ChronoMixer: Decoupled Local-Global Dynamics Learning via Adaptive Information Gating

Effective modeling of complex temporal dependencies is critical in time series analysis, particularly for simultaneously characterizing local variations and global structures in long sequences. Existing forecasting models often rely on a dominant temporal inductive bias: convolutional or MLP-style operators are effective for local variations but limited in long-range dependence modeling, whereas attention-based models capture distant interactions yet may underrepresent rapidly changing local details. This imbalance makes it difficult for a single static modeling bias to adapt robustly to complex nonstationary sequences. In this article, we propose ChronoMixer, an efficient parallel framework that explicitly decouples short- and long-term temporal modeling within a unified network. ChronoMixer performs parallel local-global modeling through a lightweight local branch and a global branch, and coordinates their contributions with a content-aware scalar gate for adaptive temporal bias allocation. This design enhances stability and generalization across varying horizons and distributions while significantly reducing parameter count and complexity. On multiple long-term forecasting benchmarks, ChronoMixer outperforms or matches representative methods in accuracy, scale, and inference efficiency. Furthermore, in a real-world stellar spectral parameter estimation task, ChronoMixer achieves competitive results without domain-specific engineering, validating its cross-task potential. Experimental results demonstrate that explicit coordination of local and global temporal inductive biases offers an effective and efficient solution for robust time series forecasting. The source code will be made publicly available upon acceptance.