🎯 What it does: A self-supervised pre-training framework based on Bird's Eye View Masked Autoencoder (BEV-MAE) is proposed for 3D object detection of LiDAR point clouds in autonomous driving scenarios.
Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory
Aydar Bulatov (Moscow Institute of Physics and Technology), Mikhail Burtsev (London Institute for Mathematical Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the Recurrent Memory Transformer (RMT), which inserts learnable memory tokens into the input/output of the Transformer and implements recursive connections at the segment level, allowing for a linear expansion of the input context length up to 2 million tokens.
Beyond Entities: A Large-Scale Multi-Modal Knowledge Graph with Triplet Fact Grounding
Jingping Liu (East China University of Science and Technology), Yunwen Chen (DataGrand Inc.)
CodeContrastive LearningImageTextMultimodality
🎯 What it does: This paper constructs ImgFact, a multimodal knowledge graph based on images, which records 247,732 triple facts and 3,730,805 related images.
🎯 What it does: This paper proposes the task of extracting fine-grained event hierarchical structures from two modalities: video and text, and constructs a new MultiHiEve dataset. It introduces a weakly supervised model called MASHER to accomplish this task.
Beyond Prototypes: Semantic Anchor Regularization for Better Representation Learning
Yanqi Ge (Shenzhen Institute for Advanced Study University of Electronic Science and Technology of China), Lixin Duan (Shenzhen Institute for Advanced Study University of Electronic Science and Technology of China)
CodeSegmentationRepresentation LearningImage
🎯 What it does: A Semantic Anchor Regularization (SAR) is designed to guide feature learning through predefined class anchors, reducing prototype learning bias and enhancing inter-class separation and intra-class compactness.
🎯 What it does: This paper proposes the Bi-directional Adapter for Multi-modal Tracking (BAT) framework, which utilizes a lightweight bidirectional adapter to achieve dynamic information mutual promotion and fusion between RGB and infrared (TIR) dual modalities within a pre-trained Transformer backbone.
🎯 What it does: This paper proposes a bias elimination framework based on adversarial training called BSSARD, which dynamically synthesizes bias conflict samples and breaks the pseudo-correlation between temporal position and semantics through a visual and textual bias generator to enhance the generalization ability of video sentence localization.
🎯 What it does: This paper proposes the Bidirectional Temporal Plan Graph (BTPG), which allows for dynamic switching of the passing order at conflict points during execution to reduce meaningless waiting and improve the execution efficiency of multi-agent path planning.
Yulai Cong (Sun Yat-sen University), Sijia Li (Sun Yat-sen University)
CodeOptimizationTabular
🎯 What it does: A BigLearn-EM algorithm based on the principle of large learning is proposed, improving the EM training of GMM and overcoming the issue of poor local optima.
Binding-Adaptive Diffusion Models for Structure-Based Drug Design
Zhilin Huang (Tsinghua University), Wenming Yang (Peng Cheng Laboratory)
CodeDrug DiscoveryGraph Neural NetworkDiffusion modelGraphBiomedical Data
🎯 What it does: A binding adaptive diffusion model based on BINDDM is proposed for structured drug design, which automatically extracts key sub-complexes from protein-ligand complexes and enhances the generation process during diffusion.
BiPFT: Binary Pre-trained Foundation Transformer with Low-Rank Estimation of Binarization Residual Polynomials
Xingrun Xing (Institute of Automation, Chinese Academy of Sciences), Jiajun Zhang (Institute of Automation, Chinese Academy of Sciences)
CodeComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This paper proposes the first binary pre-trained foundational transformer for natural language understanding tasks, BiPFT, which significantly reduces computational load and memory usage by enhancing self-attention performance through low-rank estimation of binary residual polynomials.
BLIVA: A Simple Multimodal LLM for Better Handling of Text-Rich Visual Questions
Wenbo Hu (University of California), Zhuowen Tu (University of California)
CodeRecognitionGenerationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: We propose BLIVA, a multimodal large language model that integrates learned query embeddings with patch embeddings from an image encoder, enabling better understanding of textual information within images.
🎯 What it does: A block-level image compression sensing method called BRBCN is proposed, which combines CNN local reconstruction with Transformer global communication to achieve bidirectional information interaction both within and between blocks.
🎯 What it does: This paper proposes an input transformation attack method based on elastic constraint deformation (DeCoW) called DeCoWA, which utilizes max-min optimization combined with multi-deformation averaging to enhance the transferability of adversarial samples across model families (such as CNN and ViT) and across modalities (images, videos, audio).
Boosting Multiple Instance Learning Models for Whole Slide Image Classification: A Model-Agnostic Framework Based on Counterfactual Inference
Weiping Lin (Xiamen University), Liansheng Wang (University of Hong Kong)
CodeClassificationImage
🎯 What it does: A model-agnostic MIL enhancement framework CIMIL is proposed, which obtains high-quality pseudo-labels through counterfactual inference sub-bag evaluation and hierarchical instance search, trains an instance classifier, and uses its embeddings as prompts to improve the original features, thereby enhancing the bag-level and instance-level classification performance of WSI.
Boosting Neural Cognitive Diagnosis with Student’s Affective State Modeling
Shanshan Wang (Anhui University), Xingyi Zhang (Anhui University)
CodeContrastive LearningTabular
🎯 What it does: Affective Cognitive Diagnosis Model (ACD) is proposed, which enhances the accuracy of cognitive diagnosis by predicting students' emotional states during the answering process and converting them into personalized guessing and slipping parameters.
🎯 What it does: A training framework based on grouped knowledge (GKT) is proposed to enhance the performance of residual networks through the aggregation and transfer of hierarchical subnetwork groups.
🎯 What it does: This paper proposes an iterative polygon regression framework Box2Poly based on Sparse R‑CNN, which gradually regresses to text polygons using learnable box priors, and introduces Bézier curves as an intermediate representation along with PolyAlign for efficient RoI feature extraction.
🎯 What it does: A training-free Diff-Text framework is proposed, capable of generating text images for any language scene by utilizing pre-rendered sketches and edge maps as control conditions, combined with text prompts to generate realistic scene images.
🎯 What it does: Proposes the Learngene Pool method, which constructs variable-sized networks using learngene instances distilled from a single large model, addressing issues of storage consumption, minimum model size limitations, and insufficient initialization of stitch layers in SN-Net.
🎯 What it does: A single-round federated learning algorithm is proposed, utilizing Bayesian inference to aggregate local predictive posteriors in the prediction space for model calibration and performance improvement under heterogeneous data.
Can Large Language Models Understand Real-World Complex Instructions?
Qianyu He (Fudan University), Yanghua Xiao (Fudan University)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A benchmark named CELLO has been designed and released to systematically evaluate the understanding capabilities of large language models for complex instructions.
🎯 What it does: The CARAT framework is proposed, which achieves multi-label emotion recognition through the reconstruction fusion and feature aggregation of multimodal information.
🎯 What it does: A framework is proposed and validated that inserts a Feature Calibration Module (FCM) into unsupervised object re-identification based on clustering, enhancing the synergy between pseudo-label generation and representation learning.
CatFormer: Category-Level 6D Object Pose Estimation with Transformer
Sheng Yu (Beijing Institute of Technology), Yuanqing Xia (Beijing Institute of Technology)
CodePose EstimationTransformerPoint Cloud
🎯 What it does: This paper proposes a category-level 6D object pose estimation framework based on Transformer, called CatFormer, which includes coarse and fine deformation modules, a graphical module, and a recursive refinement module, capable of accurately recovering the target's NOCS model and estimating its pose from RGB-D input;
Causal Strategic Learning with Competitive Selection
Kiet Q. H. Vo (CISPA Helmholtz Center for Information Security), Krikamol Muandet (CISPA Helmholtz Center for Information Security)
CodeTabular
🎯 What it does: This study investigates causal policy learning in a multi-decision-maker environment, proposing a competitive selection framework and designing an optimal selection rule that balances optimization and incentives during the selection process. It also introduces a cooperation protocol and Mean Shift Linear Regression (MSLR) to achieve unbiased causal parameter estimation.
🎯 What it does: This paper proposes a causal inference method based on front-door adjustment called Causal Walk, aimed at debiasing multi-hop fact verification.
Cautiously-Optimistic Knowledge Sharing for Cooperative Multi-Agent Reinforcement Learning
Yanwen Ba (Hunan University), Shigeng Zhang (Central South University)
CodeReinforcement LearningSequential
🎯 What it does: This paper proposes a cautious optimistic knowledge sharing framework called CONS, which is used for sharing positive and negative experiences and performing soft policy updates and targeted exploration in decentralized training and execution of multi-agent reinforcement learning.
🎯 What it does: We propose Ced-NeRF, a dynamic NeRF framework that combines high-resolution meshes with deformation networks, capable of achieving fast training and high-quality rendering with only a minimal number of additional parameters.
🎯 What it does: This paper proposes the Cell Graph Transformer (CGT), which classifies cell nucleus types by using both the cell nucleus and edges as input tokens and leveraging learnable adjacency relationships.
CFEVER: A Chinese Fact Extraction and VERification Dataset
Ying-Jia Lin (National Cheng Kung University), Hung-Yu Kao (National Cheng Kung University)
CodeRetrievalTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: Created the largest Chinese fact extraction and verification dataset CFEVER, which contains 30,012 manually annotated propositions based on Chinese Wikipedia and corresponding evidence.
Chain-of-Thought Improves Text Generation with Citations in Large Language Models
Bin Ji (National University of Singapore), See-Kiong Ng (National University of Singapore)
CodeGenerationRetrievalTransformerLarge Language ModelTextBenchmarkRetrieval-Augmented GenerationChain-of-Thought
🎯 What it does: This study investigates the enhancement of the correctness and citation quality of large language models in generating text with citations through Chain-of-Thought (CoT) prompts and a Citation Insurance Mechanism (CIM); systematic experiments were conducted on six major open-source LLMs.
Chinese Spelling Correction as Rephrasing Language Model
Linfeng Liu (Shanghai Jiao Tong University), Hai Zhao (Shanghai Jiao Tong University)
CodeGenerationTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: This paper studies Chinese spelling correction and proposes a rewritten language model (ReLM) that changes the model training objective to semantic rewriting of the input sentences, thereby replacing traditional character-level labeling methods.
🎯 What it does: This study investigates backdoor attacks in Split Learning on the server side and proposes a three-stage framework (Steal, Finetune, Implant) called SFI, which injects backdoors without accessing client data and models.
CK12: A Rounded K12 Knowledge Graph Based Benchmark for Chinese Holistic Cognition Evaluation
Weihao You (Tomorrow Advancing Life), Jinfeng Bai (Tomorrow Advancing Life)
CodeTransformerLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: A Chinese K12 assessment benchmark CK12 based on knowledge graphs has been constructed, covering 9 subjects, 584 primary knowledge points, and 1989 secondary knowledge points, totaling 39,452 questions. Additionally, 8 mainstream LLMs (including Chinese and English models) were evaluated under four prompting modes, and 7,334 multi-step reasoning questions were provided, with the CoT process assessed using interpretability metrics.
CodeObject DetectionRepresentation LearningVision Language ModelContrastive LearningImageText
🎯 What it does: By stitching multiple images into a mosaic and treating each image as a pseudo-region, contrastive learning is used to achieve region-text alignment, thereby enhancing open vocabulary object detection and visual-language model region representation without the need for box annotations.
🎯 What it does: This paper proposes a federated learning framework CLIP2FL for heterogeneous and long-tail data scenarios, utilizing CLIP for knowledge distillation and prototype contrastive learning to enhance the generalization ability of both server and client models.
Collaborative Consortium of Foundation Models for Open-World Few-Shot Learning
Shuai Shao (Zhejiang Lab), Bin Liu (Zhejiang Lab)
CodeClassificationTransformerLarge Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes an open-world few-shot learning framework CO3 based on a collaborative foundational model, which can achieve high-accuracy classification in scenarios with severe label noise.
🎯 What it does: A new transferable sparse input neural radiance field model, ColNeRF, is proposed, which achieves high-quality novel view rendering under sparse viewpoints by utilizing collaborative fusion between input views and self-supervised constraints at the output layer.
Mengyao Cui (University of Hong Kong), Xuelong Li (Northwestern Polytechnical University)
CodeRestorationData SynthesisTransformerImage
🎯 What it does: A framework for HDR reconstruction using a combination of color event cameras and single-exposure LDR images is proposed, along with the construction of corresponding synthetic and real datasets.
Colored Noise in PPO: Improved Exploration and Performance through Correlated Action Sampling
Jakob Hollenstein (University of Innsbruck), Justus Piater (University of Innsbruck)
CodeReinforcement LearningSequential
🎯 What it does: This paper introduces correlated (colored) noise into PPO, and experiments verify that it can significantly enhance exploration and performance;
🎯 What it does: A framework for clean-label backdoor attacks based on alternating training, called COMBAT, is proposed, which utilizes a generator to produce low-frequency, blurred triggers and optimizes them together with a surrogate model.
🎯 What it does: This paper proposes a framework called COMBHELPER, which uses Graph Neural Networks (GNN) to predict which nodes are likely to appear in the solutions of combinatorial optimization (CO), thereby pruning the search space and running traditional CO algorithms (such as linear programming, greedy algorithms, and local search) on the pruned subgraphs, significantly improving solving efficiency.
🎯 What it does: The COMMA method is proposed, which enhances CLIP's performance on zero-shot generalization tasks by generating prompts that are interrelated between the visual and language branches and constraining the feature gap between learned prompts and pre-trained prompts.
Commonsense for Zero-Shot Natural Language Video Localization
Meghana Holla (Virginia Tech), Ismini Lourentzou (University of Illinois at Urbana Champaign)
CodeGraph Neural NetworkVideoText
🎯 What it does: Developed the CORONET framework, which enhances pseudo-queries and video representations through consensus knowledge in zero-shot natural language video localization tasks, achieving cross-modal alignment.
🎯 What it does: This paper proposes DPFormer, an end-to-end Transformer framework that utilizes Douglas-Peucker (DP) points for high-precision and compact HD map construction.
Comparing the Robustness of Modern No-Reference Image- and Video-Quality Metrics to Adversarial Attacks
Anastasia Antsiferova (Moscow State University Institute for Artificial Intelligence), Dmitriy Vatolin (Moscow State University Institute for Artificial Intelligence)
CodeAdversarial AttackImageVideoBenchmark
🎯 What it does: This paper establishes a benchmark to evaluate the robustness of 15 no-reference image/video quality assessment (NR IQA/VQA) metrics under various adversarial attacks (FGSM, I-FGSM, MI-FGSM, UAP, etc.).
Composing Biases by Using CP to Decompose Minimal Functional Dependencies for Acquiring Complex Formulae
Ramiz Gindullin (IMT Atlantique), Claude-Guy Quimper (Universite Laval)
CodeTabular
🎯 What it does: A decomposition method based on constraint programming is proposed, which gradually breaks down minimal functional dependencies and combines various learning biases to learn complex formulas.
Composite Active Learning: Towards Multi-Domain Active Learning with Theoretical Guarantees
Guang-Yuan Hao (Hong Kong University of Science and Technology), Hao Wang (Rutgers University)
CodeClassificationDomain AdaptationImage
🎯 What it does: A multi-domain active learning framework called Composite Active Learning (CAL) is proposed, which constructs proxy domains by first estimating inter-domain similarity, then allocates labeling budgets in each domain, and combines instance-level query strategies to enhance overall classification performance.
Compositional Generalization for Multi-Label Text Classification: A Data-Augmentation Approach
Yuyang Chai (Wuhan University), Chong Teng (Wuhan University)
CodeClassificationGenerationTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study investigates the combinatorial generalization (CG) problem in multi-label text classification, proposing a specialized training/testing split and evaluation metrics. It introduces two generative models based on label representation disentanglement (LS-PT and LD-VAE) for data augmentation to enhance the model's ability to recognize rare combinatorial labels.
🎯 What it does: This paper proposes a Compositional Inversion method that improves the text inversion of the Stable Diffusion model using semantic anchoring and spatial regularization, addressing the overfitting and concept dominance issues caused by the original inversion.
🎯 What it does: The BoxNet module is proposed to predict the bounding boxes of each entity during the sampling process of Stable Diffusion, and based on these boxes, unique mask controls are applied to cross-attention and self-attention, achieving multi-entity and attribute semantic consistent synthesis from text to image.
🎯 What it does: This paper proposes the Compound Text-Guided Prompt Tuning (TGP-T) method, which utilizes category-level and content-level text supervision to optimize prompts, significantly reducing GPU memory usage and improving few-shot visual classification performance.
CodeRepresentation LearningData-Centric LearningGraph Neural NetworkTransformerContrastive LearningMultimodalityBiomedical Data
🎯 What it does: An unsupervised single-cell multimodal integration method called CoVEL is proposed, which can learn unified embeddings from three perspectives (regulatory relationships, cell fine-grained features, and intercellular similarity) to address the information loss problem caused by feature space mismatches between different modalities.
Computing Nash Equilibria in Potential Games with Private Uncoupled Constraints
Nikolas Patris (University of California), Ioannis Panageas (University of California)
CodeOptimizationGraph
🎯 What it does: This paper proposes a distributed gradient descent algorithm IGDλ based on the regularized Lagrangian function, aimed at solving the ε-approximate Nash equilibrium of potential games with private convex constraints.
Concept-Guided Prompt Learning for Generalization in Vision-Language Models
Yi Zhang (Harbin Institute of Technology), Zhihai He (Carnegie Mellon University)
CodeClassificationDomain AdaptationRepresentation LearningTransformerPrompt EngineeringVision Language ModelContrastive LearningImageMultimodality
🎯 What it does: This paper proposes the Concept-Guided Prompt Learning (CPL) scheme, which constructs a visual concept cache and projects multi-layer visual features into the text space to leverage the pre-trained knowledge of CLIP and generate fine-grained concept prompts, thereby enhancing the model's cross-domain generalization performance.
🎯 What it does: A framework for sign language translation without gloss based on Conditional Variational Autoencoders (CV-SLT) is proposed, which directly aligns sign language videos with cross-modal representations of text.
Confusing Pair Correction Based on Category Prototype for Domain Adaptation under Noisy Environments
Churan Zhi (Beijing Jiaotong University), Shuhui Wang (Institute of Computing Technology, Chinese Academy of Sciences)
CodeDomain AdaptationImage
🎯 What it does: Proposes an unsupervised domain adaptation method in noisy environments, utilizing category prototypes to identify and correct mislabeling of similar class pairs.
CONSIDER: Commonalities and Specialties Driven Multilingual Code Retrieval Framework
Rui Li (University of Science and Technology of China), Shijin Wang (Hefei Normal University)
CodeRetrievalTransformerContrastive LearningText
🎯 What it does: Proposes the CONSIDER framework, which enhances multilingual code retrieval performance by modeling linguistic commonality and specificity.
🎯 What it does: A post-hoc calibration method based on temperature scaling is proposed—Consistency Guided Temperature Scaling (CTS), which utilizes the style and content information of source domain samples to enhance the model's calibration performance in unknown domains.
🎯 What it does: This paper proposes ConsistentEE, an early exit method based on reinforcement learning, which requires only one internal classifier to correctly predict instances during the training phase, thereby achieving consistency between training and inference.
Constrained Bayesian Optimization under Partial Observations: Balanced Improvements and Provable Convergence
Shengbo Wang (University of Electronic Science and Technology of China), Ke Li (University of Exeter)
CodeOptimizationReinforcement LearningTabular
🎯 What it does: A Bayesian optimization framework for partially observable constraints (CBOB) is proposed, which efficiently solves POCOPs by balancing exploration and exploitation and utilizing HLGP (Heterogeneous Likelihood Gaussian Process).
Context-I2W: Mapping Images to Context-Dependent Words for Accurate Zero-Shot Composed Image Retrieval
Yuanmin Tang (Institute of Information Engineering, Chinese Academy of Sciences), Qi Wu (Australia Institute of Machine Learning, University of Adelaide)
CodeRetrievalVision Language ModelContrastive LearningImage
🎯 What it does: Proposes the Context-I2W network, which maps reference images to pseudo-words related to descriptions, achieving zero-shot composite image retrieval.
Contextual Pre-planning on Reward Machine Abstractions for Enhanced Transfer in Deep Reinforcement Learning
Guy Azran (Technion), Sarah Keren (Technion)
CodeReinforcement Learning
🎯 What it does: A Contextual PRE-Planning (C-PREP) framework is proposed, which utilizes a Reward Machine (RM) to generate task-specific abstractions for contextual MDPs and employs optimal abstraction transfer labels and potential reward shaping in deep reinforcement learning to achieve zero/few-shot transfer.
🎯 What it does: This paper proposes an unsupervised image animation method that utilizes a Continuous Piecewise Affine (CPAB) transformation model to map source images to the motion of driving videos, and enhances animation quality by extracting keypoint semantic information through SAM and structural information through DINO ViT.
Continuous Treatment Effect Estimation Using Gradient Interpolation and Kernel Smoothing
Lokesh Nagalapatti (Indian Institute of Technology Bombay), Sunita Sarawagi (Indian Institute of Technology Bombay)
CodeBiomedical Data
🎯 What it does: A method for generating unsupervised control samples based on gradient interpolation and Gaussian process kernel smoothing is proposed for the individualized estimation of continuous treatment effects.
ContraNovo: A Contrastive Learning Approach to Enhance De Novo Peptide Sequencing
Zhi Jin (Shanghai Artificial Intelligence Laboratory), Siqi Sun (Research Institute of Intelligent Complex Systems Fudan University)
CodeProtein Structure PredictionTransformerContrastive LearningBiomedical Data
🎯 What it does: A de novo peptide sequence prediction algorithm called ContraNovo based on contrastive learning is proposed, which utilizes quality information to enhance prediction accuracy during the decoding process.
🎯 What it does: A contrastive continual learning framework based on importance sampling (CCLIS) is proposed, which recovers the data distribution of previous tasks by replaying important samples from a buffer and maintains knowledge through prototype-instance relationship distillation.
🎯 What it does: A diversity measurement and constraint mechanism based on Contribution Dimension Structure (CDS) is proposed to improve the coreset selection process.
🎯 What it does: A controllable 3D face generation framework TEx-Face based on a three-stage approach is proposed, which completes 3D GAN inversion, conditional style code diffusion, and 3D face decoding.
🎯 What it does: Based on functional magnetic resonance imaging (fMRI) signals, a Controlled Visual Diffusion Model (CMVDM) is constructed to generate images that are highly similar to the original visual stimuli.
🎯 What it does: An improvement of the Forward-Forward (FF) algorithm based on convolutional channel competitive learning is proposed, using CFSE blocks and CwC loss to achieve hierarchical training without negative samples.
ConVQG: Contrastive Visual Question Generation with Multimodal Guidance
Li Mi (École Polytechnique Fédérale de Lausanne), Devis Tuia (École Polytechnique Fédérale de Lausanne)
CodeGenerationTransformerVision Language ModelContrastive LearningImageTextMultimodality
🎯 What it does: This paper proposes a visual question generation method based on dual-modal contrastive learning, named ConVQG, which can generate questions that are highly relevant to image content and rich in knowledge while satisfying textual constraints (answers, knowledge triples, or titles).
Cooper: Coordinating Specialized Agents towards a Complex Dialogue Goal
Yi Cheng (Hong Kong Polytechnic University), Yefeng Zheng (Tencent)
CodeTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The COOPER framework is proposed, which coordinates specialized multi-agents, each responsible for different aspects of complex dialogue goals, to jointly advance the dialogue process to achieve the goals.
CORECODE: A Common Sense Annotated Dialogue Dataset with Benchmark Tasks for Chinese Large Language Models
Dan Shi (Tianjin University), Deyi Xiong (Zhejiang Lab)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: A large-scale Chinese two-person dialogue commonsense knowledge annotation dataset, CORECODE, has been constructed, and six benchmark tasks have been designed to evaluate the commonsense reasoning and conflict detection capabilities of large language models.
🎯 What it does: This paper proposes CMGN (Coreference-guided Mind-Map Generation Network) to generate mind maps for documents: it first constructs a coreference graph and encodes sentence relationships through a graph neural network, then utilizes a graph contrastive learning module to enhance the representation of the graph structure, ultimately generating a sentence governance relationship graph.
Correlation Matching Transformation Transformers for UHD Image Restoration
Cong Wang (Hong Kong Polytechnic University), Jun Liu (National University of Singapore)
CodeRestorationTransformerImage
🎯 What it does: A universal Transformer named UHDformer is proposed for ultra-high-definition (UHD) image denoising, dehazing, and deblurring, achieving effective transformation of high-resolution features to low-resolution space through two modules, DualCMT and ACM, thereby enhancing recovery quality.
🎯 What it does: This paper proposes a few-shot human action counting framework based on recordable voice commands, utilizing the user reading 'one, two, three' to extract action samples and perform counting.
Counterfactual-Enhanced Information Bottleneck for Aspect-Based Sentiment Analysis
Mingshan Chang (Shenzhen Institutes of Advanced Technology), Ruifeng Xu (Harbin Institute of Technology)
CodeClassificationRecommendation SystemTransformerLarge Language ModelPrompt EngineeringText
🎯 What it does: A Contrastive Experiment Enhanced Information Bottleneck (CEIB) framework is proposed to reduce spurious correlations caused by surface features in Aspect-Based Sentiment Analysis (ABSA) and enhance model robustness.
Coupled Confusion Correction: Learning from Crowds with Sparse Annotations
Hansong Zhang (Institute of Information Engineering, Chinese Academy of Sciences), Shiming Ge (Shanghai University)
CodeData-Centric LearningMeta LearningImage
🎯 What it does: The study investigates how to learn in a sparsely annotated crowdsourcing environment, proposing the CCC model that corrects the confusion matrix through a dual model collaboration.
🎯 What it does: A curvature regularization method based on normalized Hessian trace, CR-SAM, is proposed to enhance the generalization ability of deep networks.
🎯 What it does: This paper proposes a point cloud completion network CRA-PCN based on Cross-Resolution Transformer, which can effectively aggregate features across different resolution levels and gradually generate complete point clouds.
🎯 What it does: This paper proposes an incremental learning method based on Cross-Class Feature Augmentation (CCFA), which generates samples of new classes by applying adversarial perturbations to features in the feature space of the old model, thereby compensating for the collapse of decision boundaries caused by insufficient samples from old tasks.
🎯 What it does: A CCGR cross-covariate gait recognition benchmark is proposed, providing a dataset with millions of samples, 53 types of covariates, and 33 viewpoints;
🎯 What it does: A cross-layer and cross-sample feature optimization network C2-Net is proposed for few-shot fine-grained image classification, addressing feature noise and matching mismatch issues.
🎯 What it does: A CUSA framework is proposed, which uses the soft labels of a single-modal pre-trained model to align cross-modal and single-modal soft labels for the image-text retrieval model, thereby addressing the issues of false negative samples and single-modal semantic loss.
CrossBind: Collaborative Cross-Modal Identification of Protein Nucleic-Acid-Binding Residues
Linglin Jing (Shanghai Artificial Intelligence Laboratory), Siqi Sun (Research Institute of Intelligent Complex Systems, Fudan University)
CodeRecognitionProtein Structure PredictionConvolutional Neural NetworkLarge Language ModelContrastive LearningMultimodalityPoint CloudBiomedical Data
🎯 What it does: A cross-modal framework called CrossBind is proposed, which combines protein structure (atomic point clouds) and sequence (ESM-2 language model) to identify nucleic acid binding residues.
CrystalBox: Future-Based Explanations for Input-Driven Deep RL Systems
Sagar Patel (University of California), Nina Narodytska (VMware Research)
CodeExplainability and InterpretabilitySupervised Fine-TuningReinforcement LearningSequential
🎯 What it does: The CrystalBox framework is proposed, providing post-hoc interpretability for input-driven deep reinforcement learning controllers and generating future causal explanations.
CTO-SLAM: Contour Tracking for Object-Level Robust 4D SLAM
Xiaohan Li (University of Science and Technology of China), Jun Wu (Fudan University)
CodeObject TrackingPose EstimationDepth EstimationAutonomous DrivingSimultaneous Localization and MappingImageVideo
🎯 What it does: This paper proposes a contour tracking-based object-level 4D SLAM system called CTO-SLAM, which can simultaneously estimate camera pose and dynamic object motion in dynamic scenes, and generate a sparse 4D map.
Cumulative Difference Learning VAE for Time-Series with Temporally Correlated Inflow-Outflow
Tianchun Li (Purdue University), Xiaoqian Wang (Purdue University)
CodeGenerationData SynthesisRecurrent Neural NetworkAuto EncoderGenerative Adversarial NetworkTime SeriesSequentialFinance Related
🎯 What it does: A differential learning-based variational autoencoder (DT-VAE) is proposed, which addresses the issue of error accumulation in time series generation by introducing an inflow-outflow structure, and can be combined with GANs to enhance generation quality.
🎯 What it does: This paper proposes a Residual Enhanced Soft Graph Normalization Layer (R-SoftGraphAIN) and a label smoothing-based curriculum learning framework (SmoothCurriculum) to address the issues of over-smoothing and optimization difficulties in deep graph neural networks.
CutFreq: Cut-and-Swap Frequency Components for Low-Level Vision Augmentation
Hongyang Chen (Xi'an Jiaotong University), Kaisheng Ma (Tsinghua University)
CodeRestorationSegmentationImage
🎯 What it does: A CutFreq data augmentation method is proposed for cutting and exchanging frequency components in the frequency domain to enhance the performance of low-level visual tasks.
🎯 What it does: A multi-source domain adaptation method with a cyclic self-improvement approach is proposed, utilizing source-specific networks and domain integration networks to generate high-confidence pseudo-labels through instance-level voting, and mutually guiding each other in a cycle to enhance target domain performance.
🎯 What it does: This paper proposes a user feature attention-based DAG-Aware Variational Autoencoder (DAVA), which utilizes nearly a million real social network propagation data to generate large-scale, highly realistic propagation DAG graphs.
🎯 What it does: This paper proposes an efficient 2D object detection framework called DALDet, which utilizes depth maps to output target distance information based on 2D detection.
🎯 What it does: This work proposes DanceAnyWay, a two-stage hierarchical generation framework that synchronously generates 3D dance movements using the beat information from audio.
🎯 What it does: A novel end-to-end generalized zero-shot learning framework, D3GZSL, is designed and implemented, which enhances the performance of generative models under real distributions through the combination of knowledge distillation and anomaly distribution detection.
CodeClassificationComputational EfficiencyKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringImageText
🎯 What it does: This paper proposes the Data Shunt paradigm, which uses the confidence of a small model to determine whether to invoke a large model, and achieves collaboration between small and large models through Prompt Pruning and 2-Stage Confidence Distillation, significantly improving performance while reducing the cost of invoking the large model.
Data-Driven Knowledge-Aware Inference of Private Information in Continuous Double Auctions
Lvye Cui (Beijing Institute of Technology), Haoran Yu (Beijing Institute of Technology)
CodeRecurrent Neural NetworkTabular
🎯 What it does: This paper proposes a knowledge-aware machine learning framework that infers the private cost vector of sellers in continuous double auctions by utilizing sellers' inquiry behavior and the global inquiry-cost ratio distribution.