NeurIPS paper: 1-parameter interpretable architecture beats time series foundation models zero-shot

DangerousFunny1371 · reddit · 2026-10-04

A NeurIPS paper introduces DynaBase, a minimal interpretable architecture for zero-shot reconstruction of dynamical systems, built from just two mechanisms:

A piecewise affine map with a single parameter α controlling local con-/divergence rates;
A context selector picking the data point closest to the current state, keeping generated dynamics faithful in temporal and geometric properties.

With only these, DynaBase reproduces all major dynamical regimes — fixed points (α<1), limit cycles (α=1), chaotic attractors (α>1) — and, unlike context parroting, preserves the correct regime. In zero-shot mode it outperforms most major time series and dynamical-systems foundation models on long-term statistics and even short-term prediction. Training and inference are extremely cheap: solvable analytically via linear regression or a 1-parameter grid search. The authors argue its formal simplicity offers a tractable mathematical handle for analyzing time series foundation models.

Original post →

NeurIPS paper: 1-parameter interpretable architecture beats time series foundation models zero-shot

More from Research