Neural ODEs are great as continuous-time dynamical systems, but they are slow and tedious to train.

In a new #ICML2026 paper we introduce a novel solver for continuous-time RNNs that does not rely on numerical integration, but "jumps" between piecewise closed-form solutions. It's not only way faster & more robust, but enables explicit analysis of important topological properties like fixed points and limit cycles: https://arxiv.org/abs/2602.15649

In our #ICML2026 position paper we argue a *dynamical systems perspective* is needed to drive time series (TS) modeling forward (preprint, to be revised in line with great referee suggestions: https://arxiv.org/abs/2602.16864).

Dynamical systems reconstruction (DSR) goes beyond mere forecasting and gives us an *understanding* of the dynamical rules that underlie observed time series. This in turn may enable true out-of-domain generalization and predicting a system’s *long-term* behavior, something current TS models cannot do. In the paper, we compare a variety of custom-trained and recent foundation models for TS and DSR w.r.t. short- & long-term forecasting.

Specifically, we suggest:

1) Put a focus on DSR-specific training techniques and objectives in TS model training. Proper training is more important than model architecture!

2) Pretrain TS models on *simulations from dynamical systems*, rather than on artificially created time series functions. These will yield much more natural priors for real-world TS.

3) *Move away from transformers, back to modern RNNs*. DS are defined by recursions in time. By ignoring this, transformers lose essential dynamical information.

4) *Address the hard problems in TS modeling: Topological shifts*. Although in itself tricky, the hard problem in TS forecasting is not so much mere out-of-distribution shifts, but changes that drive a system across tipping points or into different dynamical regimes.

5) *DS properties like attractors or bifurcations are universal* – acknowledging this in TS modeling will give a kind of mechanistic and transferable understanding of TS properties.

Announcing the CATS workshop #ICML2026!

Continual AdapTation at Scale:
Towards Sustainable AI
https://cats-icml.github.io

Let's enable fast, continual adaptation to drive more sustainable AI!

🗓️ Deadline: April 30, 2026

🔍 Key Topics
- Scale & Efficiency
- Lifecycle & Alignment
- Multimodality
- Deliberate Forgetting

"Kill your darlings, cut unless serving the story"
One tip from the guide.

If you needed a moment out of #ICML2026 writing and graphs, why not read some writing and figure making tips?
https://docs.google.com/document/d/14Wax8M5w8F_8miDlYJ9-I6wqpelxlXjCEUbkNzNMqqE/edit?tab=t.0#heading=h.vhbgs2r4i9p9
#AI