AAAI Conference on Artificial Intelligence Β· 2140 papers
Text-based Aerial-Ground Person Retrieval
Xinyu Zhou (Soochow University), Mang Ye (Wuhan University)
CodeRetrievalTransformerLarge Language ModelMixture of ExpertsVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: This paper proposes a text-driven aerial and ground person retrieval task (TAG-PR), and designs the corresponding dataset TAG-PEDES as well as a novel retrieval framework TAG-CLIP.
Text-Guided Channel Perturbation and Pre-Trained Knowledge Integration for Unified Multi-Modality Image Fusion
Xilai Li (Foshan University), Weijun Jiang (Foshan University)
CodeTransformerVision Language ModelTextMultimodalityMagnetic Resonance ImagingComputed TomographyPositron Emission Tomography
π― What it does: Proposed a unified multi-modal image fusion framework named UP-Fusion, combining channel perturbation with pre-trained knowledge to suppress redundant cross-modal features and enhance fusion quality.
Text-guided Controllable Diffusion for Realistic Camouflage Images Generation
Yuhang Qian (Nanjing University of Aeronautics and Astronautics), Jie Qin (Nanjing University of Aeronautics and Astronautics)
CodeGenerationPrompt EngineeringVision Language ModelDiffusion modelImageText
π― What it does: Proposed a controllable text-guided camouflage image generation method called CT-CIG, which can generate visually realistic camouflage images that are logically consistent with the background based on text prompts and masks.
CodeGenerationData SynthesisRetrievalTransformerLarge Language ModelVision Language ModelTextPoint CloudMeshRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Developed a text-to-scene synthesis framework called Reason-3D based on a large inference model, which can retrieve and place 3D objects from natural language descriptions in an unsupervised manner, generating complete indoor and outdoor environments.
TextShield-R1: Reinforced Reasoning for Tampered Text Detection
Chenfan Qu (South China University of Technology), Lianwen Jin (Ant Group)
CodeAnomaly DetectionExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningVision Language ModelImageTextMultimodalityBenchmark
π― What it does: This paper proposes TextShield-R1, a reinforcement learning-based multimodal large language model (MLLM) framework for detecting and explaining tampered text, enhancing detection, localization, and explanation capabilities through Forensic Continual Pre-training, Group Relative Policy Optimization (GRPO), and OCR Rectification, while constructing the TFR benchmark containing 45k multilingual, diverse tampering methods images;
Textual Self-Attention Network: Test-Time Preference Optimization Through Textual Gradient-Based Attention
Shibing Mo (Xidian University), Jing Liu (Xidian University)
CodeOptimizationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
π― What it does: Analyze, weight, and synthesize multiple candidate answers generated by LLMs using a parameter-free update text self-attention network (TSAN) during inference to produce answers more aligned with human preferences.
Yongqi Fan (East China University of Science and Technology), Tong Ruan (University of Chinese Academy of Sciences)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
π― What it does: This paper proposes TFRank, a pointwise reasoning ranking framework based on small-scale LLMs, which does not generate reasoning chains during the inference phase;
π― What it does: Proposes a training trajectory-guided dataset distillation method called TGDD to compress large-scale datasets while maintaining efficient performance.
The Avengers: A Routing Recipe for Collective Intelligence in Language Models
Yiqun Zhang (Northeastern University), Shuyue Hu (Singapore Management University)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelMixture of ExpertsTextBenchmarkRetrieval-Augmented Generation
π― What it does: Propose the Avengers framework, which employs an lightweight routing mechanism combining embedding-clustering-scoring-voting to aggregate the collective intelligence of 10 7B open-source models into a single answer.
The Curious Case of Analogies: Investigating Analogical Reasoning in Large Language Models
Taewhoo Lee (Korea University), Jaewoo Kang (Korea University)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
π― What it does: Analyze the internal mechanisms of large language models (LLMs) in proportion analogy and story analogy, revealing how they encode, recognize, and apply relationships.
The GATTACA Framework: Graph Neural Network-Based Reinforcement Learning for Controlling Biological Networks
Andrzej Mizera (University of Warsaw), Jakub Zarzycki (University of Warsaw)
CodeGraph Neural NetworkReinforcement LearningBiomedical Data
π― What it does: Propose a deep reinforcement learning framework called GATTACA based on graph neural networks for controlling asynchronous Boolean networks to achieve cellular reprogramming;
The Other Mind: How Language Models Exhibit Human Temporal Cognition
Lingyu Li (Shanghai Artificial Intelligence Laboratory), Yingchun Wang (Shanghai Artificial Intelligence Laboratory)
CodeExplainability and InterpretabilityRepresentation LearningTransformerLarge Language ModelText
π― What it does: Assessed large language models' (LLM) understanding of years through similarity judgment tasks, revealing that models spontaneously generate a subjective temporal reference point (β2025) and follow the Weber-Fechner Law (logarithmic compression of time distance) as their scale increases; further analyzed mechanisms of temporal cognition at neural, representational, and information exposure levels through neuron screening, activation patterns, linear probing, and semantic embedding; proposed an 'empiricism' perspective, viewing LLM cognition as a subjective construction of the external world through internal representation systems.
Haichuan Wang (Harvard University), Haifeng Xu (University of Chicago)
Code
π― What it does: Proposed the Publication Choice Problem game model, analyzing the interaction between researchers' submission strategies to different conferences/journals and venue impact factors, and proved the uniqueness of pure strategy equilibrium and the threshold effect of spotlight marking under binary types.
The Semantic Architect: How FEAML Bridges Structured Data and LLMs for Multi-Label Tasks
Wanfu Gao (Jilin University), Jun Gao (Jilin University)
CodeClassificationExplainability and InterpretabilityData-Centric LearningTransformerLarge Language ModelPrompt EngineeringTabularFinance Related
π― What it does: Designed and implemented FEAML, a closed-loop feature engineering framework that automatically generates interpretable features for multi-label learning tasks using large language models.
The Strong Lottery Ticket Hypothesis for Multi-Head Attention Mechanisms
Hikari Otsuka (Institute of Science Tokyo), Masato Motomura (Institute of Science Tokyo)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelText
π― What it does: Prove the existence of strong lottery ticket subnetworks (SLT) in randomly initialized multi-head attention mechanisms (MHA) and Transformers without normalization layers, and provide an upper bound on the approximation error for the target model; experiments verify that the error decreases exponentially with increasing hidden dimensions, is insensitive to input length, and propose an initialization scaling strategy for query/key weights.
The Structure-Equivalent Prior: Unifying Temporal Dynamics and 3D Evolution in 4D Latent Space
Jingyuan Gao (Beijing University of Chemical Technology), Kunfeng Wang (Beijing University of Chemical Technology)
CodeGenerationAutonomous DrivingAuto EncoderVideo
π― What it does: Propose a continuous 4D latent space representation (SEP-4D) based on structural equivalence prior for high-quality occupancy map reconstruction in dynamic 3D scenes.
The Visual Prism: Refracting Images into Parallel Multilingual Descriptions with Structured Visual Guidance
Chengpeng Fu (Harbin Institute of Technology), Ting Liu (Harbin Institute of Technology)
CodeGenerationTransformerSupervised Fine-TuningReinforcement LearningPrompt EngineeringVision Language ModelMultimodality
π― What it does: This paper proposes the PRISMS framework, which utilizes visual scene graphs as structured visual guidance to generate multilingual parallel image descriptions, addressing the issue of inconsistent focus across languages.
Themis: Automated Constraint-Aware Test Synthesis Framework for Code Reinforcement Learning
Shengyu Ye, Zhenya Huang (University Of Science And Technology Of China)
CodeData SynthesisData-Centric LearningLarge Language ModelTextBenchmark
π― What it does: Themis automatically generates high-quality test sets that meet constraints to enhance the reliability of reward signals in Code Reinforcement Learning (Code RL).
Theoretical and Empirical Analysis of Lehmer Codes to Search Permutation Spaces with Evolutionary Algorithms
Yuxuan Ma, Carsten Witt (Southern University of Science and Technology)
CodeOptimizationBenchmark
π― What it does: Investigate the representation advantages of Lehmer codes (inversion vectors) in evolutionary algorithms, conduct rigorous runtime analysis of RLS and (1+1)-EA in Lehmer space, and compare with classical linear permutation encoding; subsequently validate experimentally on theoretical benchmark functions and real-world linear ordering and quadratic assignment problems.
CodeRepresentation LearningGraph Neural NetworkLarge Language ModelTextGraphBenchmark
π― What it does: Proposed the first publicly available benchmark (THGB) for Text-Attributed Heterogeneous Graphs (TAHG), and evaluated multiple graph learning models on it;
Think Before You Segment: An Object-aware Reasoning Agent for Referring Audio-Visual Segmentation
Jinxing Zhou, Rao Muhammad Anwer (Mohamed Bin Zayed University Of Artificial Intelligence)
CodeSegmentationExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelVideoTextMultimodalityBenchmarkChain-of-Thought
π― What it does: This paper proposes TGS-Agent, which follows a three-step process of 'thinking-localization-segmentation'. It first explicitly reasons about the target object using the multimodal large language model Ref-Thinker, then generates bounding boxes with Grounding-DINO, followed by pixel-level segmentation with SAM2. Meanwhile, the paper constructs a more challenging benchmark called R2-AVSBench.
Think Then Rewrite: Reasoning Enhanced Query Rewriting for Domain Specific Retrieval
Ang Li (Zhejiang University), Kun Kuang (Ant Group)
CodeRetrievalLarge Language ModelReinforcement LearningContrastive LearningTextBiomedical Data
π― What it does: This paper proposes the Think-Then-Rewrite (TTR) framework, which leverages the reasoning capabilities of large language models to enhance query rewriting in specialized domains.
Think-J: Learning to Think for Generative LLM-as-a-Judge
Hui Huang (Alibaba Group), Wenbo Su (Alibaba Group)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBenchmarkChain-of-Thought
π― What it does: Propose Think-J, which first initializes the generative LLM discriminator using a small set of labeled thinking trajectories, and then optimizes its thinking process through offline/online reinforcement learning to improve preference evaluation of generated text;
Think, Speak, Decide: Language-Augmented Multi-Agent Reinforcement Learning for Economic Decision-Making
Heyang Ma (Chinese Academy of Sciences), Haifeng Zhang (Peking University)
CodeLarge Language ModelReinforcement LearningTextFinance RelatedRetrieval-Augmented GenerationChain-of-Thought
π― What it does: Constructed the LAMP framework, integrating the reasoning and reflection of large language models (LLMs) with multi-agent reinforcement learning (MARL) to achieve language-enhanced economic decision-making.
Thinking Aesthetics Assessment of Image Color Temperature: Models, Datasets and Benchmarks
Jinguang Cheng (Beijing University of Posts and Telecommunications), Anlong Ming (Beijing University of Posts and Telecommunications)
CodeClassificationTransformerVision Language ModelContrastive LearningImageTextMultimodalityBenchmark
π― What it does: This paper proposes the Image Color Temperature Aesthetics Assessment task (ICTAA) and constructs a large-scale color temperature aesthetics dataset named ICTAA240K;
Tighter Truncated Rectangular Prism Approximation for RNN Robustness Verification
Xingqi Lin, Zhenbing Zeng (East China Normal University)
CodeOptimizationExplainability and InterpretabilityRecurrent Neural NetworkImageTextAudio
π― What it does: Proposed a dual-plane compact approximation method based on truncated rectangular prisms for robustness verification of RNNs (such as LSTM), and implemented the DeepPrism verification tool.
π― What it does: By integrating tile-based perceptual feedback into the 3D Gaussian scattering rendering pipeline, dynamically densifying and pruning Gaussians to enhance detail and boundary performance;
TIM++: Transductive Information Maximization for Few-Shot CLIP
Yingping Li (Xidian University), Shuiping Gou (Xidian University)
CodeClassificationRepresentation LearningMeta LearningVision Language ModelContrastive LearningMultimodality
π― What it does: In transductive few-shot classification, joint learning is performed using textual information and visual features from the pre-trained vision-language model CLIP to enhance few-shot recognition performance.
Time Series Class-Incremental Learning via Confidence-guided Mask Distillation and Prototype-guided Contrastive Learning
Yu Liu (Dalian University of Technology), Qi Jia (Dalian University of Technology)
CodeClassificationKnowledge DistillationContrastive LearningTime Series
π― What it does: Proposes two novel techniques for incremental learning on time series: Confidence-Guided Mask Distillation (CMD) and Prototype-Guided Contrastive Learning (PCL), and jointly applies them to an incremental learning framework without rehearsal.
π― What it does: Propose the interPDN model, which directly outputs a discrete probability distribution at each time step and takes the expectation for prediction, thereby better capturing uncertainty in time series forecasting.
Time-Frequency Augmented Multi-level Contrastive Clustering for Time Series
Congyu Wang (Jiangsu Normal University), Xiang Jiang (Jiangsu Normal University)
CodeRepresentation LearningConvolutional Neural NetworkContrastive LearningTime Series
π― What it does: Proposed an end-to-end deep time series clustering framework called TFMCC, which improves clustering performance by jointly learning representations and clustering.
TimeMosaic: Temporal Heterogeneity Guided Time Series Forecasting via Adaptive Granularity Patch and Segment-wise Decoding
Kuiye Ding (Institute of Computing Technology, Chinese Academy of Sciences), Jianfeng Zhan (Institute of Computing Technology, Chinese Academy of Sciences)
CodeTransformerPrompt EngineeringTime Series
π― What it does: Proposed a time series forecasting framework called TimeMosaic that addresses the issues of encoding heterogeneity and decoding heterogeneity.
TinyChemVL: Advancing Chemical Vision-Language Models via Efficient Visual Token Reduction and Complex Reaction Tasks
Xuanle Zhao (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences), Bo Xu (Key Laboratory of Cognition and Decision Intelligence for Complex Systems, Institute of Automation, Chinese Academy of Sciences)
CodeComputational EfficiencyDrug DiscoveryTransformerVision Language ModelImageTextMultimodalityBenchmark
π― What it does: This paper proposes TinyChemVL, an efficient cheminformatics vision-language model, and designs adaptive visual token merging/cropping strategies and a new reaction-level benchmark, ChemRxn-V, achieving unified processing for molecular and reaction tasks;
π― What it does: Proposed a two-stage modal denoising and compensation framework (TMDC) for simultaneously handling missing and noisy modalities in multimodal sentiment analysis.
TO-GATE: Clarifying Questions and Summarizing Responses with Trajectory Optimization for Eliciting Human Preference
Yulin Dou (Yunnan University), Jiangming Liu (Yunnan University)
CodeReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningMixture of ExpertsContrastive LearningText
π― What it does: This work proposes a training framework called TO-GATE, which enhances the ability of LLMs to raise clarifying questions and generate final personalized answers in multi-turn dialogues through trajectory optimization.
ToC: Tree-of-Claims Search with Multi-Agent Language Models
Shuyang Yu (Columbia University), Hui Hu (AIGROW)
CodeOptimizationTransformerLarge Language ModelAgentic AITextMultimodalityChain-of-Thought
π― What it does: Proposed the Tree of Claims framework, modeling patent claim optimization as a structured search problem based on multi-agent LLM and Monte Carlo Tree Search.
Token Painter: Training-Free Text-Guided Image Inpainting via Mask Autoregressive Models
Longtao Jiang (University Of Science And Technology Of China), Zhihui Li (Mbzuai)
CodeRestorationGenerationTransformerVision Language ModelImageTextBenchmark
π― What it does: Achieve training-free text-guided image inpainting using the Mask AutoRegressive (MAR) model, preserving the background while generating content consistent with the text prompt.
CodeRecommendation SystemTransformerLarge Language ModelMixture of ExpertsAuto EncoderText
π― What it does: Propose a unified item tokenization framework called UniTok, enabling LLM generative recommendation systems to share the same tokenizer across multiple domains.
TongUI: Internet-Scale Trajectories from Multimodal Web Tutorials for Generalized GUI Agents
Bofei Zhang (Beijing Institute Of Technology), Qing Li (Bigai)
CodeData SynthesisTransformerLarge Language ModelVision-Language-Action ModelVideoTextMultimodality
π― What it does: Built a framework named TongUI, converting millions of multimodal tutorials from the internet into 1 million GUI interaction trajectories, and used these trajectories to train GUI agents.
Too Sure for Our Own Good: A User Study on AI Confidence and Human Reliance
Caterina Fregosi (University of Milano-Bicocca), Federico Cabitza (University of Milano-Bicocca)
CodeExplainability and InterpretabilityText
π― What it does: This study examined the impact of AI confidence calibration on human decision accuracy, appropriate reliance, automation bias, and conservative bias through a logical reasoning task involving 184 participants.
π― What it does: Proposes the TOPOGRAPH framework, integrating DTM outlier removal, adaptive kNN graph construction, and spectral graph coarsening, to achieve efficient and robust intrinsic persistent homology (IPH) computation, while significantly compressing data while preserving topological structures;
Topological Federated Clustering via Gravitational Potential Fields Under Local Differential Privacy
Yunbo Long (University of Cambridge), Alexandra Brintrup (University of Cambridge)
CodeFederated LearningSafty and PrivacyImageTabular
π― What it does: Developed a one-shot federated clustering method called GFC, which can cluster non-IID data with heterogeneous distributions under local differential privacy (LDP).
π― What it does: Proposes a topology-preserving class-incremental learning framework called TaKP, integrating topological distillation based on persistent homology with a dual-branch adaptive rebalancing mechanism to better preserve old class knowledge during continual learning.
π― What it does: This paper proposes a network called TANSR that integrates the topological information of superpixel graphs into Vision Transformers for scene recognition.
π― What it does: Designed a test-time adaptation framework T3A without backpropagation, based on topological consistency, integrating structure-aware modeling and box regression adaptation for medical detection.
TOSC: Task-Oriented Shape Completion for Open-World Dexterous Grasp Generation from Partial Point Clouds
Weishang Wu (National University of Defense Technology), Zhiping Cai (National University of Defense Technology)
CodeGenerationPose EstimationRobotic IntelligenceTransformerLarge Language ModelVision Language ModelDiffusion modelFlow-based ModelAuto EncoderPoint Cloud
π― What it does: This paper proposes a task-oriented shape completion framework to generate complete shapes from partial point clouds that are compatible with downstream grasping tasks, and further generates executable multi-finger grasping poses.
Toward Better EHR Reasoning in LLMs: Reinforcement Learning with Expert Attention Guidance
Yue Fang (Peking University), Yasha Wang (Peking University)
CodeTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningBiomedical DataElectronic Health Records
π― What it does: Proposed the EAG-RL two-stage training framework, which significantly enhances LLM's reasoning capabilities on electronic health records (EHR) through expert-guided Monte Carlo Tree Search initialization and attention alignment-based reinforcement learning.
π― What it does: Designed a foreground-consistent learning model, FCLM, for high-precision image matting and binary segmentation, addressing domain differences in synthetic data, foreground consistency learning, and interactive multi-instance prediction.
Toward Time-Continuous Data Inference in Sparse Urban CrowdSensing
Hao Du (Jilin University), Yuanbo Xu (Jilin University)
CodeComputational EfficiencyData-Centric LearningTime Series
π― What it does: Proposes the TIME-DMF framework for temporally continuous sparse urban sensing data completion, achieving high-precision inference at any arbitrary time point.
π― What it does: In adverse weather conditions, generating virtual LiDAR points through 4D radar guidance enhances LiDAR point cloud geometric density and structure, thereby improving 3D object detection;
Towards Authentic Movie Dubbing with Retrieve-Augmented Director-Actor Interaction Learning
Rui Liu (Inner Mongolia University), Zhenqi Jia (Inner Mongolia University)
CodeGenerationRetrievalGraph Neural NetworkLarge Language ModelMultimodalityRetrieval-Augmented Generation
π― What it does: Implemented a retrieval-enhanced director-actor interactive framework called Authentic-Dubber for generating emotion-expressive movie dubbing.
Towards Autonomous UAV Visual Object Search in City Space: Benchmark and Agentic Methodology
Yatai Ji (National University of Defense Technology), Quanjun Yin (National University of Defense Technology)
CodeAutonomous DrivingTransformerLarge Language ModelPrompt EngineeringVision Language ModelWorld ModelImageTextMultimodalityBenchmark
π― What it does: Propose the CityAVOS benchmark for urban space UAV visual target search (AVOS), and design a PRPSearcher agent based on a multi-modal large language model, achieving perception, reasoning, and planning through three specialized maps.
π― What it does: Improving performance on code understanding tasks such as code search and code clone detection by continuing contrastive learning training on a pre-trained decoder-only model.
π― What it does: Propose the Hidden Label Distribution Learning (HidLDL) problem and design a recovery method leveraging proportion constraints, graph similarity, and low-rank structure.
Towards Effective and Efficient Context-aware Nucleus Detection in Histopathology Whole Slide Images
Zhongyi Shui (Zhejiang University), Lin Yang (Westlake University)
CodeObject DetectionSegmentationComputational EfficiencyConvolutional Neural NetworkTransformerBiomedical Data
π― What it does: Proposed a context-aware method for nucleus detection on whole-slide images, enhancing detection accuracy and speed by aggregating features from historical sliding windows.
Towards Effective Offensive Security LLM Agents: Hyperparameter Tuning, LLM as a Judge, and a Lightweight CTF Benchmark
Minghao Shao (New York University), Muhammad Shafique (New York University)
CodeHyperparameter SearchTransformerLarge Language ModelAgentic AITextBenchmark
π― What it does: This paper systematically evaluates the performance of large language models in Capture the Flag (CTF) attack agents and investigates the impact of hyperparameters such as temperature, top-p, maximum token length, and multi-agent collaboration on problem-solving success rates.
Towards Effective, Stealthy, and Persistent Backdoor Attacks Targeting Graph Foundation Models
Jiayi Luo (Beihang University), Jianxin Li (Guangxi Normal University)
CodeAdversarial AttackGraph Neural NetworkGraph
π― What it does: Propose a backdoor attack framework for Graph Foundation Models (GFM) called GFM-BA, which can inject backdoors during the pre-training phase and maintain effectiveness, stealthiness, and persistence across different downstream tasks.
Towards Explainable Video Camouflaged Object Detection: SAM2 with Eventstream-Inspired Data
Hong Zhang (Beihang University), Yifan Yang (Beihang University)
CodeObject DetectionExplainability and InterpretabilitySpiking Neural NetworkTransformerPrompt EngineeringVideoBenchmark
π― What it does: Propose an event-stream inspired dual-branch framework that automatically generates prompts and integrates temporal memory to achieve interpretable video camouflage object detection.
π― What it does: Proposed the MiniShift high-resolution fine-grained 3D anomaly detection dataset and the Simple3D real-time detection framework, addressing the localization and identification of minute defects in industrial point clouds.
Towards Multimodal Continual Knowledge Embedding with Modality Forgetting Modulation
Xiaowen Jiang, Jieming Yang (Zhengzhou University)
CodeRepresentation LearningConvolutional Neural NetworkGraph Neural NetworkTransformerVision Language ModelMultimodality
π― What it does: Proposes the MoFot framework, which mitigates catastrophic forgetting in multi-modal knowledge graph continual learning through multi-modal collaborative modulation, avoiding issues of retraining and knowledge loss.
Towards Reinforcement Learning from Neural Feedback: Mapping fNIRS Signals to Agent Performance
Julia Santaniello (Tufts University), Jivko Sinapov (Tufts University)
CodeReinforcement Learning from Human FeedbackTime SeriesBiomedical Data
π― What it does: Proposed a reinforcement learning feedback framework (RLNF) based on fNIRS neural signals, achieving implicit human feedback by mapping brain signals to agent performance;
π― What it does: Propose an elastic continuous test-time adaptation framework for edge devices (Elastic Edge CTTA), which generates personalized sub-models through resource-aware architecture search before deployment, followed by lightweight online self-supervised fine-tuning using low-rank adapters on the device, and feeds adaptation experiences back to the super network via structure-aware knowledge backflow to improve subsequent search quality.
π― What it does: This study proposes a conditional control diffusion model named CtrTab for synthesizing high-dimensional sparse (low-sample) tabular data;
π― What it does: This paper proposes a multimodal ultrasound diagnostic framework based on causal inference, utilizing two causal adjustment methods, BACL and FACL, to eliminate explicit and implicit confounding biases, achieving consistency between model decisions and expert diagnostic logic.
Towards Understanding Generalization in DP-GD: A Case Study in Training Two-Layer CNNs
Zhongjie Shi (Georgia Institute of Technology), Yuan Cao (RPTU Kaiserslautern-Landau)
CodeClassificationSafty and PrivacyConvolutional Neural Network
π― What it does: This paper studies the performance of two-layer Huberized ReLU convolutional neural networks in binary classification tasks, comparing the generalization and privacy performance of non-private gradient descent (GD) and differential privacy gradient descent (DP-GD), and proves that DP-GD can achieve better test accuracy within a specific signal-to-noise ratio (SNR) range.
Towards Zero-Shot Diabetic Retinopathy Grading: Learning Generalized Knowledge via Prompt-Driven Matching and Emulating
Huan Wang (University of Wollongong), Jun Shen (University of Wollongong)
CodeClassificationTransformerPrompt EngineeringContrastive LearningBiomedical Data
π― What it does: Propose a two-stage prompt-driven framework called ProME-DR based on CLIP for zero-shot diabetic retinopathy (DR) grading on multi-view retinal images.
TRACE: A Generalizable Drift Detector for Streaming Data-Driven Optimization
Yuan-Ting Zhong (South China University Of Technology), Yue-Jiao Gong (South China University Of Technology)
CodeAnomaly DetectionOptimizationTransformerTabularTime Series
π― What it does: Proposed a transferable concept drift detector called TRACE, and integrated it as a plugin into a streaming data-driven optimization algorithm (TRACE-EA).
TRACE: Trajectory-based Activation Change Estimation for Task-specific Data Selection
Ye He (Harbin Institute of Technology), Qi Shi (Tsinghua University)
CodeOptimizationData-Centric LearningTransformerLarge Language ModelSupervised Fine-TuningText
π― What it does: Proposes the TRACE algorithm, which selects task-specific data subsets by using neural activation changes induced by a single fine-tuning step on the validation set as the selection signal.
π― What it does: Proposed the TRACE framework, which constructs a joint graph of reactants and products and performs dynamic interaction refinement to achieve conditional prediction;
TraceTrans: Translation and Spatial Tracing for Surgical Prediction
Xiyu Luo (Southern University of Science and Technology), Tianyang Zhang (University of Birmingham)
CodeImage TranslationConvolutional Neural NetworkGenerative Adversarial NetworkOptical FlowBiomedical Data
π― What it does: Propose TraceTrans, an end-to-end image-to-image translation model for postoperative prediction, capable of generating trackable spatial deformations from preoperative images while maintaining target distribution matching.
π― What it does: This paper proposes the CLEAR-HUG two-stage framework, which uses ECG self-supervised learning to capture subtle differences in cardiac conduction and simulate clinical diagnostic processes;
Tracking the Unstable: Appearance-Guided Motion Modeling for Robust Multi-Object Tracking in UAV-Captured Videos
Jianbo Ma (State Key Laboratory of Optical Field Manipulation Science and Technology, Institute of Optics and Electronics, Chinese Academy of Sciences), Ming-Hsuan Yang (University of California, Merced)
π― What it does: Proposes a multi-object tracking framework AMOT based on appearance-motion consistency, achieving robust identity association and trajectory recovery in UAV videos through the AMC matrix and MTC module.
π― What it does: Studied the overfitting problem of probabilistic circuits (PC) under limited data, proposed a sharpness regularization method based on the Hessian trace, and implemented closed-form updates in EM and gradient optimization;
Tractable Weighted First-Order Model Counting with Bounded Treewidth Binary Evidence
VΓ‘clav KΕ―la (Czech Technical University in Prague), OndΕej KuΕΎelka (Czech Technical University in Prague)
CodeComputational EfficiencyGraph
π― What it does: This paper proposes a dynamic programming algorithm based on tree decomposition, which can solve weighted first-order model counting (WFOMC) for FOΒ²/CΒ² in polynomial time with respect to domain size when the tree width of the Gaifman graph for binary predicate evidence is limited.
π― What it does: Propose a brain-augmented target speaker extraction framework (TIDENet) that utilizes trainable EEG interpolation and a structure-shared dual-path encoder, enabling the simultaneous extraction of target-related and irrelevant features from EEG and mixed speech, and achieving feature complementary fusion within a single model.
π― What it does: Propose an unsupervised, no-additional-training-data "Efficient Residual Feature Tailoring (ERFT)" framework, achieving sub-second training and inference in cross-sensor pan-sharpening tasks.
Training-Free Spatio-temporal Decoupled Reasoning Video Segmentation with Adaptive Object Memory
Zhengtong Zhu (Soochow University), Fanzhang Li (Soochow University)
CodeObject TrackingSegmentationTransformerVision Language ModelVideo
π― What it does: Proposed a training-agnostic, spatiotemporally decoupled inference video segmentation framework SDAM, achieving stable segmentation through adaptive keyframe selection and memory modules.
π― What it does: Proposed the TrajAgg framework, which achieves efficient and accurate trajectory similarity computation using aggregation Transformer and hybrid training.
π― What it does: Propose an entropy-constrained vector quantization (ECVQ)-based feature encoding framework that directly encodes deep features without using traditional transformation modules.
π― What it does: This paper proposes a Transformer-based synchronized action caption generation method, which achieves synchronized output of actions and text through a control attention mechanism.
Transformers in Pseudo-Random Number Generation: A Dual Perspective on Theory and Practice
Ran Li (Northeast Normal University), Lingshu Zeng (Northeast Normal University)
CodeGenerationTransformerLarge Language ModelSequentialChain-of-Thought
π― What it does: This paper studies the simulation of classical pseudo-random number generators (LCG and Mersenne Twister) using Transformer (decoder-only + Chain-of-Thought), and theoretically proves that it can achieve non-uniform AC0; subsequently, pseudo-random number sequences are generated using GPT-2 training, and their performance is verified through NIST tests and prediction attack experiments.
TransMamba: A Sequence-Level Hybrid Transformer-Mamba Language Model
Yixing Li (Chinese University of Hong Kong), Yu Cheng (Tencent)
CodeTransformerLarge Language ModelText
π― What it does: Proposed a sequence-level hybrid Transformer-Mamba language model called TransMamba, which dynamically switches between attention and state space models in the same layer with shared parameters, achieving efficient long-sequence modeling.
Transolver Is a Linear Transformer: Revisiting Physics-Attention Through the Lens of Linear Attention
Wenjie Hu (National University of Defense Technology), Yong Dou (National University of Defense Technology)
CodeComputational EfficiencyTransformerMeshBenchmarkPhysics Related
π― What it does: This paper proposes LinearNO, which reinterprets Transolver's Physics-Attention as linear attention and achieves a more efficient model through two-step generalization and simplification;
Tree-Based Stochastic Optimization for Solving Large-Scale Urban Network Security Games
Shuxin Zhuang (City University of Hong Kong), Youzhi Zhang (Hong Kong Institute of Science & Innovation)
CodeOptimizationGraph
π― What it does: A tree-structured stochastic optimization framework (TSO) is studied for solving Nash equilibrium in large-scale urban cyber security games.
π― What it does: Propose the TripleFDS framework, which explicitly splits and reconstructs the ternary features (text content, style, and background) in scene text editing;
CodeDomain AdaptationTransformerVision Language ModelContrastive LearningImageTextMultimodality
π― What it does: TRUST proposes a framework for unsupervised domain adaptation leveraging text robustness, enhancing the generalization of visual models in target domains through pseudo label generation, uncertainty estimation, and multimodal soft contrastive learning.
π― What it does: This paper proposes the TSBOW (Traffic Surveillance Benchmark for Occluded vehicles under various Weather conditions) dataset, which includes 32 hours of CCTV surveillance video, 3.2M frames, and annotations for 8 categories of traffic participants, with object detection experiments conducted on this dataset.
TSGDiff: Rethinking Synthetic Time Series Generation from a Pure Graph Perspective
Lifeng Shen (Chongqing University of Posts and Telecommunications), Lele Long (Chongqing University of Posts and Telecommunications)
CodeGenerationData SynthesisGraph Neural NetworkDiffusion modelAuto EncoderGraphTime Series
π― What it does: Propose a graph-based temporal data generation framework called TSGDiff, which converts multivariate time series into dynamic graphs and generates synthetic time series through diffusion models in the latent graph space;
π― What it does: Propose the TSPE-GS method, which utilizes a probabilistic deep distribution for multi-modal depth extraction of semi-transparent surfaces in 3D Gaussian Splatting, and achieves multi-layer surface reconstruction through advanced TSDF fusion.
TSPO: Temporal Sampling Policy Optimization for Long-form Video Language Understanding
Canhui Tang (Xi'an Jiaotong University), Hao Sun (China Telecom)
CodeTransformerLarge Language ModelReinforcement LearningVision Language ModelVideoMultimodality
π― What it does: Proposed the Temporal Sampling Policy Optimization (TSPO) framework, which employs reinforcement learning to end-to-end train an event-aware temporal sampler, achieving joint decision-making for sparse frame sampling and language generation in long videos;
π― What it does: Propose a training-agnostic temporal token fusion framework that intelligently merges historical and current visual tokens by leveraging gray-scale pixel differences and attention-based semantic relevance detection, thereby enhancing the inference quality of VLA models.