AAAI 2025 Papers with Code — Page 14

TinySubNets: An Efficient and Low Capacity Continual Learning Strategy

Marcin Pietron, Roberto Corizzo (American University)

🎯 What it does: This paper proposes TinySubNets (TSN), a no-forgetting continual learning framework based on pruning, variable sparsity, adaptive quantization, and weight sharing.

TNCSE: Tensor Norm Constraints for Unsupervised Contrastive Learning of Sentence Embeddings

Tianyu Zong (University of Chinese Academy of Sciences), Jungang Xu (University of Chinese Academy of Sciences)

🎯 What it does: An unsupervised contrastive learning framework TNCSE based on tensor norm constraints is proposed, and sentence embeddings are achieved through a dual-encoder integration.

TokenMatcher: Diverse Tokens Matching for Unsupervised Visible-Infrared Person Re-Identification

Xiao Wang (Wuhan University of Science and Technology), Xin Xu (Wuhan University)

🎯 What it does: A TokenMatcher framework is designed to achieve unsupervised visible-infrared person re-identification through multi-class token matching, neighborhood learning, and homogeneous fusion.

Tokenphormer: Structure-aware Multi-token Graph Transformer for Node Classification

Zijie Zhou (University of Macau), Leong Hou U (University of Macau)

🎯 What it does: A graph transformer named Tokenphormer is proposed, which learns node representations through various fine-grained and global token generation mechanisms (walk-token, SGPM-token, hop-token) to achieve efficient node classification.

Told You That Will Not Work: Optimal Corrections to Planning Domains Using Counter-Example Plans

Songtuan Lin (Australian National University), Pascal Bercher (Paris-Saclay University)

🎯 What it does: This paper proposes an optimal repair method for planning domains based on positive and negative plans, which can output the smallest set of atomic repairs under the premise of given positive and negative example plans, making all positive examples feasible and all negative examples infeasible.

ToMATO: Verbalizing the Mental States of Role-Playing LLMs for Benchmarking Theory of Mind

Kazutoshi Shinoda (NTT Corporation), Kuniko Saito (NTT Corporation)

🎯 What it does: The ToMATO benchmark dataset was constructed to generate psychological state question-answering through multi-turn LLM-LLM dialogues, assessing the Theory of Mind (ToM) capabilities of LLMs.

Topology-Aware 3D Gaussian Splatting: Leveraging Persistent Homology for Optimized Structural Integrity

Tianqi Shen (City University of Hong Kong), Ning An (China Coal Research Institute)

🎯 What it does: Based on 3D Gaussian Splatting, we propose Topology-Aware 3D Gaussian Splatting (Topology-GS), which enhances view synthesis quality through two new techniques: 1) Local Voronoi Interpolation (LPVI) based on persistent homology to fill in the sparsity of point clouds in low-curvature areas; 2) Persistent Homology Loss (PersLoss) constrains the topological features of images during training, ensuring the integrity of feature layer structures.

Toward Efficient Data-Free Unlearning

Chenhao Zhang (University of Queensland), Miao Xu (University of Queensland)

🎯 What it does: This paper proposes a data-free unlearning method based on data-free knowledge distillation (ISPF), which addresses the issue of low distillation efficiency caused by the generator producing too many forgotten class samples in traditional methods.

Toward Verifiable Instruction-Following Alignment for Retrieval Augmented Generation

Guanting Dong (Renmin University of China), Ji-Rong Wen (Renmin University of China)

🎯 What it does: Generate high-quality instruction-following data through an automated and verifiable synthesis process, enhancing the instruction adherence capability of retrieval-augmented generation systems.

Towards a Comprehensive, Efficient and Promptable Anatomic Structure Segmentation Model Using 3D Whole-Body CT Scans

Heng Guo (DAMO Academy Alibaba Group), Minfeng Xu (DAMO Academy Alibaba Group)

🎯 What it does: CT-SAM3D is proposed, a three-dimensional interactive segmentation model based on whole-body CT, capable of achieving high-precision segmentation of nearly a hundred anatomical structures.

Towards a Multimodal Large Language Model with Pixel-Level Insight for Biomedicine

Xiaoshuang Huang (Baidu Inc), Yehui Yang (Peking University)

Towards Audio-Visual Navigation in Noisy Environments: A Large-Scale Benchmark Dataset and an Architecture Considering Multiple Sound-Sources

Zhanbo Shi (Tongji University), Ying Shen (Tongji University)

🎯 What it does: A large-scale audio-visual navigation benchmark dataset BeDAViN and a navigation framework ENMuS3 for multi-source noise environments are proposed, aiming to address the challenges of audio-visual navigation in the presence of multiple sound sources.

Towards Better Robustness Against Natural Corruptions in Document Tampering Localization

Huiru Shao (Xi'an Jiaotong-Liverpool University), Qiufeng Wang (Duke Kunshan University)

🎯 What it does: Proposes adversarial forensic regularization (AFR) based on mutual information and distribution discrepancy regularization (DDR) to enhance the robustness of document tampering localization under natural degradation (noise, JPEG, social network transmission) and adversarial attacks; simultaneously constructs the TSroie-CRP dataset.

Towards Efficient Low-Order Hybrid Optimizer for Language Model Fine-Tuning

Minping Chen (Hong Kong University of Science and Technology), Zeyi Wen (Hong Kong University of Science and Technology)

🎯 What it does: A low-order hybrid optimizer, LoHO, is proposed, which integrates zero-order (MeZO) and first-order (SGD/Adam) optimizers in language model fine-tuning, achieving memory efficiency while improving accuracy and convergence speed.

Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering

Hao Lu (Hong Kong University of Science and Technology), Ying-Cong Chen

🎯 What it does: This paper proposes a general multi-camera 3D object detection framework PR-BEV based on perspective rendering, which reconstructs semantic maps from different viewpoints in BEV using implicit foreground volumes and corrects object positions through 3D boxes or pre-trained 2D detectors, thereby enhancing cross-domain generalization performance.

Towards Loss-Resilient Image Coding for Unstable Satellite Networks

Hongwei Sha (Nanjing University), Zhan Ma (Nanjing University)

🎯 What it does: A loss-robust learning-based image compression method for unstable satellite networks is proposed, achieving an evolvable multi-level bitstream;

Towards Macro-AUC Oriented Imbalanced Multi-Label Continual Learning

Yan Zhang (Shandong University), Yilong Yin (Shandong University)

🎯 What it does: This paper proposes a macro AUC optimization method for multi-label continual learning (MLCL), mainly achieved by introducing a reweighted label distribution-aware margin loss (RLDAM) and a weight retention update (WRU) mechanism;

Towards Multimodal Sentiment Analysis via Hierarchical Correlation Modeling with Semantic Distribution Constraints

Qinfu Xu (China University of Petroleum), Hengyang Zhou (China University of Petroleum)

🎯 What it does: This paper studies a hierarchical correlation modeling network called HCMNet for multimodal sentiment analysis.

Towards Open-Vocabulary Remote Sensing Image Semantic Segmentation

Chengyang Ye (Dalian University of Technology), Pingping Zhang (Dalian University of Technology)

🎯 What it does: This paper proposes the Open Vocabulary Remote Sensing Image Semantic Segmentation task (OVRSISS), constructs the LandDiscover50K dataset with 51,846 images and 40 categories, and designs the GSNet model to achieve pixel-level segmentation for arbitrary semantic categories.

Towards Practical Classical Planning Compilations of Numeric Planning

Luigi Bonassi (University of Brescia), Enrico Scala (University of Brescia)

🎯 What it does: Three methods are proposed to compile finite field numerical planning into classical planning, and their feasibility is verified.

Towards Real-Time Approximate Counting

Yash Pote (National University of Singapore), Jiong Yang (Georgia Institute of Technology)

🎯 What it does: A new approximate model counting algorithm, ApproxMC7, is proposed, aimed at quickly estimating the number of satisfying solutions for Boolean formulas under real-time or low-precision requirements.

Towards Scalable and Deep Graph Neural Networks via Noise Masking

Yuxuan Liang (Peking University), Bin Cui (Wuhan University)

🎯 What it does: A pluggable RMask module is proposed to suppress over-smoothing and enhance deep propagation effects in simplified graph neural networks through noise masking and random walks.

Towards Ship License Plate Recognition in the Wild: A Large Benchmark and Strong Baseline

Baolong Liu (Zhejiang Gongshang University), Jianfeng Dong (Zhejiang Gongshang University)

🎯 What it does: A large ship license plate recognition dataset SLP34K has been constructed, and a recognition baseline based on self-supervised pre-training and semantic enhancement has been proposed.

Towards Trustable SHAP Scores

Olivier Létoffé (University of Toulouse), Joao Marques-Silva (Institute for Catalan Research and Advanced Studies)

🎯 What it does: This paper proposes several new feature functions (υ_s, υ_a, υ_c, υ_n, etc.) by modifying the feature function of SHAP, and defines the corresponding SHAP scores based on this to eliminate unreasonable feature importance results that occur in existing SHAP calculations.

Towards Trustworthy Knowledge Graph Reasoning: An Uncertainty Aware Perspective

Bo Ni (Vanderbilt University), Tyler Derr (Vanderbilt University)

🎯 What it does: This paper proposes the UAG (Uncertainty Aware Knowledge Graph Reasoning) framework, which utilizes multi-step reasoning combining KG and LLM, and calibrates the error rates of each component through confidence prediction and Learn-Then-Test, forming a theoretically guaranteed set of answers.

Towards Unifying Evaluation of Counterfactual Explanations: Leveraging Large Language Models for Human-Centric Assessments

Marharyta Domnich (Institute of Computer Science, University of Tartu), Raul Vicente (Institute of Computer Science, University of Tartu)

🎯 What it does: Collected 30 adversarial explanations, invited 206 participants to score them across 8 dimensions, and utilized a large language model (LLM) to predict human evaluations.

Toy-GS: Assembling Local Gaussians for Precisely Rendering Large-Scale Free Camera Trajectories

Xiaohan Zhang (South China University of Technology), Qi Liu (South China University of Technology)

🎯 What it does: The Toy-GS method is proposed, which utilizes adaptive spatial partitioning, Patchmatch and PPAC, as well as local-global fusion techniques to achieve high-quality rendering of large-scale free camera trajectories, significantly reducing GPU memory usage.

TPCH: Tensor-interacted Projection and Cooperative Hashing for Multi-view Clustering

Zhongwen Wang (Nanjing University of Science and Technology), Zhenwen Ren (Sichuan University)

🎯 What it does: The TPCH (Tensor-Interacted Projection and Cooperative Hashing) framework is proposed, which stacks multi-view projection matrices and hash codes into tensors, introducing high-order cooperation in the projection and Hamming space and enhancing the tensor nuclear norm to improve the density and separability of binary representations for large-scale multi-view clustering.

TRACI: A Data-centric Approach for Multi-Domain Generalization on Graphs

Yusheng Zhao (Peking University), Ming Zhang (Peking University)

🎯 What it does: This work studies domain generalization of multi-source graph data and proposes the TRACI method to enhance the generalization performance of GNNs without a target graph.

Track the Answer: Extending TextVQA from Image to Video with Spatio-Temporal Clues

Yan Zhang (Institute of Information Engineering, Chinese Academy of Sciences), Can Ma (Institute of Information Engineering, Chinese Academy of Sciences)

🎯 What it does: For the task of Video Text Visual Question Answering (Video TextVQA), a TEA framework is proposed, based on a T5 generative model, to restore the spatiotemporal relationships between scene text and visual entities in videos, and to aggregate scene text clues related to the question to enhance the quality of answer generation.

Tradutor: Building a Variety Specific Translation Model

Hugo Sousa (University of Porto), Alipio Jorge (University of Porto)

🎯 What it does: A parallel corpus of European Portuguese and English, PTradutor, was created, and an open-source machine translation model was trained based on it.

Traffic Scenario Logic: A Spatial-Temporal Logic for Modeling and Reasoning of Urban Traffic Scenarios

Ruolin Wang (University of Science and Technology of China), Jianmin Ji (University of Science and Technology of China)

🎯 What it does: This paper proposes and implements a spatial-temporal logic for pedestrian-free urban road traffic scenarios—Traffic Scenario Logic (TSL), achieving automatic conversion from OpenDRIVE high-precision maps to logical models, and generating complete traffic scenario sequences for testing, decision-making, and control validation using the ASP+ temporal logic solver Telingo.

Training Deep Neural Networks with Virtual Smoothing Classes

Zhiyang Zhou (Chinese Academy of Sciences), Yan Cai (Chinese Academy of Sciences)

🎯 What it does: A Virtual Smoothing label is proposed, which transforms the overconfidence problem into a competition between the positive class and a virtual class by adding an additional virtual category to the last layer of the classifier, thereby reducing overconfidence while maintaining confidence levels.

Training Matting Models Without Alpha Labels

Wenze Liu (Chinese University of Hong Kong), Xiangyu Yue (Huazhong University of Science and Technology)

🎯 What it does: Using a rough trimap as training labels, we propose a Direction Distance Consistency loss (DDC loss) to train a deep image matting model without fine alpha annotations.

Training on the Benchmark Is Not All You Need

Shiwen Ni (Shenzhen Institutes of Advanced Technology), Min Yang (Shenzhen Institutes of Advanced Technology)

🎯 What it does: A leakage detection method based on multiple-choice question option replacement is proposed, and the identification of benchmark test set leakage in LLM pre-training data is achieved under gray-box conditions.

Training-and-Prompt-Free General Painterly Harmonization via Zero-Shot Disentenglement on Style and Content References

Teng-Fang Hsiao (National Yang Ming Chiao Tung University), Hong-Han Shuai (National Yang Ming Chiao Tung University)

🎯 What it does: A general painting style fusion method called TF-GPH is proposed, which can achieve seamless integration of foreground and background in three types of tasks: object insertion, object exchange, and style transfer, without the need for training or prompts.

Training-Free and Hardware-Friendly Acceleration for Diffusion Models via Similarity-based Token Pruning

Evelyn Zhang (Shanghai Jiao Tong University), Linfeng Zhang (Shanghai Jiao Tong University)

🎯 What it does: This paper proposes and implements SiTo, a similarity-based token pruning method designed to accelerate the inference of diffusion models without the need for training or labeling.

Transferable Adversarial Face Attack with Text Controlled Attribute

Wenyun Li (Harbin Institute of Technology), Dongmei Jiang (Pengcheng Laboratory)

🎯 What it does: A transferable adversarial face attack method TCA2 based on natural language guidance has been developed, utilizing StyleGAN2 to generate realistic attack samples.

Transformer Layers as Painters

Qi Sun (Sakana AI), Llion Jones (Emergence AI)

🎯 What it does: This paper explores the similarity and robustness between layers by conducting various experimental variants such as layer skipping, rearrangement, parallel execution, and cyclic execution on a frozen pre-trained Transformer.

Transtreaming: Adaptive Delay-aware Transformer for Real-time Streaming Perception

Xiang Zhang (Southeast University), Weiwei Wu (Southeast University)

🎯 What it does: This paper proposes Transtreaming, a real-time streaming perception framework that achieves multi-frame prediction through an adaptive delay-aware Transformer to compensate for computational delays.

TreeEval: Benchmark-Free Evaluation of Large Language Models through Tree Planning

Xiang Li (East China Normal University), Chao Yang (Shanghai AI Laboratory)

🎯 What it does: Proposes TreeEval, a benchmark-free evaluation method for LLMs based on tree planning.

Trigger3:Refining Query Correction via Adaptive Model Selector

Kepu Zhang (Renmin University of China), Jun Xu (Renmin University of China)

🎯 What it does: Proposes the Trigger3 framework, which utilizes a three-stage trigger (error correction trigger, LLM trigger, fallback trigger) to achieve adaptive collaborative error correction between small models and large language models.

Trust-GRS: A Trustworthy Training Framework for Graph Neural Network Based Recommender Systems Against Shilling Attacks

Lingyu Mu (Institute of Information Engineering, Chinese Academy of Sciences), Zheng Lin (Institute of Information Engineering, Chinese Academy of Sciences)

🎯 What it does: This paper proposes Trust-GRS, a two-stage, prior knowledge-free GNN recommendation system training framework designed to detect and suppress fake users injected by spoof attackers.

Trusted Unified Feature-Neighborhood Dynamics for Multi-View Classification

Haojian Huang (University of Hong Kong), Zhongjiang He (TeleAI)

🎯 What it does: The TUNED model is proposed for multi-view classification, integrating the local feature-neighborhood structure of each view with global consistency, and adaptively fusing evidence through selective Markov random fields.

TrustUQA: A Trustful Framework for Unified Structured Data Question Answering

Wen Zhang (Zhejiang University), Huajun Chen (Ant Group)

🎯 What it does: This paper proposes TrustUQA, a trustworthy unified structured data question-answering framework that can simultaneously support natural language questions for tables, knowledge graphs, and temporal knowledge graphs. It uses a Condition Graph (CG) to unify the representation of different types of data and employs a two-layer functional query approach (first generating a simplified query using LLM, then converting it into an executable CG query through rules) to achieve efficient reasoning. Additionally, it introduces dynamic demonstration retrieval to enhance prompt quality and improve model performance.

TTE: Two Tokens Are Enough to Improve Parameter-Efficient Tuning

Jiacheng Ruan (Shanghai Jiao Tong University), Yuzhuo Fu (Shanghai Jiao Tong University)

🎯 What it does: This paper studies the overfitting problem in Parameter-Efficient Tuning (PET) and proposes the TTE framework, which effectively alleviates overfitting and enhances PET performance by utilizing globally learnable tokens, instance-specific tokens, and Parameter-Free Cross Attention (PFCA) loss.

Tuning-Free Accountable Intervention for LLM Deployment – a Metacognitive Approach

Zhen Tan (Arizona State University), Huan Liu (University of North Carolina at Chapel Hill)

🎯 What it does: This paper proposes the CLEAR framework, which implements a metacognitive intervention method that can automatically identify and correct errors in large language models (LLMs) during the inference phase without the need for fine-tuning.

Two-Timescale Critic-Actor for Average Reward MDPs with Function Approximation

Prashansa Panda (Indian Institute of Science), Shalabh Bhatnagar (Indian Institute of Science)

🎯 What it does: This paper proposes and analyzes the first two-time-scale Critic-Actor algorithm for average reward MDPs using linear function approximation, providing proofs of non-asymptotic and asymptotic convergence, and determining the optimal learning rate and sample complexity.

Twofold Debiasing Enhances Fine-Grained Learning with Coarse Labels

Xin-yang Zhao (Nanjing University of Science and Technology), Yazhou Yao (Nanjing University of Science and Technology)

🎯 What it does: For the Coarse-to-Fine Few-Shot task, a Twofold Debiasing (TFB) method is proposed, which enhances fine-grained representation through multi-layer feature fusion reconstruction and intermediate layer feature alignment, and calibrates the fine-grained classifier distribution using features from the coarse label training set.

U-KAN Makes Strong Backbone for Medical Image Segmentation and Generation

Chenxin Li (Chinese University of Hong Kong), Yixuan Yuan (Chinese University of Hong Kong)

🎯 What it does: This paper proposes a new framework called U-KAN, which embeds the Kolmogorov-Arnold Network (KAN) into the U-Net backbone for medical image segmentation and generation.

UAWTrack: Universal 3D Single Object Tracking in Adverse Weather

Yuxiang Yang (Hangzhou Dianzi University), Jing Zhang (Wuhan University)

🎯 What it does: This paper proposes UAWTrack, a unified model capable of 3D single-object tracking under various weather conditions.

UCF-Crime-DVS: A Novel Event-Based Dataset for Video Anomaly Detection with Spiking Neural Networks

Yuanbin Qian (Ningbo University), Jiafei Wu (The University of Hong Kong)

🎯 What it does: The first video anomaly detection dataset based on event cameras, UCF-Crime-DVS, has been constructed, and a multi-scale fusion framework (MSF) based on Spiking Neural Networks (SNN) has been proposed for weakly supervised video anomaly detection.

UFO: Enhancing Diffusion-Based Video Generation with a Uniform Frame Organizer

Delong Liu (Beijing University of Posts and Telecommunications), Fei Su (Beijing University of Posts and Telecommunications)

🎯 What it does: The UFO plugin is proposed, utilizing a lightweight non-intrusive adapter to enhance the consistency and quality of video generation based on diffusion models;

UN-DETR: Promoting Objectness Learning via Joint Supervision for Unknown Object Detection

HaoMiao Liu, Bo Ma (Beijing Institute of Technology)

🎯 What it does: This paper proposes an unknown object detection framework called UN-DETR based on Transformer, which achieves end-to-end unknown object detection by combining Instance Presence Score (IPS) predictor, one-to-many assignment, unbiased query selection, IPS-guided post-processing, and unsupervised pre-training.

Unaligned Message-Passing and Contextualized-Pretraining for Robust Geo-Entity Resolution

Yuwen Ji (Beihang University), Yue Zhang (Westlake University)

🎯 What it does: This paper proposes a BERT-based encoder that incorporates Unaligned Message Passing (UMP) and Contextual Pre-training (CP) to enhance the robustness of geographic entity matching through neighborhood geographic context.

UniDemoiré: Towards Universal Image Demoiréing with Data Generation and Synthesis

Zemin Yang (ShanghaiTech University), Yuexin Ma (ShanghaiTech University)

🎯 What it does: This paper proposes a unified image de-moiré framework called UniDemoir, which achieves generalization for various moiré patterns through large-scale data generation and synthesis.

Unified Coding for Both Human Perception and Generalized Machine Analytics with CLIP Supervision

Kangsheng Yin (Shenzhen University), Shiqi Wang (City University of Hong Kong)

🎯 What it does: This paper proposes a unified and general image coding framework UG-ICM, which utilizes a single bitstream to meet the needs of human visual perception and unknown machine vision analysis through conditional decoding, and achieves self-supervised training via cross-modal supervision provided by CLIP.

Unified Knowledge Maintenance Pruning and Progressive Recovery with Weight Recalling for Large Vision-Language Models

Zimeng Wu (Beihang University), Yunhong Wang (Beihang University)

🎯 What it does: This paper proposes a unified structured pruning framework for large-scale vision-language models (LVLM) called UKMP, aimed at significantly reducing model parameters and computational load while maintaining zero-shot performance.

UniMuMo: Unified Text, Music, and Motion Generation

Han Yang (Chinese University of Hong Kong), Chuang Gan (Cisco Research)

🎯 What it does: A unified multimodal generation framework called UniMuMo has been constructed, capable of generating combinations of text, music, and actions.

Union Is Strength! Unite the Power of LLMs and MLLMs for Chart Question Answering

Jiapeng Liu (Institute of Information Engineering, Chinese Academy of Sciences), Can Ma (Institute of Information Engineering, Chinese Academy of Sciences)

🎯 What it does: A framework named SYNERGY is proposed, which integrates large language models (LLM) and multimodal large language models (MLLM) to accomplish chart question answering tasks (CQA), achieving more accurate answer generation through phased processing.

UniTR: A Unified Framework for Joint Representation Learning of Trajectories and Road Networks

Jie Zhao (Chongqing University), Yuxuan Liang (Hong Kong University of Science and Technology)

🎯 What it does: The UniTR framework is proposed to achieve joint representation learning of road networks and trajectories, utilizing a hierarchical propagation mechanism and three-layer contrastive learning to simultaneously optimize the representations of both types of data within the same embedding space.

Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment

Yuanfan Zheng (Institute of Automation, Chinese Academy of Sciences), Zhen Chen (Chinese University of Hong Kong)

🎯 What it does: Proposes the Dual Probabilistic Alignment (DPA) framework, which targets unified domain adaptive object detection by performing probabilistic alignment for global domain private categories and instance-level domain shared categories, and reduces negative transfer through private category constraints.

Universal Post-Processing Networks for Joint Optimization of Modules in Task-Oriented Dialogue Systems

Atsumoto Ohashi (Nagoya University), Ryuichiro Higashinaka (Nagoya University)

🎯 What it does: This paper studies a unified post-processing network (UniPPN) designed to jointly optimize the outputs of all modules in task-oriented dialogue systems, thereby improving the task completion rate of the system.

Universality of Real Minimal Complexity Reservoir

Robert Simon Fong (University of Birmingham), Peter Tino (New Mexico State University)

🎯 What it does: It is proven that Simple Cycle Reservoirs (SCR) can approximate any time-invariant dynamic filter with decaying memory over the real number field, filling a gap that was previously only known in the complex number field.

Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient

Yongliang Wu (Southeast University), Xu Yang (Nanyang Technological University)

🎯 What it does: This paper proposes a conceptual forgetting framework called DoCo for text-to-image diffusion models, aimed at thoroughly removing sensitive concepts while maintaining the overall performance of the model.

Unleashing the Potential of Model Bias for Generalized Category Discovery

Wenbin An (Xi'an Jiaotong University), Ping Chen (University of Massachusetts Boston)

🎯 What it does: Proposes the Self-Debiasing Calibration (SDC) framework, which utilizes the outputs of a pre-trained bias model to dynamically calibrate logits, generating more accurate pseudo-labels, thereby improving the recognition of new categories in the General Category Discovery (GCD) task.

Unleashing the Power of Visual Foundation Models for Generalizable Semantic Segmentation

PeiYuan Tang (Xi'an Jiaotong University), Zijiang James Yang (Singapore Management University)

🎯 What it does: The study utilizes visual foundation models for domain generalization semantic segmentation.

Unleashing the Temporal-Spatial Reasoning Capacity of GPT for Training-Free Audio and Language Referenced Video Object Segmentation

Shaofei Huang (Hefei University of Technology), Si Liu (Meituan)

🎯 What it does: A training-free Audio-Language-Referenced SAM 2 (AL-Ref-SAM 2) pipeline is proposed to achieve multi-modal video object segmentation (AVS and RVOS) with audio and language cues;

Unlocking the Game: Estimating Games in Möbius Representation for Explanation and High-Order Interaction Detection

Majid Mohammadi (Vrije Universiteit Amsterdam), Annette Ten Teije (Vrije Universiteit Amsterdam)

🎯 What it does: The GEM-FIX method is proposed, which allows for the precise calculation of Shapley values and efficient identification of significant high-order feature interactions by performing regression on the explanation game using Möbius representation.

Unlocking the Potential of Black-box Pre-trained GNNs for Graph Few-shot Learning

Qiannan Zhang (Cornell University), Xiangliang Zhang (University of Notre Dame)

🎯 What it does: This study investigates how to achieve few-shot learning for graph nodes in a black-box pre-trained Graph Neural Network (GNN) environment, proposing a lightweight meta-learner to extract task-related knowledge and utilizing sub-networks for rapid adaptation.

Unlocking the Potential of Reverse Distillation for Anomaly Detection

Xinyue Liu (Beihang University), Shuo Zhang (Beijing Jiaotong University)

🎯 What it does: This paper proposes an unsupervised anomaly detection method that incorporates an expert network and guided information injection within a reverse knowledge distillation framework, which can simultaneously enhance the teacher network's sensitivity to anomalies and the student network's denoising capability.

Unlocking the Power of LSTM for Long Term Time Series Forecasting

Yaxuan Kong (University of Oxford), Qingsong Wen (Squirrel AI)

🎯 What it does: A long-term forecasting model for time series based on sLSTM is proposed—P-sLSTM, which addresses the short memory problem of traditional LSTM by utilizing patching and channel independence techniques to enhance memory and generalization capabilities.

Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting

Peiwang Tang (University of Science and Technology of China), Weitai Zhang (University of Science and Technology of China)

🎯 What it does: This paper proposes a Patch-based MLP (PatchMLP) model that utilizes patch embedding and linear layers to predict long-period time series.

Unravelling Causal Genetic Biomarkers of Alzheimer’s Disease via Neuron to Gene-token Backtracking in Neural Architecture: A Groundbreaking Reverse-Gene-Finder Approach

Victor O. K. Li (University of Hong Kong), Jacqueline C. K. Lam (University of Hong Kong)

🎯 What it does: Based on the large gene pre-training model Geneformer, combined with single-cell microglial expression data, we propose the Reverse-Gene-Finder method, which first fine-tunes the model for early classification of Alzheimer's disease, then identifies the most causative neurons (MCN) through causal tracing, and finally traces back to the input layer to obtain the most causative gene markers (MCT) and their corresponding genes (MCG), thereby discovering several new candidate genes previously unrecognized as AD-related.

Unsupervised Anomaly Detection for Tabular Data Using Deep Noise Evaluation

Wei Dai (Chinese University of Hong Kong), Jicong Fan (Chinese University of Hong Kong)

🎯 What it does: A noise assessment-based unsupervised anomaly detection method is proposed, which directly trains the model using samples of normal data with added noise to learn the anomaly boundary.

Unsupervised Degradation Representation Aware Transform for Real-World Blind Image Super-Resolution

Sen Chen (Xidian University), Yaowei Wang (Tianjin University)

🎯 What it does: This paper designs an unsupervised degradation representation-aware transformation network (DRAT), which filters image content through a dual encoder to accurately extract degradation representations and applies them to feature transformation and global degradation-aware blocks, achieving superior blind image super-resolution.

Unsupervised Diffusion-Based Degradation Modeling for Real-World Super-Resolution

Yuying Chen (Shenzhen Campus of Sun Yat-sen University), Wenqi Ren (Shenzhen Campus of Sun Yat-sen University)

🎯 What it does: An unsupervised diffusion degradation modeling framework (UDDM) is proposed, which maps real low-resolution images to extremely low resolution through extreme downsampling, learns the degradation distribution, and generates content-aware low-resolution images using a physical dynamic degradation module (P-DDM), ultimately synthesizing HR-LR training pairs consistent with the real distribution for training SISR models.

Unsupervised Domain Adaptive Person Search via Dual Self-Calibration

Linfeng Qi (Dalian Maritime University), Yang Wang (Hefei University of Technology)

🎯 What it does: A dual self-calibration framework is proposed to eliminate the interference of pseudo-label noise in unsupervised domain adaptation for person search.

Unsupervised Photometric-Consistent Depth Estimation from Endoscopic Monocular Video

Shijie Li (Sichuan University), Zheng Li (University of Chinese Academy of Sciences)

🎯 What it does: This paper proposes an unsupervised endoscopic monocular video depth estimation method called PC-Depth, which improves depth prediction accuracy by addressing the issue of inconsistent lighting.

Unsupervised Self-Prior Embedding Neural Representation for Iterative Sparse-View CT Reconstruction

Xuanyu Tian (ShanghaiTech University), Yuyao Zhang (Shanghai Jiao Tong University)

🎯 What it does: An iterative self-supervised sparse view CT reconstruction method called Spener is proposed, which does not require external training data. It utilizes the prior features of the image from the previous iteration embedded in an implicit neural representation network to achieve robust reconstruction in extremely under-sampled and noisy scenarios.

Unveiling Multi-View Anomaly Detection: Intra-view Decoupling and Inter-view Fusion

Kai Mao (Xi'an Jiaotong University), Ping Wei (Xi'an Jiaotong University)

🎯 What it does: A multi-view anomaly detection framework IDIF is proposed, which achieves anomaly recognition for multi-view data through two steps: view decoupling and view fusion.

Unveiling the Threat of Fraud Gangs to Graph Neural Networks: Multi-Target Graph Injection Attacks Against GNN-Based Fraud Detectors

Jinhyeok Choi (KAIST), Joyce Jiyoung Whang (KAIST)

🎯 What it does: This paper studies multi-target graph injection attacks on fraud detectors based on graph neural networks and proposes a new attack model, MonTi, to inject attack nodes at once and induce fraudulent nodes to be misclassified as normal.

UrBench: A Comprehensive Benchmark for Evaluating Large Multimodal Models in Multi-View Urban Scenarios

Baichuan Zhou (Shanghai AI Laboratory), Weijia Li (Sun Yat-Sen University)

🎯 What it does: This paper presents UrBench, a multi-view, multi-task urban scene evaluation benchmark designed to test the performance of large-scale multimodal models (LMM) in urban environments.

USDRL: Unified Skeleton-Based Dense Representation Learning with Multi-Grained Feature Decorrelation

Wanjiang Weng (Southeast University), Guo-Sen Xie (Nanjing University of Science and Technology)

🎯 What it does: A USDRL framework is proposed, utilizing multi-layer feature decorrelation to achieve dense representation learning of skeleton sequences.

Utilize the Flow Before Stepping into the Same River Twice: Certainty Represented Knowledge Flow for Refusal-Aware Instruction Tuning

Runchuan Zhu (Shanghai Artificial Intelligence Laboratory), Conghui He (Shanghai Artificial Intelligence Laboratory)

🎯 What it does: To address the phenomenon of over-refusal in large language models during the RAIT process, the CRaFT method is proposed, which reduces static and dynamic conflicts by incorporating answer confidence screening and rehearsal training, thereby enhancing the reliability of the model's refusals.

V2C-CBM: Building Concept Bottlenecks with Vision-to-Concept Tokenizer

Hangzhou He (Peking University), Yanye Lu (Peking University)

🎯 What it does: This study proposes V2C-CBM, which utilizes a Vision-to-Concept (V2C) tokenizer to directly generate visual concepts from unlabeled images, constructing a concept bottleneck model to achieve interpretable image classification without the need for LLMs.

VE-Bench: Subjective-Aligned Benchmark Suite for Text-Driven Video Editing Quality Assessment

Shangkun Sun (Peking University), Wei Gao (Peking University)

🎯 What it does: A VE-Bench benchmark has been constructed, which includes a subjective alignment database (VE-Bench DB) specifically designed for text-driven video editing and a quantitative evaluation network (VE-Bench QA) based on this database, aimed at comprehensive assessment of the quality of edited videos.

VEGAS: Towards Visually Explainable and Grounded Artificial Social Intelligence

Hao Li (Wuhan University), Zheng Wang (Wuhan University)

🎯 What it does: Proposes the VEGAS framework, which combines visual interpretability with language-guided frame sampling to generate interpretable multimodal answers.

VerilogCoder: Autonomous Verilog Coding Agents with Graph-based Planning and Abstract Syntax Tree (AST)-based Waveform Tracing Tool

Chia-Tung Ho (NVIDIA Research), Brucek Khailany (NVIDIA Research)

🎯 What it does: This paper presents VerilogCoder, a multi-agent system capable of automatically generating Verilog code and fixing functional errors through syntax checking, simulation, and a novel AST waveform tracing tool.

VERO: Verification and Zero-Shot Feedback Acquisition for Few-Shot Multimodal Aspect-Level Sentiment Classification

Kai Sun (Xi'an Jiaotong University), Bo Dong (Xi'an Jiaotong University)

🎯 What it does: A self-verification-based sample collection method called VERO was designed and implemented to select challenging samples from unlabeled multimodal data, followed by few-shot fine-tuning of large visual-language models (LLaVA-7b/13b) to complete the multimodal aspect-level sentiment classification task.

VersaGen: Unleashing Versatile Visual Control for Text-to-Image Synthesis

Zhipeng Chen (Beijing University of Posts and Telecommunications), Yi-Zhe Song (University of Surrey)

🎯 What it does: Proposes VersaGen, a control framework for text-to-image synthesis that can be guided by user-drawn inputs at multiple levels.

VERSE: Verification-based Self-Play for Code Instructions

Hao Jiang (University of Science and Technology of China), Yu Su (Hefei Normal University)

🎯 What it does: The VERSE method is proposed, which utilizes code LLMs to perform code verification during the self-generation of instructions, responses, and validation scripts, and selects high-quality data for self-fine-tuning based on execution results and self-consistency scores.

VG-TVP: Multimodal Procedural Planning via Visually Grounded Text-Video Prompting

Muhammet Furkan Ilaslan (National University of Singapore), Qianli Xu (Institute for Infocomm Research, Agency for Science, Technology, and Research)

🎯 What it does: This paper proposes a multimodal program planning framework called VG‑TVP based on large language models, which can generate coherent text and video program steps according to user-defined task objectives, addressing the limitations of traditional unimodal planning.

VHM: Versatile and Honest Vision Language Model for Remote Sensing Image Analysis

Chao Pang (Wuhan University), Conghui He (Shanghai Artificial Intelligence Laboratory)

🎯 What it does: A multi-task visual language model VHM aimed at remote sensing image analysis has been developed, and it achieves multifunctional understanding and honest responses by combining a newly constructed multi-content description dataset VersaD and a truthfulness instruction set HnstD containing both factual and deceptive questions.

Video Repurposing from User Generated Content: A Large-scale Dataset and Benchmark

Yongliang Wu (Southeast University), Xu Yang (Opus AI Research)

🎯 What it does: A research framework for video repurposing tasks is proposed, and a large-scale user-generated content (UGC) video repurposing dataset, Repurpose-10K, is constructed. An end-to-end baseline model based on a multi-modal Transformer is developed, capable of automatically generating short video clips of about 60 seconds from long videos.

VideoElevator: Elevating Video Generation Quality with Versatile Text-to-Image Diffusion Models

Yabo Zhang (Harbin Institute of Technology), Wangmeng Zuo (Harbin Institute of Technology)

🎯 What it does: A training-agnostic, plug-and-play method called VideoElevator is proposed, which utilizes existing text-to-image diffusion models to enhance the quality of text-to-video diffusion models. This method explicitly splits each sampling step into two sub-steps: temporal motion refinement and spatial quality enhancement.

ViFactCheck: A New Benchmark Dataset and Methods for Multi-Domain News Fact-Checking In Vietnamese

Tran Thai Hoa (University of Information Technology), Kiet Van Nguyen (University of Information Technology)

🎯 What it does: This study first constructed the ViFactCheck dataset, a Vietnamese multi-domain fact-checking benchmark translated from Chinese, and conducted model training and evaluation on this dataset.

ViG: Linear-complexity Visual Sequence Learning with Gated Linear Attention

Bencheng Liao (Huazhong University of Science and Technology), Chang Huang (Horizon Robotics)

🎯 What it does: ViG proposes a visual sequence learning framework with linear complexity.

VIoTGPT: Learning to Schedule Vision Tools Towards Intelligent Video Internet of Things

Yaoyao Zhong (Beijing University of Posts and Telecommunications), Huadong Ma (Beijing University of Posts and Telecommunications)

🎯 What it does: A framework called VIoTGPT based on large language models is proposed, which can automatically invoke video tools according to human queries, enabling unified and intelligent analysis of large-scale videos in the Video Internet of Things (VIoT) and providing responses.

Virtual Nodes Can Help: Tackling Distribution Shifts in Federated Graph Learning

Xingbo Fu (University of Virginia), Jundong Li (University of Virginia)

🎯 What it does: The FedVN framework is proposed, which eliminates the shift caused by inconsistent data distribution in federated graph learning by learning adaptive graph enhancement strategies for each client (using shared multiple virtual nodes and personalized edge generators), thereby achieving efficient training of the global GNN model.