---
id: liquid-machine-learning
name: Liquid Machine Learning
family: technological
tier: emerging
url: https://futuremap.talin.digital/e/liquid-machine-learning/
license: CC-BY-NC-4.0
---

# Liquid Machine Learning

## In short
Most machine-learning models are trained, frozen and deployed; they do not change again until somebody retrains them. Liquid neural networks alter their own governing equations as new data arrives, so the model keeps adapting after deployment.

The attraction is small models that behave well in changing conditions — the demonstrations are mostly in control and robotics, where a system has to cope with a situation its training set did not contain, and where the model has to fit on the machine rather than in a datacentre.

It sits at 4 on technical readiness with societal readiness slightly ahead, which is a mild version of the hype-risk condition. The appetite for models that keep learning is running a little ahead of the evidence that they can be relied on — and a system that changes after certification raises a question every safety regulator will ask and none has yet answered.

**Common misconception:** that continuous adaptation is straightforwardly better. A model that keeps learning is a model whose behaviour last week does not guarantee its behaviour today.

## Readiness
TRL 4 · SRL 5 · gap -1

Source: WKO Innovation Map 26/27 (2026)

## Candidate megatrends (UNREVIEWED)
- Automation of Cognitive Work (automation-of-cognitive-work)
- Digital Sovereignty (digital-sovereignty)
- Synthetic Media and the Verification Collapse (synthetic-media-and-the-verification-collapse)
- The Application Gap (the-application-gap)

These are derived from this technology's cluster, not from a judgement about this technology. They are a starting point for review, not a claim.
