Falcon-X: New AI Model for Heterogeneous Time Series Forecasting

A new time series foundation model called Falcon-X has been proposed to address limitations in existing models. Most current time series foundation models are univariate, and recent attempts at multivariate modeling still operate in raw variate space, which hampers semantic alignment and relational expressivity. Falcon-X decouples variates from raw space into a unified latent prototype space. It uses a Unified Prototype Diff-Attention mechanism to evaluate both positive and negative semantic affinities, enabling explicit alignment of heterogeneous physical quantities. This approach aims to capture complex synergistic and antagonistic interactions in real-world systems. The model is described in a paper on arXiv (2605.27286).