distribution with the original data through ranking and data shuffling.
Tackling Dimensional Collapse toward Comprehensive Universal Domain Adaptation
Hung-Chieh Fang (National Taiwan University), Hsuan-Tien Lin (National Taiwan University)
Domain AdaptationContrastive LearningImage
🎯 What it does: This paper studies the failure of traditional Partial Domain Matching (PDM) methods in extreme Universal Domain Adaptation (UniDA) scenarios, identifying the fundamental cause as dimensional collapse (DC) of target domain features.
Tackling View-Dependent Semantics in 3D Language Gaussian Splatting
Jiazhong Cen (Shanghai Jiao Tong University), Qi Tian (Huawei Technologies Company)
Object DetectionSegmentationContrastive LearningGaussian SplattingPoint Cloud
🎯 What it does: An open-source vocabulary scene understanding method based on 3D Gaussian Splatting, called LaGa, is proposed to address the perspective-dependent semantic issues.
Taming Diffusion for Dataset Distillation with High Representativeness
Lin Zhao (Northeastern University), Xue Lin (Northeastern University)
Data SynthesisKnowledge DistillationConvolutional Neural NetworkDiffusion modelAuto EncoderImage
🎯 What it does: A dataset distillation framework D3HR based on DDIM reverse mapping is proposed, which maps the VAE latent space to a high-normality noise domain and samples representative latent variables in that domain to generate a compressed dataset.
Taming Knowledge Conflicts in Language Models
Gaotang Li (University of Illinois), Hanghang Tong (University of Illinois)
TransformerLarge Language ModelText
🎯 What it does: This study conducts a systematic analysis of knowledge conflicts in language models, revealing the phenomenon of the superposition of contextual information and parameter memory, and proposes a dual-round testing intervention method called JUICE that does not require fine-tuning.
Taming Rectified Flow for Inversion and Editing
Jiangshan Wang (Tsinghua University), Ying Shan (ARC Lab, Tencent PCG)
Diffusion modelRectified FlowImageVideoOrdinary Differential Equation
🎯 What it does: Designed RF-Solver to enhance the inversion and sampling accuracy of the Rectified Flow model, and based on it, proposed RF-Edit for high-quality image and video editing.
TANGO: Clustering with Typicality-Aware Nonlocal Mode-Seeking and Graph-Cut Optimization
Haowen Ma (Southwest Jiaotong University), Hua Meng (Southwest Jiaotong University)
OptimizationImageTabular
🎯 What it does: A new clustering framework called TANGO is proposed, which achieves threshold-free pattern recognition through a typicality index and uses graph cuts to merge subclusters.
Target Concrete Score Matching: A Holistic Framework for Discrete Diffusion
Ruixiang ZHANG, Navdeep Jaitly (Apple)
Knowledge DistillationDiffusion modelScore-based ModelText
🎯 What it does: Proposes Target Concrete Score Matching (TCSM), a unified framework for target discrete diffusion training and fine-tuning;
Targeted control of fast prototyping through domain-specific interface
Yu-Zhe Shi (Peking University), Qining Wang (Peking University)
Large Language ModelText
🎯 What it does: This paper proposes an interface architecture based on a domain-specific DSL, which converts the natural language intentions of industrial designers into low-level modeling commands through a large language model, achieving precise control for rapid prototyping.
Targeted Low-rank Refinement: Enhancing Sparse Language Models with Precision
Li Shen (Shenzhen Campus of Sun Yat-sen University), Xiaochun Cao
TransformerLarge Language ModelText
🎯 What it does: This paper proposes an iterative low-rank correction method to restore performance after pruning large language models, maintaining a sparse pattern and compensating for the information that has been pruned.
Targeted Unlearning with Single Layer Unlearning Gradient
Zikui Cai (University of California Riverside), M. Salman Asif (University of California Riverside)
Computational EfficiencyData-Centric LearningVision Language ModelDiffusion modelImageMultimodality
🎯 What it does: This paper proposes a Single-Layer Unlearning Gradient (SLUG) method that achieves target information unlearning in large-scale multimodal models using a single gradient update.
TAROT: Targeted Data Selection via Optimal Transport
Lan Feng (École Polytechnique Fédérale de Lausanne), Alexandre Alahi (École Polytechnique Fédérale de Lausanne)
SegmentationDomain AdaptationAutonomous DrivingData-Centric LearningImageVideo
🎯 What it does: This paper presents TAROT, a target data selection framework based on optimal transport theory, designed to select a subset of data from a candidate pool that matches the target distribution.
Task Generalization with Autoregressive Compositional Structure: Can Learning from $D$ Tasks Generalize to $D^T$ Tasks?
Amirhesam Abedsoltan (University of California San Diego), Mikhail Belkin (University of California San Diego)
TransformerTextChain-of-Thought
🎯 What it does: This study investigates and verifies that under the autoregressive compositional structure, training a small number (≈O(D)) of tasks can achieve global generalization for an exponential family of tasks D^T; by introducing Chain-of-Thought (CoT), the generalization ability of Transformers on synthetic tasks is further enhanced.
Task-Agnostic Pre-training and Task-Guided Fine-tuning for Versatile Diffusion Planner
Chenyou Fan (Northwestern Polytechnical University), Zhen Wang (Northwestern Polytechnical University)
Robotic IntelligenceReinforcement LearningDiffusion modelMultimodality
🎯 What it does: This paper proposes a multi-task planner SODP based on sub-optimal data with unconditional diffusion pre-training and reward-guided fine-tuning.
Task-Aware Virtual Training: Enhancing Generalization in Meta-Reinforcement Learning for Out-of-Distribution Tasks
Jeongmo Kim (Ulsan National Institute of Science and Technology), Seungyul Han (Ulsan National Institute of Science and Technology)
Meta LearningReinforcement LearningGenerative Adversarial Network
🎯 What it does: The Task-Aware Virtual Training (TAVT) algorithm is proposed in meta reinforcement learning, utilizing virtual tasks to enhance generalization capabilities for out-of-distribution (OOD) tasks.
Task-Gated Multi-Expert Collaboration Network for Degraded Multi-Modal Image Fusion
Yiming Sun (Southeast University), Xinzhong Zhu (Zhejiang Normal University)
RestorationTransformerMixture of ExpertsImageBenchmark
🎯 What it does: A unified multi-task network TG-ECNet is proposed, capable of simultaneously performing denoising/deblurring/dehazing and fusion tasks for multi-modal images, addressing the issues of information loss and conflict in the traditional two-stage process of recovery followed by fusion.
TCP-Diffusion: A Multi-modal Diffusion Model for Global Tropical Cyclone Precipitation Forecasting with Change Awareness
Cheng Huang (Zhejiang University of Technology), Peter AG Watson
Diffusion modelMultimodalityTime Series
🎯 What it does: A multi-modal diffusion model, TCP-Diffusion, is proposed to predict precipitation around the center of tropical cyclones at any location globally for the next 12 hours.
Teaching Language Models to Critique via Reinforcement Learning
Zhihui Xie (University of Hong Kong), Lingpeng Kong (University of Hong Kong)
AI Code AssistantReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: Train large language models as critics to generate actionable feedback on code generation outputs through reinforcement learning, achieving iterative improvements of the model.
Teaching Physical Awareness to LLMs through Sounds
Weiguo Wang (NIO), Chengchen Hu (Peking University)
TransformerLarge Language ModelSupervised Fine-TuningPhysics RelatedAudio
🎯 What it does: By using sound to enable LLMs to acquire physical awareness, the ACORN framework is proposed, which includes a physical channel simulator, an audio encoder, and the AQA-PHY dataset. It trains LLMs to perform tasks such as LOS detection, multipath analysis, Doppler estimation, DOA estimation, and distance estimation.
Teaching Transformers Causal Reasoning through Axiomatic Training
Aniket Vashishtha (Microsoft Research India), Amit Sharma (Microsoft Research)
TransformerLarge Language ModelSupervised Fine-TuningTextGraph
🎯 What it does: Through 'axiomatic training', the Transformer model learns causal reasoning abilities from symbolic examples of causal axioms and transfers the trained model to more complex causal graphs and real-world causal reasoning benchmarks.
TeDS: Joint Learning of Diachronic and Synchronic Perspectives in Quaternion Space for Temporal Knowledge Graph Completion
Jiujiang Guo (Tianjin University), Yu Ruiguo
GraphTime Series
🎯 What it does: The TeDS model is proposed, which performs dual spatiotemporal perception of time knowledge graphs through quaternion embedding to complete knowledge graph completion.
Telling Peer Direct Effects from Indirect Effects in Observational Network Data
Xiaojing Du (University of South Australia), Ziqi Xu (RMIT University)
Recommendation SystemGraph Neural NetworkGraph
🎯 What it does: A framework and algorithm called gDIS is proposed to distinguish and estimate peer direct effects (PDE), peer indirect effects (PIE), and self-treatment effects (STE) in observational network data.
TeLoGraF: Temporal Logic Planning via Graph-encoded Flow Matching
Yue Meng (Massachusetts Institute of Technology), Chuchu Fan (Massachusetts Institute of Technology)
OptimizationRobotic IntelligenceGraph Neural NetworkReinforcement LearningFlow-based ModelGraphSequential
🎯 What it does: A framework called TeLoGraF is proposed, which utilizes graph neural networks to encode Signal Temporal Logic (STL) syntax and generates trajectories through flow matching to satisfy general STL specifications.
Temperature-Annealed Boltzmann Generators
Henrik Schopmans (Karlsruhe Institute of Technology), Pascal Friederich (Karlsruhe Institute of Technology)
GenerationOptimizationDrug DiscoveryFlow-based ModelSequentialBiomedical Data
🎯 What it does: Train a reversible flow network by first fitting the Boltzmann distribution at high temperatures using reverse KL divergence, and then recursively cooling down through importance sampling and fine-tuning with forward KL to achieve molecular sampling without mode collapse.
Temporal Difference Flows
Jesse Farebrother (McGill University), Ahmed Touati (Meta)
GenerationReinforcement LearningFlow-based ModelTime SeriesSequentialOrdinary Differential Equation
🎯 What it does: A generative model based on temporal difference flow is proposed to learn the Geometric Horizon Model (GHM) for accurately modeling future state distributions in long-term predictions.
Temporal Distance-aware Transition Augmentation for Offline Model-based Reinforcement Learning
Dongsu Lee (Carnegie Mellon University), Minhae Kwon (Soongsil University)
Reinforcement LearningAuto EncoderSequential
🎯 What it does: This paper proposes a transfer-enhanced offline model-based reinforcement learning framework called TempDATA, designed to address the performance degradation issues associated with sparse rewards and long-horizon tasks.
Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons
Velibor Bojkovic (Mohamed bin Zayed University of Artificial Intelligence), Bin Gu (Jilin University)
ClassificationSpiking Neural NetworkImage
🎯 What it does: This study investigates the phenomenon of temporal misalignment that occurs during the ANN-SNN conversion process and proposes a Two-Phase Probabilistic Spiking Neuron (TPP) to achieve temporal rearrangement, thereby enhancing the classification performance of the converted SNN.
Temporal Query Network for Efficient Multivariate Time Series Forecasting
Shengsheng Lin (South China University of Technology), Weiwei Lin (South China University of Technology)
TransformerTime Series
🎯 What it does: A novel Temporal Query (TQ) technique is proposed, and based on this, a lightweight TQNet is constructed for multivariate time series forecasting.
Tensor Decomposition Based Memory-Efficient Incremental Learning
Yuhang Li (Guangdong University of Technology), Qibin Zhao (RIKEN AIP)
CompressionKnowledge DistillationImage
🎯 What it does: A tensor compression method based on CP decomposition is proposed for replay memory in incremental learning, significantly reducing storage costs while maintaining discriminative information.
Tensor Product Neural Networks for Functional ANOVA Model
Seokhun Park (Seoul National University), Yongdai Kim (Seoul National University)
Tabular
🎯 What it does: The ANOVA-TPNN (Tensor Product Neural Network) model is proposed for stable estimation of the components of the functional ANOVA model, achieving component uniqueness through the sum-to-zero constraint.
Tensor-Var: Efficient Four-Dimensional Variational Data Assimilation
Yiming Yang (University College London), Yukun Hu (University College London)
OptimizationAuto EncoderTime Series
🎯 What it does: This paper proposes the Tensor-Var framework, which maps four-dimensional variational data assimilation (4D-Var) to a kernel feature space. It achieves the linearization of nonlinear dynamics and observation models through conditional mean embedding, thereby transforming the originally non-convex optimization into convex optimization. Furthermore, deep feature learning is employed in this feature space to enhance scalability.
Tensorized Multi-View Multi-Label Classification via Laplace Tensor Rank
Qiyu Zhong (Beijing University of Technology), Gengyu Lyu (Beijing University of Technology)
ClassificationOptimizationMultimodality
🎯 What it does: A tensor-based multi-view multi-label classification method TMvML is proposed, which utilizes low-rank tensor constraints to simultaneously mine cross-view feature correlations and label semantic relationships.
Test-Time Adaptation for Online Vision-Language Navigation with Feedback-based Reinforcement Learning
Sungjune Kim, Sangpil Kim (Korea University)
Large Language ModelReinforcement LearningMultimodality
🎯 What it does: Proposes the FEEDTTA framework, which utilizes binary feedback at termination as a reward to achieve online visual-language navigation adaptation during testing through REINFORCE reinforcement learning.
Test-time Adaptation on Graphs via Adaptive Subgraph-based Selection and Regularized Prototypes
Yusheng Zhao (Peking University), Ming Zhang (Peking University)
Domain AdaptationGraph Neural NetworkGraph
🎯 What it does: A method called ASSESS is proposed for adaptive tuning of graph data during testing without the need to access training data.
Test-Time Adaptation with Binary Feedback
Taeckyung Lee (KAIST), Sung-Ju Lee (KAIST)
Domain AdaptationReinforcement LearningImage
🎯 What it does: This paper proposes a testing-time adaptive method called BiTTA, which requires only a small amount of binary feedback to improve model performance in extreme domain shift environments.
Test-time Adapted Reinforcement Learning with Action Entropy Regularization
Shoukai Xu (South China University of Technology), Peilin Zhao (Tencent AI Lab)
Reinforcement LearningTabularBenchmark
🎯 What it does: Test-Time Adapted Reinforcement Learning (TARL) is proposed based on offline reinforcement learning. It fine-tunes the parameters of normalization layers using observed states only during testing, through unsupervised entropy minimization and KL constraints, to adapt to distribution shifts in online environments.
Test-Time Canonicalization by Foundation Models for Robust Perception
Utkarsh Singhal (University of California Berkeley), Atul Prakash (University of Michigan)
ClassificationImage TranslationSegmentationDiffusion modelImage
🎯 What it does: The FOCAL framework is proposed, which transforms the input image into the most typical view during the inference phase by generating and ranking various transformations, utilizing the visual priors of foundational models (CLIP, Stable Diffusion) to achieve robustness against multiple visual transformations.
Test-time Correlation Alignment
Linjing You (Institute of Automation, Chinese Academy of Sciences), Xiayuan Huang (Beijing Forestry University)
Domain AdaptationImage
🎯 What it does: Proposes the Test-time Correlation Alignment (TCA) method, which constructs a pseudo-source correlation matrix using high-confidence test samples, achieving feature covariance alignment through linear transformation without updating model parameters, thus enabling adaptation during testing.
Test-Time Graph Neural Dataset Search With Generative Projection
Xin Zheng (Griffith University), Shirui Pan (Griffith University)
ClassificationDomain AdaptationDrug DiscoveryGraph Neural NetworkDiffusion modelScore-based ModelGraphBiomedical Data
🎯 What it does: To address the issue of distribution drift in graph neural networks during testing, a method called 'Test-Time Graph Neural Dataset Search' is proposed, which uses a diffusion model to project the unknown test graph distribution back to the training distribution, thereby enhancing the inference performance of the pre-trained GNN.
Test-Time Learning for Large Language Models
Jinwu Hu (South China University of Technology), Mingkui Tan (South China University of Technology)
TransformerLarge Language ModelTextBenchmarkAgriculture RelatedFinance Related
🎯 What it does: This paper proposes a Test-Time Learning (TTL) framework called TLM, which dynamically updates large language models (LLMs) during inference using only unlabeled test data, achieving self-supervised adaptation through input perplexity minimization.
Test-Time Multimodal Backdoor Detection by Contrastive Prompting
Yuwei Niu (Chongqing University), Lei Feng (Southeast University)
Anomaly DetectionTransformerLarge Language ModelPrompt EngineeringContrastive LearningImageTextMultimodality
🎯 What it does: A multimodal method called BDetCLIP is proposed for detecting backdoor samples during the inference phase of CLIP using contrastive prompts.
Test-Time Preference Optimization: On-the-Fly Alignment via Iterative Textual Feedback
Yafu Li (Shanghai AI Laboratory), Yu Cheng (Chinese University of Hong Kong)
OptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelReinforcement LearningPrompt EngineeringText
🎯 What it does: A method for real-time preference alignment of large language models during inference (Test-Time Preference Optimization, TPO) is proposed, which does not require updating model parameters.
Test-Time Selective Adaptation for Uni-Modal Distribution Shift in Multi-Modal Data
MingCai Chen (Nanjing University of Posts and Telecommunications), Bingkun BAO
Domain AdaptationVideoMultimodalityAudio
🎯 What it does: This study investigates the unimodal distribution drift during multimodal testing and proposes a selective adaptation method.
Test-Time Training Provably Improves Transformers as In-context Learners
Halil Alperen Gozeten (University of Michigan), Samet Oymak (University of Michigan)
TransformerSupervised Fine-TuningTabular
🎯 What it does: The theoretical analysis and experimental validation of training Transformer with one-step testing (TTT) demonstrate that it can significantly enhance contextual learning effects.
Testing Conditional Mean Independence Using Generative Neural Networks
Yi Zhang (University of Illinois), Xiaofeng Shao (Washington University)
ClassificationRecognitionGenerative Adversarial NetworkImage
🎯 What it does: A new Conditional Mean Independence (CMI) testing method is proposed, utilizing deep generative neural networks to estimate the involved conditional mean functions, and introducing a bootstrap-based testing procedure.
Testing the Limits of Fine-Tuning for Improving Visual Cognition in Vision Language Models
Luca M. Schulze Buschoff (Institute for Human-Centered AI), Eric Schulz (Institute for Human-Centered AI)
Representation LearningTransformerSupervised Fine-TuningVision Language ModelImageMultimodality
🎯 What it does: The study conducts task-specific parameter-efficient fine-tuning on visual language models to enhance performance in intuitive physics and causal reasoning, and evaluates its alignment and generalization capabilities with human behavior.
Text-to-CAD Generation Through Infusing Visual Feedback in Large Language Models
Ruiyu Wang (University of Toronto), Jiang Bian (Microsoft Research Asia)
GenerationOptimizationTransformerLarge Language ModelReinforcement LearningVision Language ModelText
🎯 What it does: The CADFusion framework is proposed, dividing the training process of converting text to CAD into two stages: Sequence Learning (SL) and Visual Feedback (VF), which are alternately conducted, utilizing LLM to generate and optimize CAD parameter sequences.
Text-to-LoRA: Instant Transformer Adaption
Rujikorn Charakorn (Sakana AI), Robert Tjarko Lange
CompressionDomain AdaptationOptimizationKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Trained a hypernetwork (T2L) that can instantly generate low-rank adapters (LoRA) based solely on natural language task descriptions, enabling rapid, training-free task adaptation for large models.
TextCenGen: Attention-Guided Text-Centric Background Adaptation for Text-to-Image Generation
Tianyi Liang (East China Normal University), Chenhui Li (East China Normal University)
GenerationData SynthesisDomain AdaptationTransformerDiffusion modelImageText
🎯 What it does: Using text prompts and predefined blank areas to generate images that are semantically complete and have space for text placement.
Textual Unlearning Gives a False Sense of Unlearning
Jiacheng Du (Zhejiang University), Kui Ren (Zhejiang University)
Safty and PrivacyTransformerLarge Language ModelText
🎯 What it does: This paper reveals the failure of textual unlearning in language models and the potential privacy leakage risks in practical applications through rigorous auditing and novel attack methods.
Textural or Textual: How Vision-Language Models Read Text in Images
Hanzhang Wang (Shanghai University), Qingyuan Ma (Shanghai University)
ClassificationAdversarial AttackTransformerVision Language ModelImageText
🎯 What it does: This paper constructs the ToT dataset and utilizes Intrinsic Dimension (ID) analysis to study whether visual-language models rely more on texture features or semantic understanding when processing text in images. It also explores the impact of factors such as font size and synonyms/homophones on model performance.
TGDPO: Harnessing Token-Level Reward Guidance for Enhancing Direct Preference Optimization
Mingkang Zhu (Chinese University of Hong Kong), Jiaya Jia (Hong Kong University of Science and Technology)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes the introduction of token-level reward guidance in Direct Preference Optimization (DPO) to improve the alignment training of LLMs.
The Batch Complexity of Bandit Pure Exploration
Adrienne Tuynman (Inria), Rémy Degenne (Inria)
Optimization
🎯 What it does: This paper studies the fixed confidence pure exploration problem in stochastic multi-armed bandits, proposing a batch algorithm and providing upper bounds for sample complexity and batch complexity.
The Berkeley Function Calling Leaderboard (BFCL): From Tool Use to Agentic Evaluation of Large Language Models
Shishir G Patil, Joseph E. Gonzalez (University of California)
OptimizationAI Code AssistantTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: The Berkeley Function Calling Leaderboard (BFCL) has been proposed and released, which is a large-scale function call evaluation benchmark covering single-turn, multi-source, multi-turn, and agent-based scenarios, aimed at systematically assessing the interaction capabilities of large language models (LLMs) with external tools/APIs.
The Best of Both Worlds: Bridging Quality and Diversity in Data Selection with Bipartite Graph
Minghao Wu (Monash University), Gholamreza Haffari (Monash University)
Supervised Fine-TuningTextBenchmark
🎯 What it does: A data subset selection method based on a bipartite graph, GRAPHFILTER, is proposed to simultaneously consider data quality and diversity during supervised fine-tuning.
The Brain's Bitter Lesson: Scaling Speech Decoding With Self-Supervised Learning
Dulhan Jayalath (University of Oxford), Oiwi Parker Jones (University of Oxford)
ClassificationRecognitionConvolutional Neural NetworkSupervised Fine-TuningContrastive LearningBiomedical DataMagnetic Resonance ImagingAudio
🎯 What it does: By using neuroscience-inspired self-supervised pre-training tasks on nearly 400 hours of heterogeneous MEG data from nearly 900 subjects, followed by linear probe fine-tuning on a small amount of labeled auditory speech data, efficient decoding of auditory speech and phonemes was achieved.
The Butterfly Effect: Neural Network Training Trajectories Are Highly Sensitive to Initial Conditions
Gül Sena Altıntaş (University of Toronto), David Rolnick (McGill University)
Convolutional Neural NetworkTransformerImageText
🎯 What it does: This paper studies the extreme sensitivity to initial conditions during the training process of neural networks and demonstrates that applying minimal perturbations in the early stages of training can lead to bifurcations in the training trajectory, illustrating the butterfly effect.
The Canary’s Echo: Auditing Privacy Risks of LLM-Generated Synthetic Text
Matthieu Meeus (Imperial College London), Reza Shokri (National University of Singapore)
Data SynthesisSafty and PrivacyTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper conducts a privacy risk audit on synthetic data generated using LLMs for text synthesis, proposing a membership inference attack (MIA) that can be completed solely by observing the synthetic data.
The Case for Learned Provenance-based System Behavior Baseline
Yao Zhu (Zhejiang University), Shouling Ji (Zhejiang University)
Anomaly DetectionConvolutional Neural NetworkRecurrent Neural NetworkSupervised Fine-TuningTime Series
🎯 What it does: This paper proposes a learning-based intrinsic graph behavior baseline that maps events to vectors using adaptive embedding and predicts event normality through a lightweight regression model. The predicted results are then used as labels and combined with a label propagation framework to complete anomaly path mining and alerting in real-time streams.
The Complexity of Learning Sparse Superposed Features with Feedback
Akash Kumar (University of California)
OptimizationRepresentation LearningAuto EncoderTabular
🎯 What it does: This paper studies obtaining feedback through relative triplet comparisons (triplet feedback) from models or human agents, thereby learning and retrieving hidden sparse dictionary features in neural networks or sparse autoencoders (SAE) layers.
The Courage to Stop: Overcoming Sunk Cost Fallacy in Deep Reinforcement Learning
Jiashun Liu (Hong Kong University of Science and Technology), Ling Pan (Hong Kong University of Science and Technology)
Reinforcement LearningSequential
🎯 What it does: An adaptive early termination mechanism called LEAST, based on Q-values and gradient statistics, is proposed to overcome the sunk cost fallacy in deep reinforcement learning.
The dark side of the forces: assessing non-conservative force models for atomistic machine learning
Filippo Bigi (Ecole Polytechnique Federale de Lausanne), Michele Ceriotti (Ecole Polytechnique Federale de Lausanne)
Graph Neural NetworkTabularPhysics Related
🎯 What it does: The behavior and issues of non-conservative force models in molecular dynamics and geometric optimization were studied, and a strategy for combining non-conservative and conservative forces was proposed.
The Devil Is in the Details: Tackling Unimodal Spurious Correlations for Generalizable Multimodal Reward Models
Zichao Li (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences), Le Sun (Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences)
Recommendation SystemReinforcement Learning from Human FeedbackTransformerReinforcement LearningVision Language ModelMultimodality
🎯 What it does: A Shortcut-aware MM-RM learning algorithm is proposed, which reduces the dependence of the multimodal reward model on text shortcuts by dynamically reweighting samples.
The Diffusion Duality
Subham Sekhar Sahoo (Cornell University), Volodymyr Kuleshov (Cornell University)
GenerationComputational EfficiencyKnowledge DistillationDiffusion modelText
🎯 What it does: This paper conducts a theoretical analysis of the Unified State Discrete Diffusion Model (USDM) and proposes the Duo framework, establishing a Diffusion Duality with the Gaussian diffusion model, thereby enhancing training and sampling efficiency.
The Disparate Benefits of Deep Ensembles
Kajetan Schweighofer (Johannes Kepler University Linz), Nuria M Oliver
ClassificationRecognitionConvolutional Neural NetworkSupervised Fine-TuningImageBiomedical DataComputed Tomography
🎯 What it does: This study investigates the impact of Deep Ensembles on algorithmic fairness, finding that it produces unequal benefits for different protected groups. It proposes and validates that differences in predictive diversity are the main reason for this effect, and demonstrates that the Hardt post-processing method can mitigate the decline in fairness while maintaining performance.
The Double-Ellipsoid Geometry of CLIP
Meir Yossef Levi (Technion Israel Institute of Technology), Guy Gilboa (Technion Israel Institute of Technology)
GenerationRepresentation LearningTransformerVision Language ModelContrastive LearningImageText
🎯 What it does: This study investigates the original geometric structure of the CLIP embedding space and finds that images and texts are located on offset elliptical shells;
The Elicitation Game: Evaluating Capability Elicitation Techniques
Felix Hofstätter, Francis Rhys Ward (Imperial College London)
Large Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: By implementing password-locking and circuit-breaking techniques on language models, a 'model organism' with hidden capabilities was constructed. Various capability mining methods (few-shot, pre-filling, multi-turn, concept/persona steering, and supervised/anti-refusal fine-tuning) were systematically evaluated on multiple-choice questions (WMDP, MMLU) and code generation (APPS) benchmarks, comparing their effectiveness in recovering hidden capabilities.
The Emperor's New Clothes in Benchmarking? A Rigorous Examination of Mitigation Strategies for LLM Benchmark Data Contamination
Yifan Sun (University of Illinois Urbana-Champaign), Huan Zhang (University of Illinois Urbana-Champaign)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: This study investigates the impact of Benchmark Data Contamination on the evaluation of LLMs, proposing two fine-grained metrics: fidelity and contamination resistance, and designing a rigorously controlled experimental process.
The Empirical Mean is Minimax Optimal for Local Glivenko-Cantelli
Doron Cohen (Ben-Gurion University of the Negev), Roi Weiss (Ariel University)
🎯 What it does: The study investigates the mean estimation problem within the local Glivenko-Cantelli framework, proving the optimality of the empirical mean estimator in the learnable distribution family LGC, and exploring the possibility of extending the learnable family through different estimators.
The Energy Loss Phenomenon in RLHF: A New Perspective on Mitigating Reward Hacking
Yuchun Miao (Wuhan University), Dacheng Tao (Nanyang Technological University)
Reinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: In RLHF training, it was discovered that the increase in energy loss at the final layer of LLM is related to reward hacking. The EPPO algorithm is proposed to mitigate reward hacking by penalizing the growth of energy loss.
The Four Color Theorem for Cell Instance Segmentation
Ye Zhang (Harbin Institute of Technology), Jianxu Chen (Leibniz Institute for Analytical Sciences)
SegmentationConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper proposes a cell instance segmentation method based on the four-color theorem, transforming instance segmentation into four types of semantic segmentation.
The Generalized Skew Spectrum of Graphs
Armando Bellante (Politecnico di Milano), Alessandro Luongo (National University of Singapore)
Graph Neural NetworkGraph
🎯 What it does: A graph embedding method based on group theory harmonic analysis is proposed, which generalizes the skew spectrum introduced by Kondor and Borgwardt (2008) and can handle more complex graph structures, including attributed graphs, multilayer graphs, and hypergraphs.
The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independence
Tom Wollschläger (Technical University of Munich), Johannes Gasteiger
OptimizationAdversarial AttackLarge Language ModelPrompt EngineeringText
🎯 What it does: This study investigates the rejection mechanism in large language models (LLMs) and proposes a gradient-based representation engineering method to identify and manipulate rejection directions.