These 1442 AAAI 2025 papers come with a code repository. Each shows an AI one-line summary below — get the verified repo link + the full 6-part summary (innovation, method, data, results, limitations) and search every AAAI 2025 paper, free trial on arXivSub.
3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection
Enquan Yang (Shanghai University), Dan Zeng (Shanghai University)
CodeAnomaly DetectionKnowledge DistillationImage
🎯 What it does: The first large-scale industrial defect detection dataset for 3C products, 3CAD, is proposed, and an unsupervised anomaly detection framework, CFRG, is developed based on this dataset.
🎯 What it does: A two-stage 3D annotation-free learning framework AFOV is constructed, which generates pseudo-labels for point cloud learning through a 2D open vocabulary segmentation model and performs knowledge distillation.
3D-RPE: Enhancing Long-Context Modeling Through 3D Rotary Position Encoding
Xindian Ma (Tianjin University), Nan Xu (Beijing Wenge Technology)
CodeTransformerLarge Language ModelText
🎯 What it does: Proposes a 3D Rotational Position Encoding (3D-RPE) on the three-dimensional sphere to enhance the long-context modeling capabilities of large language models.
3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering
Qingyuan Zhou (Fudan University), Ying He (Nanyang Technological University)
CodeRestorationGraph Neural NetworkPoint Cloud
🎯 What it does: This paper proposes an iterative point cloud filtering framework based on Mamba, called 3DMambaIPF, and introduces a differentiable rendering loss to enhance surface detail recovery.
3DPGS: 3D Probabilistic Graph Search for Archaeological Piece Grouping
Junfeng Cheng (Imperial College London), Tania Stathaki (Imperial College London)
CodeGraph Neural NetworkPoint CloudBenchmark
🎯 What it does: A new benchmark called 'Archaeological Fragment Grouping' is proposed, along with the corresponding dataset ArcPie and new evaluation metrics;
3SAT: A Simple Self-Supervised Adversarial Training Framework
Jiang Fang (Institute of Information Engineering, Chinese Academy of Sciences), Wei Ma (Institute of Information Engineering, Chinese Academy of Sciences)
🎯 What it does: A 3SAT framework is proposed, utilizing raw unaugmented samples for self-supervised adversarial training, significantly enhancing the model's robustness and generalization performance.
🎯 What it does: A black-box semantic robustness evaluation framework is designed to quantify the worst-case robustness of BEV detection models under three types of semantic disturbances: geometric transformations, color shifts, and motion blur.
A Complete Algorithm for Optimization Modulo Nonlinear Real Arithmetic
Fuqi Jia (Institute of Software Chinese Academy of Sciences), Jian Zhang (Institute of Software Chinese Academy of Sciences)
CodeOptimization
🎯 What it does: The first complete OMT(NRA) solver is proposed, with the core algorithm being Optimization Cylindrical Algebraic Covering (OCAC), which is embedded in the CDCL(T) framework to form the CDCL(OCAC) solution.
A Comprehensive Overhaul of Multimodal Assistant with Small Language Models
Minjie Zhu (East China Normal University), Yaxin Peng (Shanghai University)
CodeOptimizationTransformerLarge Language ModelSupervised Fine-TuningVision Language ModelTextMultimodality
🎯 What it does: This paper proposes and implements a series of multimodal assistants based on small language models (≤3B parameters) — Mipha, systematically studying the impact of language models, visual representations, and optimization strategies on the performance of multimodal small language models (MSLMs), and validating their effectiveness through extensive benchmarking.
A Deep Probabilistic Framework for Continuous Time Dynamic Graph Generation
Ryien Hosseini (University of Chicago), Henry Hoffmann (University of Chicago)
CodeGenerationData SynthesisGraph Neural NetworkGraphTime Series
🎯 What it does: This paper proposes a continuous-time dynamic graph generation framework called DG-Gen, which directly models event probabilities and supports autoregressive generation and link prediction.
🎯 What it does: A simplified CNN model is trained using knowledge distillation, and hierarchical interpretability is achieved through average feature maps, intuitively demonstrating the diagnostic decision-making process.
A Lottery Ticket Hypothesis Approach with Sparse Fine-tuning and MAE for Image Forgery Detection and Localization
Jiaying Zhu (University of Science and Technology of China), Zheng-Jun Zha (University of Science and Technology of China)
CodeAnomaly DetectionTransformerMixture of ExpertsAuto EncoderImage
🎯 What it does: Proposes the Forgery Masked Autoencoder (FMAE), which utilizes the Lottery Ticket Hypothesis to select forgery-sensitive parameters for sparse fine-tuning based on MAE, and incorporates a multi-source noise extractor to achieve image forgery detection and localization;
A Matching-Based Algorithm for the Traveling Tournament Problem
Jingyang Zhao (University of Electronic Science and Technology of China), Mingyu Xiao (University of Electronic Science and Technology of China)
CodeOptimization
🎯 What it does: An effective algorithm based on minimum weight matching is designed to solve the travel tournament problem with a maximum of 4 consecutive away/home games (TTP-4).
🎯 What it does: This paper proposes an optimizer called MIAdam, which adds multiple integral terms based on Adam. It utilizes the low-pass filtering effect of integrals to suppress high-frequency noise in gradients, guiding the training process towards flat minima, thereby enhancing the model's generalization ability and robustness to label noise while maintaining Adam's fast convergence characteristics.
🎯 What it does: This paper proposes MFMB-Net, a multi-focus multi-branch network that combines fusion and reconstruction to achieve robust multimodal emotion analysis, particularly targeting scenarios with missing modalities.
A New Federated Learning Framework Against Gradient Inversion Attacks
Pengxin Guo (University of Hong Kong), Liangqiong Qu (University of Hong Kong)
CodeFederated LearningSafty and PrivacyImage
🎯 What it does: This paper proposes a federated learning framework called HyperFL that utilizes hypernetworks to generate local model parameters, aiming to break the direct link between shared gradients and local private data, thereby defending against gradient inversion attacks.
A Practical Approach to Causal Inference over Time
Martina Cinquini (University of Pisa), Isabel Valera (Saarland University)
CodeTime Series
🎯 What it does: This paper studies the causal effects over time in discrete-time dynamic systems, defines temporal interventions, and maps vector autoregression (VAR) models to structural causal models (SCM), achieving causal inference based on observed time series data.
CodeClassificationExplainability and InterpretabilityConvolutional Neural NetworkImage
🎯 What it does: This paper proposes an improved Component-Based Classification Network (CBC), closely related to RBF networks, and achieves interpretability and robustness through negative reasoning;
A Sample-Level Evaluation and Generative Framework for Model Inversion Attacks
Haoyang Li (Hong Kong Polytechnic University), Jianliang Xu (Hong Kong Baptist University)
CodeGenerationData SynthesisSafty and PrivacyAdversarial AttackConvolutional Neural NetworkGenerative Adversarial NetworkImage
🎯 What it does: This study investigates sample-level privacy assessment in model inversion attacks (MI) and proposes a new evaluation metric, DDCS, along with a GAN transfer learning enhancement framework based on natural gradient and entropy loss.
🎯 What it does: A new graph model GECO is proposed to replace the self-attention mechanism of traditional Graph Transformers, significantly improving computational efficiency and scalability while maintaining high-quality predictions.
A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics
Jiaxin Qi (Computer Network Information Center, Chinese Academy of Sciences), Gaogang Xie (Computer Network Information Center, Chinese Academy of Sciences)
🎯 What it does: This paper systematically identifies and addresses three overlooked issues in single-cell transcriptome data: long-tail distribution, model selection, and evaluation, proposing weighted sampling, model structure adaptation, and a unified benchmark.
🎯 What it does: A contrastive learning framework for short text classification called SimSTC is proposed, which does not require data augmentation and utilizes multi-view graph structures.
🎯 What it does: A method for selecting synthetic data without training (SDS) is proposed, utilizing CLIP for perturbation similarity assessment (PCS) and mIoU-based annotation filtering (ASF) to filter high-quality samples from a large number of synthetic image-annotation pairs for training semantic segmentation models.
A Unified Framework for Human-Allied Learning of Probabilistic Circuits
Athresh Karanam (University of Texas at Dallas), Sriraam Natarajan (University of Texas at Dallas)
CodeOptimizationData-Centric LearningPoint CloudTabularBiomedical Data
🎯 What it does: A unified framework is proposed to encode domain knowledge as differentiable equality/inequality constraints, and to achieve constraint satisfaction in the parameter learning of probabilistic circuits (PC) through penalty terms.
A Unifying Information-theoretic Perspective on Evaluating Generative Models
Alexis Fox (Duke University), Abhijin Adiga (University of Virginia)
CodeGenerationData SynthesisDiffusion modelImage
🎯 What it does: A unified information theory-based k-nearest neighbors precision/recall evaluation framework is proposed, and three-dimensional metrics (PCE, RCE, RE) are designed to simultaneously assess the authenticity, mode loss, and mode collapse of generative models.
A Video-grounded Dialogue Dataset and Metric for Event-driven Activities
Wiradee Imrattanatrai (National Institute of Advanced Industrial Science and Technology), Teruko Mitamura (Language Technologies Institute Carnegie Mellon University)
CodeLarge Language ModelVision Language ModelVideoText
🎯 What it does: The VDAct video dialogue dataset and VDEval evaluation metrics are proposed to support event-driven multi-scenario video dialogue tasks.
🎯 What it does: This paper proposes an adversarial attack-resistant network A2RNet for the fusion of infrared and visible light images, which can maintain high-quality fusion results and sustain downstream task performance in the presence of adversarial perturbations.
ABC3: Active Bayesian Causal Inference with Cohn Criteria in Randomized Experiments
Taehun Cha (Korea University), Donghun Lee (Korea University)
CodeOptimizationTabular
🎯 What it does: A Bayesian active learning-based random experiment design algorithm ABC3 is proposed for efficiently estimating the Conditional Average Treatment Effect (CATE) under limited samples.
Accurate and Regret-Aware Numerical Problem Solver for Tabular Question Answering
Yuxiang Wang (University of Melbourne), Junhao Gan (University of Melbourne)
CodeLarge Language ModelSupervised Fine-TuningTabularChain-of-Thought
🎯 What it does: To address the issue of inaccurate numerical calculations in table-based question answering, the TabLaP model is proposed: it uses a large language model (LLM) as a planner to generate Python calculation scripts, while the actual calculations are performed by an interpreter; it also integrates a state-of-the-art (SOTA) question answering branch, an answer selector (AnsSelector), and a credibility evaluator (TwEvaluator), forming a multi-branch, multi-LLM framework.
Accurate Estimation of Feature Importance Faithfulness for Tree Models
Mateusz Gajewski (University of Warsaw), Piotr Sankowski (University of Warsaw)
CodeTabular
🎯 What it does: This paper proposes a new prediction accuracy evaluation metric PGI² and provides an exact quadratic time complexity algorithm for tree models, while also designing a greedy feature ranking method based on this metric.
CodeDomain AdaptationExplainability and InterpretabilityProtein Structure PredictionGraph Neural NetworkSupervised Fine-TuningBiomedical Data
🎯 What it does: Developed the GeSite model, which combines domain-adaptive protein language models and E(3) equivariant graph neural networks to predict the binding residues of proteins with DNA or RNA;
🎯 What it does: A lightweight real-time super-resolution framework CGSR is proposed, which integrates rendering information (depth, normals, edges) into the network and achieves low parameters and efficient SR through NAS and dynamic pruning.
🎯 What it does: A 3D occupancy prediction method based on geometric-semantic decoupling, GSD-Occ, is proposed to achieve both real-time performance and high accuracy.
CodeLarge Language ModelPrompt EngineeringTextBenchmarkChain-of-Thought
🎯 What it does: This paper presents ACPBench—a large-scale benchmark based on PDDL, focused on seven reasoning tasks related to actions, changes, and planning, for evaluating the capabilities of LLMs in planning-related reasoning.
🎯 What it does: A motion-agnostic point-level (AAPL) annotation scheme without human intervention is proposed, along with a corresponding detection model and training method, achieving action instance detection under lightweight annotation.
Active Fourier Auditor for Estimating Distributional Properties of ML Models
Ayoub Ajarra (Inria), Debabrota Basu (Max Planck Institute for Software Systems)
CodeTabular
🎯 What it does: This paper proposes the Active Fourier Auditor (AFA), a unified auditing framework for black-box machine learning models that estimates the robustness, individual fairness, and group fairness of the model's distributional properties using Fourier coefficients.
Active Large Language Model-Based Knowledge Distillation for Session-Based Recommendation
Yingpeng Du (Nanyang Technological University), Yew-Soon Ong (Nanyang Technological University)
CodeRecommendation SystemKnowledge DistillationTransformerLarge Language ModelPrompt EngineeringSequential
🎯 What it does: A knowledge distillation method for LLM based on active learning (ALKDRec) is proposed, which efficiently distills LLM knowledge to lightweight recommendation models using a small number of instances in the conversational recommendation task.
🎯 What it does: This paper proposes an active reinforcement learning strategy that enhances the performance of offline policies by actively collecting information-rich trajectories using existing offline data under a limited budget.
🎯 What it does: This paper proposes a method for discovering active symbolic ODEs (Ordinary Differential Equations) through phase diagram sketching, utilizing an active learning strategy to select information-rich regions in phase space, and then sampling batches of initial conditions from these regions to query real trajectory data, thereby more efficiently discovering the symbolic dynamical equations of the system.
AD4CD: Causal-Guided Anomaly Detection for Enhancing Cognitive Diagnosis
Haiping Ma (Anhui University), Yong Yang (Anhui University)
CodeAnomaly DetectionAuto EncoderTime Series
🎯 What it does: The AD4CD framework is proposed, which combines causal inference and anomaly detection to capture abnormal states of students and questions through response time distribution, thereby improving the accuracy of cognitive diagnosis.
AdaCo: Overcoming Visual Foundation Model Noise in 3D Semantic Segmentation via Adaptive Label Correction
Pufan Zou (Xiamen University), Cheng Wang (Xiamen University)
CodeSegmentationAutonomous DrivingPoint Cloud
🎯 What it does: A label-free 3D semantic segmentation framework called AdaCo is proposed, which utilizes visual foundation models to generate cross-modal pseudo-labels and iteratively improves segmentation accuracy through adaptive noise correction and adaptive robust loss.
🎯 What it does: Proposes AdaGK-SGD, an algorithm that guides global knowledge in distributed training without additional communication through a maximum lifetime global knowledge module.
🎯 What it does: A multi-armed bandit (MAB) driven retrieval-enhanced generation framework is proposed, utilizing a knowledge graph as the underlying knowledge base to achieve dynamic selection and real-time optimization of various retrieval methods.
🎯 What it does: A training-free ANN-to-SNN conversion framework is proposed, combining Adaptive Firing Neurons (AdaFire), Layer Sensitivity Threshold Compression (SSC), and Input-Aware Adaptive Time Steps (IAT), significantly improving the accuracy and energy efficiency of the converted SNN.
Adaptive Computation Modules: Granular Conditional Computation for Efficient Inference
Bartosz Wójcik (Jagiellonian University), Simone Scardapane (Sapienza University of Rome)
CodeComputational EfficiencyKnowledge DistillationTransformerSupervised Fine-TuningMixture of ExpertsImageAudio
🎯 What it does: This paper proposes Adaptive Computation Modules (ACM) for the Transformer model, enabling dynamic adjustment of computation width for each token as needed, thereby reducing inference costs.
🎯 What it does: Proposes Adaptive Dataset Quantization (ADQ), which compresses datasets using an adaptive sampling method by evaluating the representativeness, richness, and importance of each bin based on Dataset Quantization.
🎯 What it does: A pluggable adaptive decision boundary strategy (ADBS) is proposed for few-shot incremental learning (FSCIL) to dynamically allocate decision boundaries for each category and further enhance inter-class separability through inter-class constraint (IC).
Adaptive Draft-Verification for Efficient Large Language Model Decoding
Xukun Liu (Northwestern University), Dongkuan (DK) Xu (North Carolina State University)
CodeGenerationComputational EfficiencyTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper presents Adaptix, an adaptive draft-validation scheme that does not require fine-tuning. It utilizes a triplet matrix to approximate the output distribution of LLMs and generates drafts that are highly consistent with the LLM distribution through MCTS, significantly accelerating LLM decoding.
Adaptive Multi-Scale Decomposition Framework for Time Series Forecasting
Yifan Hu (Tsinghua University), Tao Dai (Shenzhen University)
CodeOptimizationMixture of ExpertsTime Series
🎯 What it does: An adaptive multi-scale decomposition framework (AMD) based on MLP is proposed, which achieves finer time series prediction through multi-scale splitting, mixing, dual dependency interaction, and adaptive expert fusion of time series.
🎯 What it does: Proposes the Adapter method to address the catastrophic forgetting problem caused by model representation shift during prototype replay in class-incremental semantic segmentation, and enhances classification performance through uncertainty constraints and prototype distinction.
Adaptive Sampling to Reduce Epistemic Uncertainty Using Prediction Interval-Generation Neural Networks
Giorgio Morales (Montana State University), John W. Sheppard (Montana State University)
CodeTabularTime SeriesAgriculture Related
🎯 What it does: This paper proposes an adaptive sampling method based on a predictive interval generating neural network (ASPINN), aimed at reducing the model's epistemic uncertainty through active sampling.
🎯 What it does: This paper proposes an Approximation Decision Boundary Approach (ADBA) and its variant ADBA-md, which quickly compares and optimizes perturbation directions in black-box decision attacks, significantly reducing the number of queries.
Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling
Wenqiao Zhu (HiThink Research), Jun Wu (HiThink Research)
CodeRecommendation SystemDiffusion modelTabular
🎯 What it does: A supervised diffusion model (CSDM) was designed and implemented to generate warm-up embeddings for cold-start projects, thereby improving the accuracy of CTR prediction.
🎯 What it does: The ADELA method is proposed, which accelerates the design of machine learning pipelines in evolutionary AutoML by constructing an accompanying surrogate model.
🎯 What it does: A distributionally robust recommendation loss DrRL based on R' enyi divergence is proposed to unify the advantages of Softmax Loss and Cosine Contrastive Loss while overcoming their limitations.
🎯 What it does: A retrieval-augmented molecular graph generation framework RARB is proposed for template-free single-step backward synthesis prediction, implemented on the RetroBridge base model.
CodeClassificationSpiking Neural NetworkImageTime Series
🎯 What it does: A pluggable multi-scale attention module (SMA) and attention-based zone-out regularization (AZO) are proposed, enabling spiking neural networks to simultaneously utilize multi-scale features and spatiotemporal correlations, thereby enhancing the learning effectiveness on event stream data.
🎯 What it does: This paper proposes an Adversarial Contrastive Graph Augmentation framework (ACGA), which automatically generates positive samples containing minimal sufficient information and difficult negative samples through a conditional variational graph autoencoder, thereby enhancing model robustness in unsupervised graph representation learning.
🎯 What it does: AeroGTO is designed and implemented, an operator that combines graph neural networks and transformers to quickly and accurately predict surface pressure and drag coefficients on large-scale meshes of three-dimensional vehicle geometries.
AFFAKT: A Hierarchical Optimal Transport Based Method for Affective Facial Knowledge Transfer in Video Deception Detection
Zihan Ji (South China University of Technology), Ye Liu (Beijing Normal University)
CodeClassificationDomain AdaptationAnomaly DetectionExplainability and InterpretabilityRecurrent Neural NetworkSupervised Fine-TuningVideoMultimodalityAudio
🎯 What it does: A hierarchical optimal transport-based emotional facial knowledge transfer method (AFFAKT) is proposed to enhance video deception detection performance.
Affirm: Interactive Mamba with Adaptive Fourier Filters for Long-term Time Series Forecasting
Yuhan Wu (Zhejiang University), Dongming Lu (Zhejiang University)
CodeTime Series
🎯 What it does: A lightweight time series forecasting model called Affirm is proposed, which combines adaptive Fourier filters and a dual-interaction Mamba module to achieve efficient modeling of long sequences.
AFiRe: Anatomy-Driven Self-Supervised Learning for Fine-Grained Representation in Radiographic Images
Yihang Liu (Tongji University), Hongzhou Chen (Tongji University)
CodeClassificationSegmentationAnomaly DetectionRepresentation LearningTransformerContrastive LearningImageBiomedical Data
🎯 What it does: The AFiRe framework is proposed, combining anatomy-driven self-supervised learning with ViT-based token-level contrastive learning and pixel-level anomaly removal reconstruction to achieve fine-grained representation of chest X-rays.
Against All Odds: Overcoming Typology, Script, and Language Confusion in Multilingual Embedding Inversion Attacks
Yiyi Chen (Aalborg University), Johannes Bjerva (Aalborg University)
CodeAdversarial AttackTransformerLarge Language ModelText
🎯 What it does: Evaluate embedding inversion attacks on multilingual large language models, studying vulnerabilities across languages and scripts, and exploring language confusion phenomena.
Agent4Edu: Generating Learner Response Data by Generative Agents for Intelligent Education Systems
Weibo Gao (University of Science and Technology of China), Zhenya Huang (University of Science and Technology of China)
CodeGenerationData SynthesisRecommendation SystemTransformerLarge Language ModelReinforcement LearningAgentic AITextSequentialChain-of-Thought
🎯 What it does: Using large language model-driven generative agents to simulate personalized learner practice response data on online education platforms and record their learning processes.
AI-generated Image Quality Assessment in Visual Communication
Yu Tian (City University of Hong Kong), Sam Kwong (Lingnan University)
CodeGenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageMultimodality
🎯 What it does: This study constructed the AIGI-VC dataset, which includes 2,500 AI-generated images covering 14 advertising themes and 8 types of emotions, and provides preference annotations from coarse to fine, used to evaluate the clarity of information and emotional interaction in visual communication.
Tian Li (Henan Key Laboratory of Quantum Information and Cryptography), He-Liang Huang (Henan Key Laboratory of Quantum Information and Cryptography)
CodeOptimizationReinforcement LearningTabularBenchmarkPhysics Related
🎯 What it does: Dynamically designing the topology of quantum processors and mapping quantum circuits through reinforcement learning significantly reduces circuit depth.
AIM: Let Any Multimodal Large Language Models Embrace Efficient In-Context Learning
Jun Gao (Soochow University), Wenjie Li (Hong Kong Polytechnic University)
CodeTransformerLarge Language ModelVision Language ModelImageTextMultimodality
🎯 What it does: The AIM framework is proposed, enabling any multimodal large language model to achieve efficient context learning without updating model parameters.
AIR: Unifying Individual and Collective Exploration in Cooperative Multi-Agent Reinforcement Learning
Guangchong Zhou (Chinese Academy of Sciences), Guoliang Fan (Chinese Academy of Sciences)
CodeReinforcement Learning
🎯 What it does: The AIR (Adaptive exploration via Identity Recognition) framework is proposed, which implements individual and collective exploration in value-based multi-agent reinforcement learning through a unified identity recognizer, and dynamically adjusts the exploration method via adaptive temperature.
Aligning and Prompting Anything for Zero-Shot Generalized Anomaly Detection
Jitao Ma (Xidian University), Leyuan Fang (Hunan University)
CodeAnomaly DetectionTransformerPrompt EngineeringVision Language ModelImageBiomedical DataMagnetic Resonance ImagingComputed Tomography
🎯 What it does: This paper proposes a zero-shot generalized anomaly detection (ZGAD) method based on Text Prompt Shunt (TPS), which can simultaneously achieve image-level anomaly classification and pixel-level anomaly segmentation.
Aligning Language Models Using Follow-up Likelihood as Reward Signal
Chen Zhang (National University of Singapore), Haizhou Li (Tencent AI Lab)
CodeTransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a new reward mechanism called 'Follow-up Likelihood as Reward (FLR)', which uses the likelihood of user follow-up statements after generating replies as an unannotated reward signal. It constructs preference data by automating the annotation of online generated replies and further enhances model helpfulness using DAP (such as DPO).
Alignment-Free RGB-T Salient Object Detection: A Large-Scale Dataset and Progressive Correlation Network
Kunpeng Wang (Anhui University), Bin Luo (Anhui University)
CodeObject DetectionTransformerImageMultimodality
🎯 What it does: This paper proposes a large-scale unaligned RGB-Thermal visual salient object detection dataset named UVT20K, and designs a Progressive Correlation Network (PCNet) based on this dataset for salient object detection in unaligned image pairs.
AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors
Hao Shi (University of Chinese Academy of Sciences), Luis Angel Seco (University of Toronto)
CodeRecommendation SystemOptimizationExplainability and InterpretabilityComputational EfficiencyReinforcement LearningGenerative Adversarial NetworkTabularTime SeriesFinance Related
🎯 What it does: The AlphaForge framework is proposed, which is divided into two stages: a generative predictive network is used to mine formalized Alpha factors, and then a dynamic factor combination model dynamically adjusts the weights based on the latest performance of the factors to form a Mega-Alpha signal.
🎯 What it does: This paper proposes a dynamic learning rate adjustment method based on memory rate, ALRMR-GEC, to improve the generalization performance of edit-based grammar error correction models.
Amplifier: Bringing Attention to Neglected Low-Energy Components in Time Series Forecasting
Jingru Fei (Beijing Institute of Technology), Zhendong Niu (Beijing Institute of Technology)
CodeTransformerTime Series
🎯 What it does: Proposes energy amplification and recovery technology, and constructs an Amplifier model based on this technology to enhance the learning of low-energy spectral components, thereby improving time series forecasting.
An Automatic Sound and Complete Abstraction Method for Generalized Planning with Baggable Types
Hao Dong (Sun Yat-sen University), Yongmei Liu (Sun Yat-sen University)
Code
🎯 What it does: This paper proposes an automated abstract method with soundness and completeness for generating Bounded QNP (BQNP) abstractions from STRIPS domain instances containing countable types (baggable types) to solve generalized planning problems.
🎯 What it does: A dynamic sparse training framework based on parameter freezing and server-side sparse mask re-tuning is proposed in federated learning, significantly reducing communication and computation costs.
🎯 What it does: A framework based on diffusion models (DBMEF) is proposed, which provides a 're-thinking' function for existing discriminative models without the need for additional training, significantly improving classification accuracy.
An Elite-guided Weighted Simulated Annealing Algorithm for the Clique Partitioning Problem
Baiyu Chen (Huazhong University of Science and Technology), Zhipeng Lü (Huazhong University of Science and Technology)
CodeOptimizationGraph
🎯 What it does: An elite-guided weighted simulated annealing algorithm named EWSA is proposed and implemented to solve the clique partition problem (CPP) of complete graphs, improving search efficiency and solution quality through the alternating use of two search configurations, a weighted scoring function, and partition constraint strategies.
An Evaluation Framework for Product Images Background Inpainting Based on Human Feedback and Product Consistency
Yuqi Liang (Ant Group), Jianqi Bi (Ant Group)
CodeImage TranslationRestorationSegmentationReinforcement Learning from Human FeedbackTransformerImageMultimodality
🎯 What it does: This paper proposes an automatic evaluation framework named HFPC for detecting the background appropriateness and product consistency of product images after AI background filling.
An Exemplar-based Framework for Chinese Text Recognition
Zhao Zhou (Fudan University), Cheng Jin (Fudan University)
CodeRecognitionRetrievalTransformerText
🎯 What it does: A Chinese text recognition framework based on sample retrieval is proposed—DECTR, which first discovers character samples through a weakly supervised feature extraction network, and then retrieves similar characters from an external sample library to correct recognition errors.
CodeAnomaly DetectionAutonomous DrivingOptimizationLarge Language ModelPrompt EngineeringMultimodality
🎯 What it does: An adaptive evolutionary algorithm based on LLM (µ MOEA) is proposed for the detection of security violations in multi-component deep learning systems.
An Optimal Transport-based Latent Mixer for Robust Multi-modal Learning
Fengjiiao Gong, Hongteng Xu (Renmin University of China)
CodeClassificationFederated LearningSafty and PrivacyRepresentation LearningAuto EncoderMultimodality
🎯 What it does: Designed and implemented an OTM mixer based on optimal transport, which performs privacy-friendly and robust feature fusion and enhancement of multi-head Wasserstein autoencoders on distributed, unaligned multimodal data;
Analytical-Chemistry-Informed Transformer for Infrared Spectra Modeling
Shiluo Huang (Southwestern University of Finance and Economics), Ying Mu (Zhejiang University)
CodeTransformerTabularAgriculture Related
🎯 What it does: An Analytical-Chemistry-Informed Transformer (ACT) aimed at infrared spectrum modeling is proposed, achieving domain-invariant representation through a learnable baseline correction module and spectral attention mechanism to address the calibration transfer problem.
Anchor Learning with Potential Cluster Constraints for Multi-view Clustering
Yawei Chen (Dalian Maritime University), Yang Wang (Hefei University of Technology)
CodeOptimizationImageVideo
🎯 What it does: This paper proposes an Anchor Learning with Potential Cluster Constraints (ALPC) method, which unifies the selection of anchors and the construction of anchor graphs into a single framework, and achieves a uniform distribution of anchors across clusters and alignment with the original data cluster centers through potential cluster constraints.
AnchorInv: Few-Shot Class-Incremental Learning of Physiological Signals via Feature Space-Guided Inversion
Chenqi Li (University of Oxford), Tingting Zhu (University of Oxford)
CodeClassificationData SynthesisConvolutional Neural NetworkTransformerTime SeriesBiomedical Data
🎯 What it does: This paper proposes the AnchorInv method, which utilizes anchor points in the feature space to guide model inversion, generating synthetic samples of previously learned categories to achieve few-shot class incremental learning and prevent catastrophic forgetting.
Anti-Diffusion: Preventing Abuse of Modifications of Diffusion-Based Models
Li Zheng (University of Macau), Jinyu Tian (Macau University of Science and Technology)
CodeSafty and PrivacyTransformerPrompt EngineeringDiffusion modelImage
🎯 What it does: Proposes the Anti-Diffusion privacy protection system, which adds small adversarial noise to images before release to prevent their misuse by diffusion model-based tuning and editing methods.
🎯 What it does: This paper proposes a new single destruction heuristic called ADDRESS, which utilizes constrained Thompson Sampling to select the agent with the highest delay as a seed, improving the LNS neighborhood generation method of traditional MAPF-LNS.
APAR: Modeling Irregular Target Functions in Tabular Regression via Arithmetic-Aware Pre-Training and Adaptive-Regularized Fine-Tuning
Hong-Wei Wu (National Yang Ming Chiao Tung University), Wen-Chih Peng (National Yang Ming Chiao Tung University)
CodeTransformerTabular
🎯 What it does: By using arithmetic-aware pre-training and adaptive regularization fine-tuning, we improve the table regression model to address the overfitting problem caused by irregular objective functions.
APKGC: Noise-enhanced Multi-Modal Knowledge Graph Completion with Attention Penalty
Yue Jian (Hubei University), Xiaoju Hou (Guangdong Industry Polytechnic University)
CodeGraph Neural NetworkTransformerLarge Language ModelVision Language ModelMultimodalityGraph
🎯 What it does: The paper proposes the APKGC method, which uses attention penalty and adaptive noise sampling to complete multimodal knowledge graph completion.
Rayna Andreeva (University of Edinburgh), Rik Sarkar (University of Edinburgh)
CodeOptimizationTabular
🎯 What it does: This study proposes various algorithms for fast approximation of the size of metric spaces (Magnitude), including convex optimization frameworks, iterative normalization, greedy subset selection, and discrete center hierarchy, and applies them to neural network regularization and clustering tasks.
🎯 What it does: A two-stage self-supervised denoising framework APR-RD is proposed: first, the Adjacent Pixel Replacer (APR) generates decorrelated noise pairs without downsampling, and then the Recharged Distillation (RD) enhances detail recovery through multi-target distillation, all trained on a single noisy image.