Latest Papers: Time Series, Traffic, Spatial-Temporal - Nov 2025

Stay up-to-date with the cutting edge research in time series analysis, traffic management, and spatio-temporal modeling. This article summarizes the latest papers from HeyZiy's DailyArXiv, providing a quick overview of the most recent advancements in these fields. For a better reading experience and access to even more papers, be sure to check out the Github page. This curated list covers publications up to November 28, 2025, ensuring you're informed about the newest developments.

Time Series Forecasting

Time series forecasting remains a crucial area of research, with applications spanning finance, weather prediction, and many other domains. The latest papers explore innovative techniques, including the use of Large Language Models (LLMs) and foundation models, to enhance forecasting accuracy and efficiency. Researchers are also focusing on handling the complexities of multivariate time series data and addressing non-stationarity challenges.

One notable trend is the use of Large Language Models (LLMs) to enhance time series forecasting. The paper "Augur: Modeling Covariate Causal Associations in Time Series via Large Language Models" explores how LLMs can model causal relationships within time series data, leading to more accurate predictions. Another paper, "Empowering Time Series Forecasting with LLM-Agents," investigates the use of LLM-agents to further improve forecasting performance. These approaches leverage the ability of LLMs to understand complex patterns and dependencies, making them valuable tools for time series analysis.

Another key area of focus is the development of foundation models for time series. The paper "TiCT: A Synthetically Pre-Trained Foundation Model for Time Series Classification" introduces a foundation model specifically designed for time series classification tasks. Additionally, "TS-RAG: Retrieval-Augmented Generation based Time Series Foundation Models are Stronger Zero-Shot Forecaster" highlights the potential of retrieval-augmented generation techniques in building more robust and accurate forecasting models. These foundation models aim to provide a general-purpose framework that can be adapted to various time series applications, reducing the need for task-specific training.

Multivariate time series present unique challenges due to the complex interdependencies between different time series. "MSTN: Fast and Efficient Multivariate Time Series Model" proposes a novel model for handling multivariate data efficiently. Similarly, "RED-F: Reconstruction-Elimination based Dual-stream Contrastive Forecasting for Multivariate Time Series Anomaly Prediction" addresses the problem of anomaly detection in multivariate time series using a contrastive forecasting approach. These studies contribute to the ongoing effort to develop models that can effectively capture the dynamics of complex systems represented by multiple time series.

Addressing non-stationarity is also a critical aspect of time series forecasting. The paper "Rethinking Nonstationarity in Time Series: A Deterministic Trend Perspective" offers a fresh perspective on this issue by focusing on deterministic trends within time series data. By understanding and modeling these trends, researchers aim to improve the accuracy and reliability of long-term forecasts. Furthermore, "Domain Fusion Controllable Generalization for Cross-Domain Time Series Forecasting from Multi-Domain Integrated Distribution" explores methods for generalizing time series models across different domains, enhancing their applicability in diverse scenarios.

Traffic

Traffic management and analysis are essential for urban planning, transportation optimization, and ensuring road safety. Recent research leverages advanced techniques such as computer vision, machine learning, and simulation to address various challenges in this domain. From anomaly detection to traffic flow prediction and autonomous driving, these papers offer insights into the future of intelligent transportation systems.

Anomaly detection in traffic flow is a critical task for ensuring the safety and efficiency of transportation systems. The paper "Hybrid SIFT-SNN for Efficient Anomaly Detection of Traffic Flow-Control Infrastructure" introduces a hybrid approach combining Scale-Invariant Feature Transform (SIFT) and Spiking Neural Networks (SNN) for detecting anomalies in traffic infrastructure. This method offers an efficient way to identify unusual patterns and potential issues in real-time.

Multi-camera traffic video analysis is another area gaining significant attention. "TrafficLens: Multi-Camera Traffic Video Analysis Using LLMs" explores the use of Large Language Models (LLMs) for analyzing video feeds from multiple cameras, providing a comprehensive understanding of traffic conditions. This approach enables the extraction of valuable insights from visual data, enhancing traffic monitoring and management capabilities.

The integration of computer vision and microscopic simulation is also being investigated to improve traffic signal control. "A Simple Framework Towards Vision-based Traffic Signal Control with Microscopic Simulation" presents a framework that uses vision-based techniques in conjunction with microscopic simulation to optimize traffic signal timings. This approach has the potential to reduce congestion and improve traffic flow in urban environments.

Privacy concerns related to traffic data are addressed in "Quantifying the Privacy Implications of High-Fidelity Synthetic Network Traffic." This paper quantifies the privacy risks associated with the use of high-fidelity synthetic network traffic data, highlighting the importance of implementing appropriate privacy measures. Furthermore, "Traffic Modeling for Network Security and Privacy: Challenges Ahead" discusses the challenges and future directions in traffic modeling for network security and privacy.

Autonomous driving is heavily reliant on accurate traffic scene understanding. "Enhancing LLM-based Autonomous Driving with Modular Traffic Light and Sign Recognition" explores the use of Large Language Models (LLMs) to improve traffic light and sign recognition in autonomous driving systems. Additionally, "HABIT: Human Action Benchmark for Interactive Traffic in CARLA" introduces a benchmark for evaluating human action prediction in interactive traffic scenarios within the CARLA simulation environment.

Spatio-temporal Analysis

Spatio-temporal analysis combines spatial and temporal dimensions to understand phenomena that evolve over time and space. This field is crucial for applications like weather forecasting, climate modeling, and video analysis. The latest research covers a wide range of topics, including generative models, video understanding, and privacy in federated learning within spatio-temporal contexts.

Large Language Models (LLMs) are making inroads in spatio-temporal analysis, as demonstrated by the "Qwen3-VL Technical Report." This report highlights the capabilities of the Qwen3-VL model, which can process and understand both visual and textual information, making it suitable for various spatio-temporal tasks. The integration of LLMs in this field opens up new possibilities for analyzing complex spatio-temporal data.

Generative models are also a significant focus, with "ENMA: Tokenwise Autoregression for Generative Neural PDE Operators" introducing a novel approach for generating neural partial differential equation (PDE) operators. These models are crucial for simulating and predicting the behavior of dynamic systems in various scientific and engineering applications. Similarly, "FaithFusion: Harmonizing Reconstruction and Generation via Pixel-wise Information Gain" presents a method for harmonizing reconstruction and generation tasks, enhancing the quality of generated spatio-temporal data.

Video understanding is a key application area for spatio-temporal analysis. The paper "Thinking With Bounding Boxes: Enhancing Spatio-Temporal Video Grounding via Reinforcement Fine-Tuning" explores how bounding boxes can be used to improve video grounding, which involves locating specific events or objects within a video. Additionally, "Towards an Effective Action-Region Tracking Framework for Fine-grained Video Action Recognition" focuses on developing frameworks for recognizing fine-grained actions in videos by tracking relevant action regions. These studies contribute to the advancement of video analysis techniques for various applications, including surveillance, robotics, and human-computer interaction.

Privacy in federated learning is an increasingly important concern, particularly in the context of spatio-temporal data. "Privacy in Federated Learning with Spiking Neural Networks" addresses this issue by investigating the use of spiking neural networks (SNNs) in federated learning settings to protect sensitive information. This research highlights the need for privacy-preserving techniques in spatio-temporal data analysis.

In conclusion, the latest research papers in time series forecasting, traffic management, and spatio-temporal analysis demonstrate significant advancements in these critical fields. From the use of Large Language Models and foundation models to innovative approaches for anomaly detection and video understanding, these studies offer valuable insights for researchers and practitioners alike. For further exploration, check out arXiv for a vast repository of scientific papers.