AAAI Conference on Artificial Intelligence · 1442 papers
Can We Get Rid of Handcrafted Feature Extractors? SparseViT: Nonsemantics-Centered, Parameter-Efficient Image Manipulation Localization Through Spare-Coding Transformer
Lei Su (Sichuan University), Ji-Zhe Zhou (Sichuan University)
🎯 What it does: This paper proposes SparseViT, an image tampering localization network that achieves feature extraction without manual intervention through a sparse self-attention mechanism.
Yi-Kai Zhang (Nanjing University), Han-Jia Ye (Nanjing University)
CodeLarge Language ModelSupervised Fine-TuningTextMultimodality
🎯 What it does: This paper studies a method for instruction-level model routing achieved through capability instruction tuning and proposes the MODEL-SAT framework.
🎯 What it does: A class-incremental learning method based on Circular Prompt Aggregation (CAPrompt) is proposed to eliminate the prompt inconsistency problem caused by task ID prediction.
Capture Global Feature Statistics for One-Shot Federated Learning
Zenghao Guan (Institute of Information Engineering Chinese Academy of Sciences), Xiaoyan Gu (Institute of Information Engineering Chinese Academy of Sciences)
🎯 What it does: Utilizing pre-trained models to extract features, clients upload statistical data for each type of feature, and the server aggregates to obtain global mean and covariance, directly constructing a Gaussian Naive Bayes classifier, thus achieving global and personalized federated learning with a single communication.
CareBot: A Pioneering Full-Process Open-Source Medical Language Model
Lulu Zhao (Beijing Academy of Artificial Intelligence), Hua Zhou (Beijing University of Posts and Telecommunications)
CodeReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextBiomedical Data
🎯 What it does: A bilingual medical large model, CareBot, has been developed, employing a two-stage continuous pre-training, supervised fine-tuning, and reinforcement learning based on human feedback, aimed at enhancing medical diagnosis, treatment planning, and medical education capabilities.
🎯 What it does: A model named CasFT has been developed, which utilizes observed propagation cascades, neural ODE to predict growth rates, and diffusion models to generate future trends, combined with spatiotemporal features to achieve content popularity prediction.
Qingyang Zhou (University of Waterloo), Meng Xu (University of Waterloo)
CodeFlow-based ModelTabularFinance Related
🎯 What it does: A multi-layer causal consistent normalization flow model (CCNF) is proposed, achieving universal approximation under complex distributions.
🎯 What it does: This paper proposes a robustness certification method for speaker recognition models based on Randomized Smoothing, which can provide a provable radius regarding the invariance of model predictions under additive perturbations constrained by the L₂ norm for a given audio sample.
Certifying Bounds Propagation for Integer Multiplication Constraints
Matthew J. McIlree (University of Glasgow), Ciaran McCreesh (University of Glasgow)
CodeTabular
🎯 What it does: Log the proof of bounds-consistency propagation for integer multiplication constraints, providing pseudo-Boolean (PB) proof steps that can be checked in the VeriPB verifier.
Chain-of-Instructions: Compositional Instruction Tuning on Large Language Models
Shirley Anugrah Hayati (University of Minnesota), Dongyeop Kang (Amazon)
CodeTransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: This paper introduces the concept of Chain-of-Instructions (CoI) and enhances the model's understanding and execution capabilities for multi-subtask combination instructions by fine-tuning using CoI instructions on large language models.
ChatterBox: Multimodal Referring and Grounding with Chain-of-Questions
Yunjie Tian (University of Chinese Academy of Sciences), Qixiang Ye (University of Chinese Academy of Sciences)
CodeRecognitionObject DetectionTransformerSupervised Fine-TuningVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: A multi-modal continuous reference and localization task (MCQ) and the corresponding benchmark dataset CB-300K have been constructed, with the ChatterBox model proposed as a baseline, and evaluation metrics for chain-based question answering have been provided.
ChatTime: A Unified Multimodal Time Series Foundation Model Bridging Numerical and Textual Data
Chengsen Wang (Beijing University of Posts and Telecommunications), Jianxin Liao
CodeTransformerLarge Language ModelSupervised Fine-TuningMultimodalityTime Series
🎯 What it does: ChatTime is proposed, a multimodal temporal foundation model that treats time series data as a foreign language and utilizes LLMs, supporting zero-shot prediction, context-assisted prediction, and temporal question answering.
CodeTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelImageTextMultimodality
🎯 What it does: ChemVLM, a multimodal large language model for the field of chemistry, has been developed to simultaneously process chemical images and text information.
Cherry-Picking in Time Series Forecasting: How to Select Datasets to Make Your Model Shine
Luis Roque (Universidade do Porto), Luís Torgo
CodeTime SeriesBenchmark
🎯 What it does: This paper proposes a framework to assess the performance bias caused by dataset selection (cherry-picking) in time series forecasting experiments, demonstrating that a small number of non-representative datasets can significantly exaggerate model performance.
Cirbo: A New Tool for Boolean Circuit Analysis and Synthesis
Daniil Averkov (St. Petersburg State University), Aleksey Vorobiev (Neapolis University Pafos)
CodeOptimizationBenchmark
🎯 What it does: This paper presents the open-source tool Cirbo, designed for the analysis (such as satisfiability checking) and synthesis (such as minimum circuit solving) of Boolean circuits, achieving efficient circuit optimization and size reduction through various algorithms.
Yifan Ding (University of Notre Dame), Tim Weninger (University of Notre Dame)
CodeTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: This study evaluates the impact of citations (0, 1, 5 citations) on the credibility of responses from large language models through a designed RCT experiment, and observes the relationship between users' citation retrieval behavior and their self-reported trust levels.
CITI: Enhancing Tool Utilizing Ability in Large Language Models Without Sacrificing General Performance
Yupu Hao (Institute of Automation Chinese Academy of Sciences), Jun Zhao (Institute of Automation Chinese Academy of Sciences)
CodeTransformerLarge Language ModelSupervised Fine-TuningMixture of ExpertsText
🎯 What it does: A framework named CITI has been developed, combining Mixture-of-LoRA and selective full parameter fine-tuning, significantly enhancing the tool invocation capability while maintaining the general performance of large language models.
CLIMB-ReID: A Hybrid CLIP-Mamba Framework for Person Re-Identification
Chenyang Yu (Dalian University of Technology), Huchuan Lu (Dalian University of Technology)
CodeRecognitionRetrievalTransformerVision Language ModelContrastive LearningImageVideo
🎯 What it does: A new method called CLIMB-ReID is proposed, which integrates the CLIP visual-language model with the Mamba lightweight sequence modeling framework for person re-identification.
CLIP-MSM: A Multi-Semantic Mapping Brain Representation for Human High-Level Visual Cortex
Guoyuan Yang (Beijing Institute of Technology), Xuesong Li (Beijing Institute of Technology)
CodeRepresentation LearningTransformerVision Language ModelContrastive LearningImageBiomedical DataMagnetic Resonance Imaging
🎯 What it does: By combining the CLIP model with CLIP Dissection, a multi-semantic mapping framework called CLIP-MSM was constructed, capable of voxel-level predictions and multi-semantic interpretations of the human high-level visual cortex without prior assumptions.
CodeClassificationRetrievalTransformerPrompt EngineeringVision Language ModelContrastive LearningPoint Cloud
🎯 What it does: This paper proposes a no-reference point cloud quality assessment method based on CLIP, called CLIP-PCQA, which aligns visual features with multi-level quality descriptions using a language retrieval mapping and predicts opinion distribution (OSD) to provide a final quality score.
🎯 What it does: This paper proposes an exposure correction method called CLIP-RestoreX based on the CLIP structure and perceptual priors. It utilizes the shallow structural features and deep perceptual features of CLIP to construct priors, enhances the damaged features through a frequency domain diffusion model (FFEDM), and finally injects the enhanced priors into the Restormer network to achieve exposure correction.
CNC: Cross-modal Normality Constraint for Unsupervised Multi-class Anomaly Detection
Xiaolei Wang (Xi'an Jiaotong-Liverpool University), Jimin Xiao (Beijing Jiaotong University)
CodeAnomaly DetectionTransformerMixture of ExpertsVision Language ModelImageMultimodality
🎯 What it does: A reverse distillation framework based on cross-modal regularization is proposed, utilizing learnable prompts in the visual-language model to guide the decoder to only reconstruct normal features, thereby suppressing overgeneralization (OG) caused by multi-class training and enhancing anomaly detection and localization performance.
Cobra: Extending Mamba to Multi-Modal Large Language Model for Efficient Inference
Han Zhao (Westlake University), Donglin Wang (Westlake University)
CodeComputational EfficiencyRepresentation LearningTransformerLarge Language ModelVision Language ModelImageMultimodality
🎯 What it does: A multimodal large language model named Cobra is proposed, which integrates the Mamba state space model with a visual encoder to achieve efficient inference.
🎯 What it does: Using a pre-trained diffusion model for unsupervised repair and editing of out-of-distribution (OoD) images, generating images that retain input details while achieving high realism.
CodeHalu: Investigating Code Hallucinations in LLMs via Execution-based Verification
Yuchen Tian (Hong Kong Baptist University), Dawn Song (University of California, Berkeley)
CodeAI Code AssistantTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper defines the concept of 'code hallucination' in code generation, proposes a dynamic detection algorithm called CodeHalu based on execution verification, and constructs a large-scale evaluation benchmark named CodeHaluEval, followed by a systematic evaluation of 17 mainstream large language models.
🎯 What it does: The CognitionCapturer framework is proposed, which utilizes a multimodal expert encoder for cross-modal alignment of EEG and maps the EEG embedding into the CLIP space through a diffusion prior, thereby reconstructing high-fidelity visual stimuli without the need to fine-tune the generative model.
Coherency Improved Explainable Recommendation via Large Language Model
Shijie Liu (East China Normal University), Wei Zhang (WeChat AI Tencent)
CodeRecommendation SystemExplainability and InterpretabilityTransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Utilize a large language model to first predict ratings, then map the ratings to word vectors, and finally generate interpretable recommendations consistent with the ratings using this vector along with user-item information.
CohEx: A Generalized Framework for Cohort Explanation
Fanyu Meng (University of California), Xin Chen (Georgia Institute of Technology)
CodeExplainability and InterpretabilityImageTabular
🎯 What it does: This paper presents CohEx, a general framework that transforms existing local feature importance explanation methods into group explanations, enhancing the interpretability of group explanations through supervised clustering and iterative recalculation of importance.
🎯 What it does: The Collaborative Semantic Consistency Alignment (CSCA) method is proposed, which combines slice Wasserstein distance distribution alignment, graph neural network cross-domain semantic consistency alignment, and dual consistency regularization to address the mixed target domain adaptation (BTDA) problem.
🎯 What it does: The paper proposes a color transfer method based on reversible regularized flows (ModFlows), which maps the color distribution of the target image to that of the reference image in the RGB color space.
COLUMBUS: Evaluating COgnitive Lateral Understanding Through Multiple-Choice reBUSes
Koen Kraaijveld (Vrije Universiteit Amsterdam), Filip Ilievski (Vrije Universiteit Amsterdam)
CodeTransformerLarge Language ModelVision Language ModelTextMultimodalityBenchmark
🎯 What it does: A COLUMBUS visual lateral thinking benchmark is proposed, which evaluates the performance of VLM in creative reasoning by automatically generating multiple-choice puzzles through automated rule generation.
Yihao Sun (Syracuse University), Kristopher Micinski (Washington State University)
CodeGraphTabularBenchmark
🎯 What it does: FVLOG has been implemented, a columnar Datalog engine based on NVIDIA H100 GPU, supporting columnar storage, dual indexing (sorting + hashing), and parallel relational algebra operations, completing the full reasoning process from rules to fixed points.
🎯 What it does: A method is proposed to model using a binomial process and employ maximum likelihood estimation to fuse prior distributions with empirical data, resulting in a continuous confidence calibration curve.
🎯 What it does: This paper proposes a graph de-learning framework GSMU based on community structure mapping and implements a community-centered graph eraser CGE, which can achieve node-level de-learning quickly and efficiently while maintaining model performance.
🎯 What it does: A collaborative music filling model CMI is designed, implementing a Human-in-the-Loop (HITL) framework for symbolic music filling, capable of local filling and iterative improvement based on user-marked missing segments.
ComprehendEdit: A Comprehensive Dataset and Evaluation Framework for Multimodal Knowledge Editing
Yaohui Ma (Harbin Institute of Technology), Zhiheng Ma (Shenzhen Institutes of Advanced Technology)
CodeKnowledge DistillationTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: Proposed the ComprehendEdit benchmark and two new evaluation metrics, KGI/KPI, and designed a Hierarchical Context Editing (HICE) method to improve knowledge editing in multimodal large language models.
🎯 What it does: Prova is proposed, a multimodal prototype classifier for large vocabulary object detection, which can significantly improve detection accuracy.
🎯 What it does: This paper proposes an incremental learning method based on pre-trained models (PTMCIL), which achieves feature adaptation and inter-class discrimination for different tasks by freezing the Equiangular Tight Frame (ETF) classifier and combining task-related adaptation, feature compression, and structural alignment.
Compressing Streamable Free-Viewpoint Videos to 0.1 MB per Frame
Luyang Tang (Peking University), Ronggang Wang (Peking University)
CodeCompressionGaussian SplattingVideo
🎯 What it does: A streaming FVV compression framework iFVC based on 3D Gaussians is proposed, capable of online training, real-time rendering, and low storage costs;
CoMT: A Novel Benchmark for Chain of Multi-modal Thought on Large Vision-Language Models
Zihui Cheng (Central South University), Libo Qin (Harbin Institute of Technology)
CodeTransformerLarge Language ModelVision Language ModelImageTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: The CoMT benchmark is proposed to evaluate the performance of large visual-language models in chain-based multimodal reasoning, requiring outputs that include both textual and visual (image) reasoning steps.
Concept Matching with Agent for Out-of-Distribution Detection
Yuxiao Lee (Jilin University), Yi Chang (Jilin University)
CodeAnomaly DetectionTransformerAgentic AIVision Language ModelContrastive LearningImage
🎯 What it does: Proposes the Concept Matching and Agent (CMA) method in an untrained zero-shot environment, using neutral text agents to construct a triangular relationship that improves OOD detection performance.
ConceptSearch: Towards Efficient Program Search Using LLMs for Abstraction and Reasoning Corpus (ARC)
Kartik Singhal (Indraprastha Institute of Information Technology Delhi), Gautam Shroff (Indraprastha Institute of Information Technology Delhi)
CodeOptimizationComputational EfficiencyAI Code AssistantConvolutional Neural NetworkTransformerLarge Language ModelContrastive LearningText
🎯 What it does: A concept-based function search algorithm called ConceptSearch is proposed, which utilizes LLM to generate programs and guides the search through concept scoring to solve the ARC task.
Concurrent Planning and Execution in Lifelong Multi-Agent Path Finding with Delay Probabilities
Yue Zhang (Monash University), Peter J. Stuckey (Monash University)
CodeOptimizationRobotic Intelligence
🎯 What it does: This paper proposes a new parallel planning and execution framework, PIE-D, to maintain synchronization between planning and execution and maximize throughput in the lifelong multi-agent pathfinding (LMAPF) problem with execution delays.
Conditional Diffusion Models Based Conditional Independence Testing
Yanfeng Yang (East China Normal University), Renming Zhang (Boston University)
CodeDiffusion modelScore-based ModelTabular
🎯 What it does: A conditional independence testing method based on conditional diffusion models is proposed, utilizing a conditional randomization testing framework to generate approximate samples of $X|Z$ and estimating conditional mutual information with a classifier.
Conditional Feature Importance with Generative Modeling Using Adversarial Random Forests
Kristin Blesch (Leibniz Institute for Prevention Research and Epidemiology), Marvin N. Wright
CodeExplainability and InterpretabilityAdversarial AttackGenerative Adversarial NetworkTabularTime Series
🎯 What it does: The cARFi method is proposed, which utilizes Adversarial Random Forests (ARF) to evaluate conditional feature importance for any subset of features without model fitting and with low hyperparameter tuning.
🎯 What it does: A conditional latent coding (CLC) framework based on an external image dictionary is proposed, which dynamically generates latent references to achieve deep image compression.
ConDo: Continual Domain Expansion for Absolute Pose Regression
Zijun Li (Xiamen University), Cheng Wang (Xiamen University)
CodePose EstimationDomain AdaptationKnowledge DistillationTransformerSimultaneous Localization and MappingImage
🎯 What it does: This paper proposes the Continual Domain Expansion (ConDo) method, which continuously updates the absolute pose regression model using unlabeled inference images after deployment, and obtains supervisory signals from scene-independent localization methods through knowledge distillation, constructing a large-scale benchmark aimed at long-term environmental changes.
ConDSeg: A General Medical Image Segmentation Framework via Contrast-Driven Feature Enhancement
Mengqi Lei (China University of Geosciences), Xin Wang (Baidu Inc)
CodeSegmentationConvolutional Neural NetworkTransformerImageBiomedical Data
🎯 What it does: A general medical image segmentation framework named ConDSeg is proposed, addressing soft boundaries and co-occurrence phenomena, and improving segmentation quality in low-light and low-contrast environments.
Confidence Estimation for Error Detection in Text-to-SQL Systems
Oleg Somov (Artificial Intelligence Research Institute), Elena Tutubalina (Sber Artificial Intelligence)
CodeClassificationAnomaly DetectionTransformerLarge Language ModelTextBiomedical DataElectronic Health Records
🎯 What it does: Proposed the incorporation of a rejectable selective classifier in the Text-to-SQL system to leverage model uncertainty for detecting erroneous generations and unanswerable questions.
Xiuwen Gong (University of Technology Sydney), Guandong Xu (Education University of Hong Kong)
CodeClassificationSupervised Fine-TuningImage
🎯 What it does: A CP-PLL method is proposed to quantify model uncertainty and provide reliable confidence guarantees within the framework of Partial Label Learning (PLL) using Conformal Prediction.
Conformal Thresholded Intervals for Efficient Regression
Rui Luo (City University of Hong Kong), Zhixin Zhou (Alpha Benito Research)
CodeTabular
🎯 What it does: This paper proposes a new conformal prediction method - Conformal Thresholded Intervals (CTI), which estimates the length of conditional interval ranges through multi-output quantile regression and constructs a minimal prediction set based on a threshold to ensure coverage.
Conformalized Interval Arithmetic with Symmetric Calibration
Rui Luo (City University of Hong Kong), Zhixin Zhou (Alpha Benito Research)
CodeTabularTime Series
🎯 What it does: This paper studies a new method for predicting confidence intervals for the sum/average of unknown label sets—Symmetric Calibration of Synthetic Interval Arithmetic (CIA).
Ruijiang Gao (University of Texas at Dallas), Mingzhang Yin (University of Florida)
CodeRecommendation SystemOptimizationReinforcement Learning from Human FeedbackReinforcement LearningTabularBiomedical DataFinance Related
🎯 What it does: This paper studies strategy learning in human-machine collaboration in the presence of unobserved confounding and proposes the 'Confounding-Robust Deferral Policy Learning' method.
ConSense: Continually Sensing Human Activity with WiFi via Growing and Picking
Rong Li (Soochow University), Juncheng Jia (Soochow University)
CodeRecognitionTransformerTime Series
🎯 What it does: A lightweight, example-free incremental learning framework called ConSense is proposed, specifically for continuous human activity recognition using WiFi signals.
Consistency of Compositional Generalization Across Multiple Levels
Chuanhao Li (Beijing Institute of Technology), Yunde Jia (Shenzhen MSU-BIT University)
CodeOptimizationMeta LearningLarge Language ModelVideoText
🎯 What it does: A multi-layer optimization framework based on meta-learning is proposed, which gradually trains the model in a sequence from simple to complex by introducing multiple meta-weight networks to achieve consistent combinatorial generalization across multiple levels (phrase-phrase, phrase-word, word-word).
🎯 What it does: The COOREM method is proposed to address the dual challenges of OOD and constraint risks in constrained offline black-box optimization, achieving dynamic assessment and management of risks.
Content-free Logical Modification of Large Language Model by Disentangling and Modifying Logic Representation
Xin Wu (South China University of Technology), Yi Cai (South China University of Technology)
CodeRecognitionGenerationTransformerLarge Language ModelContrastive LearningText
🎯 What it does: A logical control framework (LCF) is proposed, which splits the hidden layers of LLM into content and logic spaces, and transfers representations to valid areas in the logic space through contrastive learning, thereby enhancing the logical correctness of generated conclusions.
🎯 What it does: To address the issue of homogeneity shift in graph transfer learning, we propose the Contextual Structural Graph Neural Network (CS-GNN), which achieves cross-domain structural knowledge transfer and node classification by constructing local EGO networks, quantifying structural diversity based on feature smoothing moment, employing a contextual attention mechanism, and implementing group fairness loss.
🎯 What it does: Utilizing fixed base model features, combined with kernel functions and random Fourier features, class incremental learning is achieved through linear discriminant analysis as new tasks are continuously added.
Continual Unsupervised Generative Modelling via Online Optimal Transport
Fei Ye (University of Electronic Science and Technology of China), Kun Zhang (Carnegie Mellon University)
CodeGenerationData SynthesisDiffusion modelImage
🎯 What it does: This paper proposes an online unsupervised generative modeling framework that utilizes Sinkhorn distance-driven dual dynamic memory (SDDM) to select and maintain short-term and long-term samples, thereby achieving a task-free generative model in continual learning.
🎯 What it does: The paper proposes and validates the concept of 'harmless perturbation', which states that there exists a continuous perturbation subspace in deep neural networks, where any perturbation from this space, regardless of its magnitude, will not change the network output.
🎯 What it does: This paper proposes an auxiliary contrastive learning model CALHG that combines edge perturbation, graph diffusion, and category-guided multi-view contrastive learning to learn node representations on heterogeneous graphs with no or few features.
Contrastive Functional Principal Component Analysis
Eric Zhang (University of North Carolina at Chapel Hill), Didong Li (University of North Carolina at Chapel Hill)
CodeContrastive LearningTime SeriesFinance Related
🎯 What it does: Proposes Contrastive Functional Principal Component Analysis (CFPCA) to identify low-dimensional structures unique to the foreground group relative to the background group, reducing common variation;
Controllable Protein Sequence Generation with LLM Preference Optimization
Xiangyu Liu (Nanjing University), Wei Hu (Nanjing University)
CodeGenerationOptimizationDrug DiscoveryTransformerLarge Language ModelSupervised Fine-TuningTextBiomedical Data
🎯 What it does: The CtrlProt method is proposed, achieving controllable protein sequence generation through prefix fine-tuning of a pre-trained protein LLM and multi-list priority optimization.
Controlling Equational Reasoning in Large Language Models with Prompt Interventions
Jordan Meadows (University of Manchester), André Freitas (Idiap Research Institute)
CodeTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A 30k fine-grained equation derivation dataset was constructed using a symbolic data generation framework, and large language models were fine-tuned and evaluated through various prompt interventions (variable renaming, expression swapping, target replacement, step deletion) to explore the impact of interventions on the distribution of reasoning errors.
🎯 What it does: A lightweight CoPEFT framework is proposed to achieve rapid adaptation of trained multi-agent collaborative perception models in newly deployed environments.
CoPRA: Bridging Cross-domain Pretrained Sequence Models with Complex Structures for Protein-RNA Binding Affinity Prediction
Rong Han (Tsinghua University), Ting Chen (Tsinghua University)
CodeDrug DiscoveryProtein Structure PredictionTransformerContrastive LearningBiomedical Data
🎯 What it does: This study proposes the CoPRA model, which integrates protein language models, RNA language models, and the three-dimensional structural information of protein-RNA complexes to predict protein-RNA binding affinity.
CoRA: Collaborative Information Perception by Large Language Model’s Weights for Recommendation
Yuting Liu (Northeastern University), Xingwei Wang (Northeastern University)
CodeRecommendation SystemTransformerLarge Language ModelTabular
🎯 What it does: Proposes the CoRA framework, which integrates collaborative filtering information in the LLM parameter space to achieve personalized recommendations without fine-tuning and without additional tokens.
🎯 What it does: A new text embedding learning method called CoRe is proposed, which enhances personalized text alignment and identity preservation from text to image through contextual regularization.
Correlation-Attention Masked Temporal Transformer for User Identity Linkage Using Heterogeneous Mobility Data
Ziang Yan (Ocean University of China), Junyu Dong (Ocean University of China)
CodeTransformerTime SeriesSequential
🎯 What it does: This paper proposes the MT-Link framework for cross-platform user identity linking, utilizing spatiotemporal co-occurrence information.
COSEE: Consistency-Oriented Signal-Based Early Exiting via Calibrated Sample Weighting Mechanism
Jianing He (Tongji University), Duoqian Miao (Tongji University)
CodeClassificationComputational EfficiencyTransformerLarge Language ModelSupervised Fine-TuningTextBenchmark
🎯 What it does: This paper proposes the COSEE framework, which simulates different acceleration thresholds during the training phase and generates sample weights for each layer, allowing the internal classifier to focus on samples that may exit at that layer during training, thus achieving consistency between training and inference and supporting adjustable rates.
CoT4Rec: Revealing User Preferences Through Chain of Thought for Recommender Systems
Weiqi Yue (Hangzhou Dianzi University), Jian Wan (Hangzhou Dianzi University)
CodeRecommendation SystemLarge Language ModelTabularChain-of-Thought
🎯 What it does: Designed and implemented a user preference analysis and dual-stage recommendation framework CoT4Rec based on chain-of-thought (CoT) reasoning, utilizing LLM to generate user preferences and construct interpretable recommendation paths.
Seungheun Baek (Korea University), Jaewoo Kang (AIGEN Sciences)
CodeGenerationData SynthesisAuto EncoderBiomedical Data
🎯 What it does: This paper proposes CRADLE-VAE, a causal modeling framework based on variational autoencoders, which utilizes adversarial and counterfactual reasoning to achieve decoupling of technical artifacts and realignment of baseline states, thereby enhancing the generation quality and reliability of single-cell gene perturbation predictions.
CodeRobotic IntelligenceMeta LearningReinforcement Learning from Human FeedbackSequentialBenchmark
🎯 What it does: The CraftFactory benchmark is proposed for evaluating the compositional generalization of conditional control strategies in the interactive control environment of Minecraft.
CrAM: Credibility-Aware Attention Modification in LLMs for Combating Misinformation in RAG
Boyi Deng (University of Science and Technology of China), Fuli Feng (University of Science and Technology of China)
CodeGenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: The CrAM (Credibility-aware Attention Modification) method is proposed, which identifies important attention heads and dynamically adjusts their attention weights based on document credibility within the Retrieval-Augmented Generation (RAG) framework, thereby reducing the negative impact of low-credibility documents on the generation results of large language models (LLMs).
CriSPO: Multi-Aspect Critique-Suggestion-guided Automatic Prompt Optimization for Text Generation
Han He (Amazon Web Services Artificial Intelligence Labs), Katrin Kirchhoff (Amazon Web Services Artificial Intelligence Labs)
CodeGenerationOptimizationTransformerLarge Language ModelPrompt EngineeringTextChain-of-Thought
🎯 What it does: A multi-dimensional criticism-suggestion driven automatic prompt optimization framework named CriSPO is proposed, further introducing AST suffix tuning to achieve joint optimization of multiple metrics;
🎯 What it does: Designed and implemented a coarse-fine hierarchical adversarial patch attack framework for RGB-T cross-modal trackers, achieving covert attacks across modalities.
Cross-modulated Attention Transformer for RGBT Tracking
Yun Xiao (Anhui University), Cong Liu (iFLYTEK Company)
CodeObject TrackingTransformerMultimodality
🎯 What it does: This paper proposes a Cross-Modal Modulation Attention Transformer (CAFormer), which simultaneously performs self-attention and cross-attention within a single attention module, achieving unified feature extraction and fusion of RGB and TIR modalities.
🎯 What it does: A cross-model feature space alignment method called FedFSA is proposed to address the issue of feature space inconsistency caused by client data imbalance in federated learning.
🎯 What it does: This paper proposes a method for estimator selection and hyperparameter tuning based on cross-validation for offline policy evaluation (OPE), which allows for estimator comparison and tuning using a dataset collected from a single logging policy.
Sijia Chen (Huazhong University of Science and Technology), Wenbing Tao (Huazhong University of Science and Technology)
CodeObject DetectionObject TrackingTransformerVision Language ModelVideoTextBenchmark
🎯 What it does: This paper proposes the Cross-Perspective Reference Multi-Object Tracking (CRMOT) task and presents an end-to-end method called CRTracker to achieve object detection, tracking, and language matching across perspectives.
🎯 What it does: Proposes the Crossfire framework for detecting and recovering from Bit Flip Attacks (BFA) on GNNs, capable of restoring the model to its pre-attack state without retraining, using unlabeled data, and without posterior testing.
CROSSNEWS: A Cross-Genre Authorship Verification and Attribution Benchmark
Marcus Ma (Georgia Institute of Technology), Wei Xu (Georgia Institute of Technology)
CodeRecognitionLarge Language ModelPrompt EngineeringTextBenchmark
🎯 What it does: A CROSSNEWS cross-genre author identification dataset was constructed, and various author identification models were evaluated, proposing a zero-shot LLM embedding method called SELMA.
CSL-L2M: Controllable Song-Level Lyric-to-Melody Generation Based on Conditional Transformer with Fine-Grained Lyric and Musical Controls
Li Chai (Westlake University), Donglin Wang (Westlake University)
CodeGenerationTransformerTextAudio
🎯 What it does: A controllable full lyric-to-melody generation method called CSL-L2M is proposed, which can generate complete and structurally reasonable melodies based on lyrics and user-specified musical attributes.
🎯 What it does: Designed and implemented a chunk-based streaming singing synthesis system CSSinger based on conditional VAE, and launched a fully streaming version CSSinger-FS;
🎯 What it does: Proposes the CtrlAvatar framework, which utilizes separable invertible networks to achieve decoupled generation of realistic and customizable 3D human avatars with distinct poses and textures;
🎯 What it does: A recommendation framework CUGF based on conformal prediction is proposed to generate prediction sets with adjustable probabilities of covering real items while ensuring fairness among different user groups.
🎯 What it does: The CustomCrafter framework is proposed, which can generate customized videos using subject learning while maintaining the motion generation and concept combination capabilities of the original video diffusion model, without the need for additional videos or re-fine-tuning.
CVLUE: A New Benchmark Dataset for Chinese Vision-Language Understanding Evaluation
Yuxuan Wang (Zhejiang Lab), Hongyang Chen (Zhejiang Lab)
CodeRetrievalVision Language ModelImageTextMultimodalityBenchmark
🎯 What it does: This paper proposes the CVLUE benchmark for comprehensively evaluating the understanding ability of visual-language models (VLM) in the context of Chinese culture.
CyberPal.AI: Empowering LLMs with Expert-Driven Cybersecurity Instructions
Matan Levi (IBM Research), Anton Puzanov (IBM Research)
CodeSafty and PrivacyAdversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: This paper constructs a security instruction dataset, SecKnowledge, with approximately 400,000 entries through a two-stage process assisted by expert-defined schemas and LLMs. The dataset is used to fine-tune LLMs, resulting in the CyberPal.AI series of security expert models, and the SecKnowledge-Eval evaluation set is designed for comprehensive assessment of the models.
DABL: Detecting Semantic Anomalies in Business Processes Using Large Language Models
Wei Guan (Shanghai Jiao Tong University), Shiyou Qian (University of Shanghai for Science and Technology)
CodeAnomaly DetectionTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextSequential
🎯 What it does: DABL is a method that uses large language models to detect semantic anomalies in business processes and explains the reasons for these anomalies in natural language.
Weihai Lu (Peking University), Zhiqiu Ye (Anhui University)
CodeClassificationDomain AdaptationTransformerLarge Language ModelImageTextMultimodality
🎯 What it does: This paper proposes the Domain-Aware Multi-Modal Multi-View Fake News Detection (DAMMFND) framework to address issues such as inaccurate domain recognition, negative transfer, and uneven contributions of domain heterogeneous modalities in multi-modal multi-domain fake news detection.