← sources

A Minimal Interpretable Architecture for Zero-Shot Reconstruction of Dynamical Systems [R]

source: reddit (reddit:MachineLearning)published 2026-10-04 12:49ingested 2026-10-05 02:38

retained text

In our #NeurIPS2026 paper “A Minimal Interpretable Architecture for Zero-Shot Reconstruction of Dynamical Systems (DS)” (preprint: [https://arxiv.org/abs/2607.14937](https://arxiv.org/abs/2607.14937)) we reduce a **DS foundation model** to the ingredients **minimally necessary to faithfully reproduce long-term statistical and geometrical properties of DS**:

1) A **piecewise affine map** with only a **single (!!) parameter** α that controls local con-/divergence rates, and …

2) … a **context selector** that chooses from the provided context signal **the data point closest to the current state** of the map, thus ensuring the generated dynamics stays close to the context in its temporal and geometrical properties.

With just these two mechanisms, this minimal form – which we coined *DynaBase* – **can reproduce all major dynamical regimes**, including fixed points (α<1), limit cycles (α=1), and chaotic attractors (α>1). Thus, unlike other simple mechanisms like context parroting, DynaBase even **preserves the correct dynamical regime**!

Surprisingly, it turns out that this simple context-driven 1-parameter map **outperforms most major time series and DS foundation models**, as well as custom-trained models, in both long-term statistics and even short-term predictions, even when run in **zero-shot mode**.

**Both inference and training are extremely cheap** – training can be done either analytically in one step by linear regression on forward-predictions, or by 1-parameter grid search directly on DS reconstruction objectives → this reveals interesting performance differences induced by different training mechanisms.

Most importantly in our minds, DynaBase owing to its formal simplicity may thus provide a tractable mathematical handle on analyzing, improving & understanding the performance and training of some time series and DS foundation models.

https://preview.redd.it/jmlcnocq5gth1.png?width=3000&format=png&auto=webp&s=1fc322e17a77575ddb0be3525482b45c70fa28f1

document fetched 2026-10-05 00:26 · canonical https://reddit.com/r/MachineLearning/comments/1wxex8n/a_minimal_interpretable_architecture_for_zeroshot