A new arXiv preprint, Test-Time Forecasting: Recurrent Depth Enables In-Context Dynamics Prediction, addresses a core question about foundation models for dynamical systems. According to the abstract, such models can already forecast unseen partial differential equations (PDEs), ordinary differential equations (ODEs), and time series during inference.
The abstract then states that it remains an open question whether these models can do so in-context, with the title suggesting that recurrent depth may be the key mechanism. However, the source text cuts off mid-sentence, so this digest cannot describe the proposed method, experiments, or conclusions.
Readers should treat this as a motivation-only summary. The full paper, arXiv:2610.19552, would be needed to evaluate the actual contribution and results.