In recent years, time series data has been widely utilized across various domains, and accurate time series forecasting plays a vital role in practical applications such as energy management, meteorology, and finance. Most Transformer-based methods for multivariate time series forecasting rely on point-wise attention and Channel-Dependent modeling, leading to high computational costs and over-smoothed representations due to uncontrolled information mixing. Moreover, these methods struggle with high-frequency signals and fail to capture inherent periodic patterns, which are essential for accurate forecasting. To address these issues, we propose an Adaptive Cross-Variable Spectral Filtering (ACSF) framework that extracts variable-specific features in the frequency domain using a learnable adaptive filter. ACSF introduces a cross-variable attention mechanism for updating filter weights, enabling effective fusion of frequency information across variables and mitigating the limitations of both Channel-Dependent and Channel-Independent approaches. Experiments on multiple benchmarks demonstrate that ACSF achieves superior accuracy with lower computational complexity, offering a novel solution for multivariate long-term forecasting.