Latest AI Papers: Time Series, GNNs, Diffusion Models (Oct 10, 2025)

Hey guys! Check out the freshest finds in AI research! This article highlights the latest papers from October 10, 2025, across several exciting categories. Big thanks to JeremyChou28 and the Daily-Arxiv-Tools for curating this list. For the best reading experience, don't forget to visit the Github page. Let's dive in!

Time Series

Time series analysis remains a hot topic, and these papers showcase some of the latest innovations. From parameter-efficient neural networks to advanced forecasting techniques, there's something for everyone interested in this field. We will break down the latest research and what makes them essential for you to stay ahead.

NdLinear: Preserving Multi-Dimensional Structure for Parameter-Efficient Neural Networks

NdLinear focuses on improving neural network efficiency by preserving multi-dimensional structures. Guys, imagine creating neural networks that are not only powerful but also incredibly efficient! This paper introduces a novel approach to reduce the parameter count without sacrificing performance. The code is available on GitHub, making it easy to implement and experiment with this technique. The main idea is to maintain the integrity of multi-dimensional data while minimizing computational overhead, leading to faster and more scalable models. This approach is particularly useful in applications where resources are limited, such as edge computing and mobile devices. Moreover, the preservation of multi-dimensional structure ensures that the essential features of the data are retained, resulting in more accurate predictions and better overall performance. Researchers and practitioners alike can benefit from this work by incorporating NdLinear into their models to achieve state-of-the-art results with fewer resources. So, if you're looking to optimize your neural networks, NdLinear is definitely worth checking out!

GTCN-G: A Residual Graph-Temporal Fusion Network for Imbalanced Intrusion Detection

GTCN-G introduces a residual graph-temporal fusion network designed to tackle imbalanced intrusion detection. This preprint, submitted to IEEE TrustCom 2025, presents a cutting-edge solution for identifying security threats in complex systems. The key innovation lies in its ability to fuse graph and temporal information effectively, allowing it to detect subtle patterns that might be missed by traditional methods. The residual connections in the network help to mitigate the vanishing gradient problem, ensuring that the model can learn from long sequences of data. Additionally, the graph component enables the network to capture relationships between different entities in the system, providing a more holistic view of the security landscape. This is particularly important in detecting coordinated attacks that involve multiple actors. The imbalanced nature of intrusion detection datasets, where normal behavior far outweighs malicious activity, is also addressed by the GTCN-G architecture, making it a robust and reliable solution for real-world applications. By combining these advanced techniques, GTCN-G sets a new standard for intrusion detection systems.

HTMformer: Hybrid Time and Multivariate Transformer for Time Series Forecasting

HTMformer is a hybrid time and multivariate transformer tailored for time series forecasting. This model combines the strengths of both time-domain and multivariate analysis to deliver enhanced forecasting accuracy. By leveraging the transformer architecture, HTMformer can capture long-range dependencies in the data, which is crucial for predicting future values accurately. The hybrid approach allows it to handle complex time series data with multiple variables, making it suitable for a wide range of applications. For example, in financial forecasting, HTMformer can consider various economic indicators and their relationships over time to predict stock prices or market trends. Similarly, in weather forecasting, it can integrate temperature, humidity, wind speed, and other factors to provide more reliable predictions. The ability to process multivariate data efficiently also makes HTMformer ideal for industrial applications, such as predicting equipment failures or optimizing production processes. This model represents a significant step forward in time series forecasting, offering a powerful tool for researchers and practitioners seeking to improve their predictive capabilities.

Causal Feedback Discovery using Convergence Cross Mapping on Sea Ice Data

Causal Feedback Discovery explores the use of convergence cross mapping to uncover causal relationships in sea ice data. Accepted at the ACM Sigspatial Conference, PolDS Workshop, this research provides valuable insights into the complex dynamics of sea ice and its impact on the environment. By applying convergence cross mapping, the study identifies feedback loops and causal links between different variables, such as temperature, salinity, and ice thickness. Understanding these relationships is crucial for predicting sea ice behavior and its effects on climate change. The findings can help scientists develop more accurate models for forecasting sea ice extent and thickness, which is essential for navigation, resource management, and climate policy. Moreover, the methodology used in this paper can be applied to other complex systems, making it a valuable tool for causal discovery in various scientific domains. The use of real-world sea ice data adds to the practical relevance of this research, making it a significant contribution to the field of environmental science.

Lossless Compression of Time Series Data: A Comparative Study

Lossless Compression provides a comparative analysis of techniques for lossless compression of time series data. In this era of big data, the ability to efficiently store and transmit time series data is paramount. This study evaluates various lossless compression algorithms, assessing their performance in terms of compression ratio, speed, and memory usage. By identifying the most effective methods, this research helps practitioners optimize their data storage and transmission strategies. The insights gained from this comparative study are particularly valuable in applications where data integrity is critical, such as medical monitoring, financial analysis, and scientific research. Lossless compression ensures that no information is lost during the compression process, preserving the accuracy and reliability of the data. The findings of this study can guide researchers and engineers in selecting the most appropriate compression techniques for their specific needs, leading to more efficient and cost-effective data management practices.

Spatio Temporal

Spatio-temporal data combines spatial and temporal dimensions, offering unique challenges and opportunities. These papers cover everything from humanoid world modeling to predicting COVID-19 dynamics, showcasing the breadth of research in this area. Stay tuned as we uncover their important points and impacts to real world applications.

AbsoluteNet: A Deep Learning Neural Network to Classify Cerebral Hemodynamic Responses of Auditory Processing

AbsoluteNet is a deep learning neural network designed to classify cerebral hemodynamic responses during auditory processing. This innovative approach leverages the power of deep learning to analyze brain activity patterns, providing valuable insights into how the brain processes sound. By accurately classifying hemodynamic responses, AbsoluteNet can aid in the diagnosis of auditory disorders and the development of more effective treatments. The network's ability to capture complex relationships in the data makes it a powerful tool for neuroscience research. Moreover, AbsoluteNet has the potential to be used in real-time applications, such as brain-computer interfaces, allowing individuals with hearing impairments to interact with their environment more effectively. The development of such advanced neural networks represents a significant step forward in our understanding of the brain and its functions, paving the way for new and improved diagnostic and therapeutic interventions.

Generative World Modelling for Humanoids: 1X World Model Challenge Technical Report

Generative World Modelling focuses on creating world models for humanoids, as detailed in the 1X World Model Challenge Technical Report. This 6-page report outlines the challenges and solutions involved in building AI systems that can understand and interact with the world like humans do. The report covers various aspects of world modeling, including perception, action, and reasoning. By developing generative models, researchers aim to create humanoids that can not only perceive their environment but also predict future states and plan accordingly. This is crucial for enabling humanoids to perform complex tasks in real-world settings. The technical report provides valuable insights into the current state of the art in world modeling, highlighting both the successes and the remaining challenges. The 1X World Model Challenge serves as a benchmark for evaluating the performance of different approaches, driving innovation and progress in the field of robotics and artificial intelligence.

Fully Spiking Neural Networks for Unified Frame-Event Object Tracking

Fully Spiking Neural Networks presents an approach for unified frame-event object tracking using spiking neural networks. Accepted by NeurIPS 2025, this research explores the use of biologically inspired neural networks for processing both traditional frames and event-based data. By combining these two modalities, the network can track objects more accurately and efficiently, even in challenging conditions such as low light or high motion. Spiking neural networks offer several advantages over traditional deep learning models, including energy efficiency and temporal processing capabilities. This makes them particularly well-suited for real-time applications, such as autonomous driving and robotics. The ability to process event-based data also allows the network to respond quickly to changes in the environment, improving its robustness and adaptability. This work represents a significant step forward in the development of more intelligent and efficient object tracking systems.

Memory-Augmented Generative AI for Real-time Wireless Prediction in Dynamic Industrial Environments

Memory-Augmented Generative AI introduces a system for real-time wireless prediction in dynamic industrial environments. This system combines generative AI with memory augmentation to provide accurate and timely predictions of wireless network performance. By leveraging past data and current conditions, the system can anticipate changes in the environment and optimize network parameters accordingly. This is crucial for maintaining reliable wireless connectivity in industrial settings, where disruptions can lead to significant losses. The memory component allows the system to remember past events and adapt its predictions over time, improving its accuracy and robustness. Generative AI enables the system to generate realistic scenarios and simulate different operating conditions, allowing it to proactively identify and mitigate potential problems. This innovative approach has the potential to transform industrial wireless network management, ensuring seamless connectivity and optimal performance.

Progressive Gaussian Transformer with Anisotropy-aware Sampling for Open Vocabulary Occupancy Prediction

Progressive Gaussian Transformer presents an anisotropy-aware sampling method for open vocabulary occupancy prediction. The project page showcases the capabilities of this innovative approach, which combines a Gaussian transformer with progressive sampling techniques. By considering the anisotropy of the data, the method can more accurately predict occupancy in complex environments. This is particularly useful in applications such as autonomous driving and robotics, where understanding the surrounding environment is critical. The open vocabulary approach allows the system to handle a wide range of objects and scenarios, making it more versatile and adaptable. The progressive sampling technique improves the efficiency of the prediction process, allowing it to operate in real-time. This work represents a significant advancement in the field of occupancy prediction, enabling more accurate and reliable environmental understanding.

Time Series Imputation

Time series imputation is essential for handling missing data in time series analysis. These papers delve into various techniques, from information bottleneck approaches to state transition diffusion frameworks, providing a comprehensive overview of the latest advancements. Lets check them out!

Glocal Information Bottleneck for Time Series Imputation

Glocal Information Bottleneck introduces a novel approach to time series imputation that leverages both global and local information. This method aims to fill in missing values in time series data by considering the broader context as well as the immediate surroundings of the missing data points. By combining these two perspectives, the Glocal Information Bottleneck can achieve more accurate and robust imputations. The information bottleneck principle is used to extract the most relevant information from the data, reducing noise and improving the quality of the imputations. This approach is particularly useful in applications where the time series data is noisy or contains complex patterns. The Glocal Information Bottleneck represents a significant advancement in time series imputation, offering a powerful tool for researchers and practitioners seeking to improve the quality of their data analysis.

A Structure-Preserving Assessment of VBPBB for Time Series Imputation Under Periodic Trends, Noise, and Missingness Mechanisms

Structure-Preserving Assessment evaluates the performance of VBPBB for time series imputation under various conditions. This 24-page paper examines how well VBPBB preserves the underlying structure of the time series data when imputing missing values. The assessment considers periodic trends, noise, and different missingness mechanisms, providing a comprehensive evaluation of VBPBB's capabilities. The results of this study can help practitioners understand the strengths and limitations of VBPBB, guiding them in selecting the most appropriate imputation method for their specific needs. Structure preservation is crucial for maintaining the integrity of the time series data, ensuring that subsequent analyses are accurate and reliable. This research contributes to the growing body of knowledge on time series imputation, offering valuable insights for both researchers and practitioners.

STDiff: A State Transition Diffusion Framework for Time Series Imputation in Industrial Systems

STDiff presents a state transition diffusion framework designed for time series imputation in industrial systems. This framework leverages the power of diffusion models to fill in missing values in time series data, taking into account the underlying state transitions of the system. By modeling the dynamics of the industrial system, STDiff can generate more realistic and accurate imputations. The state transition component allows the framework to capture complex relationships between different variables, improving its ability to handle noisy or incomplete data. This approach is particularly useful in industrial applications, where the time series data often reflects the state of complex machinery and processes. STDiff represents a significant advancement in time series imputation, offering a powerful tool for improving the reliability and accuracy of industrial data analysis.

Irregular Time Series

Irregular time series pose unique challenges due to their unevenly spaced data points. These papers explore innovative methods for forecasting, classification, and causal discovery in such data, pushing the boundaries of what's possible. These papers are a must read if you are dealing with irregular data.

ASTGI: Adaptive Spatio-Temporal Graph Interactions for Irregular Multivariate Time Series Forecasting

ASTGI introduces an adaptive spatio-temporal graph interactions approach for forecasting irregular multivariate time series. This method leverages graph neural networks to capture the complex relationships between different variables in the time series, while also considering the spatial and temporal aspects of the data. The adaptive component allows the network to adjust its parameters based on the characteristics of the data, improving its ability to handle irregular time series. This approach is particularly useful in applications where the time series data is sparse or contains missing values. ASTGI represents a significant advancement in irregular time series forecasting, offering a powerful tool for researchers and practitioners seeking to improve their predictive capabilities.

Mind the Missing: Variable-Aware Representation Learning for Irregular EHR Time Series using Large Language Models

Mind the Missing explores variable-aware representation learning for irregular EHR time series using large language models. This research focuses on developing techniques to handle missing data in electronic health records (EHR) by leveraging the power of large language models. By learning representations that are sensitive to the presence or absence of different variables, the method can more accurately analyze irregular EHR time series. This is crucial for improving the quality of healthcare data analysis and decision-making. The use of large language models allows the method to capture complex relationships between different variables, improving its ability to handle noisy or incomplete data. Mind the Missing represents a significant advancement in irregular time series analysis, offering a powerful tool for researchers and practitioners in the healthcare domain.

DeNOTS: Stable Deep Neural ODEs for Time Series

DeNOTS presents stable deep neural ODEs for time series analysis. This method leverages the power of ordinary differential equations (ODEs) to model the dynamics of time series data. By using deep neural networks to parameterize the ODEs, DeNOTS can capture complex relationships in the data. The stability of the ODEs is crucial for ensuring that the model produces reliable and accurate results. This approach is particularly useful in applications where the time series data is noisy or contains missing values. DeNOTS represents a significant advancement in time series analysis, offering a powerful tool for researchers and practitioners seeking to improve their understanding of complex systems.

Diffusion Model

Diffusion models are gaining traction in various AI applications. These papers explore their use in video processing, image generation, and even robot manipulation, highlighting the versatility and potential of this technology. Get ready to learn about the next gen of diffusion model applications.

CaRDiff: Video Salient Object Ranking Chain of Thought Reasoning for Saliency Prediction with Diffusion

CaRDiff introduces a chain-of-thought reasoning approach for saliency prediction in videos using diffusion models. Accepted to AAAI 2025, this research explores how diffusion models can be used to identify the most salient objects in a video by reasoning through a series of steps. The chain-of-thought approach allows the model to consider the context and relationships between different objects, improving its ability to identify the most important elements in the scene. This is particularly useful in applications such as video summarization and object tracking. CaRDiff represents a significant advancement in video processing, offering a powerful tool for researchers and practitioners seeking to improve their understanding of complex scenes.

Security-Robustness Trade-offs in Diffusion Steganography: A Comparative Analysis of Pixel-Space and VAE-Based Architectures

Security-Robustness Trade-offs analyzes the security and robustness of diffusion steganography in pixel-space and VAE-based architectures. This research explores the trade-offs between hiding information securely and ensuring that the hidden information is robust to attacks. By comparing different architectures, the study provides insights into the strengths and weaknesses of each approach. This is crucial for developing secure and reliable steganographic systems. The analysis considers various attack scenarios, providing a comprehensive evaluation of the security and robustness of the different architectures. This work represents a significant contribution to the field of steganography, offering valuable guidance for researchers and practitioners seeking to develop more secure and robust systems.

Generative AI for Cel-Animation: A Survey

Generative AI for Cel-Animation provides a comprehensive survey of the use of generative AI in cel-animation. Accepted by ICCV 2025 AISTORY Workshop, this survey covers various techniques for generating cel-animated content using AI, including generative adversarial networks (GANs) and variational autoencoders (VAEs). The survey also discusses the challenges and opportunities in this field, providing a valuable resource for researchers and practitioners. By summarizing the current state of the art, this survey helps to identify promising directions for future research. The use of generative AI has the potential to transform the cel-animation industry, enabling the creation of high-quality content more efficiently and cost-effectively.

Graph Neural Networks

Graph Neural Networks (GNNs) are revolutionizing how we process graph-structured data. From traffic anomaly detection to multi-omics data integration, these papers showcase the diverse applications and ongoing advancements in GNN research. Here are some of the hottest topics on GNNs.

GTCN-G: A Residual Graph-Temporal Fusion Network for Imbalanced Intrusion Detection (Preprint)

GTCN-G presents a residual graph-temporal fusion network for imbalanced intrusion detection. This preprint, submitted to IEEE TrustCom 2025, introduces a novel approach to detect intrusions in complex systems by combining graph and temporal information. The residual connections in the network help to mitigate the vanishing gradient problem, ensuring that the model can learn from long sequences of data. The graph component enables the network to capture relationships between different entities in the system, providing a more holistic view of the security landscape. This is particularly important in detecting coordinated attacks that involve multiple actors. The imbalanced nature of intrusion detection datasets, where normal behavior far outweighs malicious activity, is also addressed by the GTCN-G architecture, making it a robust and reliable solution for real-world applications.

AMBER: Adaptive Mesh Generation by Iterative Mesh Resolution Prediction

AMBER introduces an adaptive mesh generation method based on iterative mesh resolution prediction. Accepted to NeurIPS 2025, this research explores how to generate high-quality meshes for computer graphics and simulations by iteratively refining the mesh resolution. The adaptive component allows the method to adjust the mesh resolution based on the complexity of the underlying geometry, improving the accuracy and efficiency of the mesh generation process. This is particularly useful in applications such as finite element analysis and 3D modeling. AMBER represents a significant advancement in mesh generation, offering a powerful tool for researchers and practitioners seeking to improve the quality of their meshes.

GNN-enhanced Traffic Anomaly Detection for Next-Generation SDN-Enabled Consumer Electronics

GNN-enhanced Traffic Anomaly Detection presents a system for detecting traffic anomalies in next-generation SDN-enabled consumer electronics using graph neural networks. This paper has been accepted for publication in IEEE Transactions on Consumer Electronics. This system leverages the power of GNNs to analyze network traffic patterns and identify unusual behavior that may indicate a security threat. By modeling the network as a graph, the system can capture complex relationships between different devices and users, improving its ability to detect anomalies. The use of SDN enables the system to dynamically adjust network parameters in response to detected anomalies, mitigating the impact of potential attacks. This approach is particularly relevant in the context of the Internet of Things (IoT), where the increasing number of connected devices poses new security challenges. This work represents a significant contribution to the field of network security, offering a powerful tool for protecting consumer electronics from cyber threats.

That's all for today, guys! Stay tuned for more updates on the latest AI research.

External Links:

• For more information on Graph Neural Networks, check out Distill's explanation of Graph Neural Networks.