World Models · Sep 26 · 3 min read
Ergodic Wins Innovate UK Backing for Causal World Models
Dr. Andre Franca will spend six months on causal world models that can learn a task in one environment and perform it reliably in another.
We're delighted to share that Ergodic has been awarded funding through Innovate UK's AI Champions: Frontier AI competition. Over the next six months, Dr. Andre Franca will lead a research project on one of the hardest open problems in applied AI.
The problem is transfer. An AI system learns a task in one place, gets deployed somewhere else, and often stops performing reliably.
Why world models don't travel
World models are AI systems that learn an internal, predictive representation of their environment. With that representation, they can anticipate what happens next and plan against it. They're fast becoming the substrate for how robots and industrial systems make decisions.
But every deployed world model today is trained on observational data from a single environment. Its learned representation gets entangled with site-specific and operator-specific confounders. It picks up the habits of one site along with the causal structure of the task itself.
Picture a model trained on a single production line. It learns that output drops when a particular shift is on, when the real cause is the older machine that shift happens to run. Move the model to a site with different staffing and different equipment, and that shortcut sends its predictions the wrong way.
Our view is that this is a structural issue and a consequence of the training objective. No amount of extra data or extra parameters closes it. A bigger model trained on the same kind of data learns the same confounders, only more confidently.
Going after the objective
So we're going after the objective itself. Andre's work focuses on causal world models, which learn the structure of a task separately from the quirks of the place they were trained.
If a model knows which relationships are causal and which are local accidents, it has something it can carry to a new environment. That's what separates a model that has memorised one site from a model that understands the task.
This work is central to the longer arc of what Ergodic is building towards. We want world models that predict what comes next and answer what would happen under intervention. And we want them to carry that understanding into environments they've never seen, across the enterprises that use them.
More to share as the work progresses.