Title: PATRA: Pattern-Aware Alignment and Balanced Reasoning for Time Series Question Answering

Original: arXiv:2602.23161v4 Announce Type: replace Abstract: Time series reasoning demands both the perception of complex dynamics and logical depth. However, existing LLM-based approaches exhibit two limitations: they often treat time series merely as text or images, failing to capture the patterns like trends and seasonalities needed to answer specific questions; and when trained on a mix of simple and complex tasks, simpler objectives often dominate the learning process, hindering the development of deep reasoning capabilities. To address these limitations, we propose the Pattern-Aware Alignment and Balanced Reasoning model (PATRA), introducing a pattern-aware mechanism that extracts trend and seasonality patterns from time series to achieve deep alignment. Furthermore, we design a task-aware balanced reward to harmonize learning across tasks of varying difficulty, incentivizing the generation of coherent Chains of Thought. Extensive experiments show that PATRA outperforms strong baselines across diverse Time Series Question Answering (TSQA) tasks, demonstrating superior cross-modal understanding and reasoning capability.

Rewrite: Effective reasoning over time series requires a dual capacity for interpreting intricate dynamics and engaging in profound logical analysis. Current large language model (LLM) methodologies, however, are constrained by two primary shortcomings. First, they frequently reduce time series data to static text or visual formats, thereby overlooking critical structural elements such as trends and seasonal fluctuations that are essential for addressing specific inquiries. Second, during training on heterogeneous datasets containing both simple and complex problems, the learning process tends to be skewed by easier objectives, which stifles the cultivation of advanced reasoning skills. To overcome these challenges, we introduce the Pattern-Aware Alignment and Balanced Reasoning (PATRA) framework. This approach incorporates a specialized mechanism designed to isolate trend and seasonality patterns from time series data, facilitating a more profound alignment. Additionally, we have developed a task-specific balanced reward system that stabilizes learning across tasks of different complexities, thereby encouraging the formation of logical Chains of Thought. Comprehensive experimental results indicate that PATRA surpasses robust baseline models in various Time Series Question Answering (TSQA) scenarios, highlighting its enhanced ability to perform cross-modal comprehension and sophisticated reasoning.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC