Science & Tech · Science & Environment · 57 mins ago
Researchers develop a method for better predictions by AI world models
Researchers have proposed a new technique to help AI systems predict how objects behave under different actions.
World models are used to train physical AI, including robots, by predicting what may happen after an action.
Existing models can struggle to predict scenarios that were not observed, because they often average possible outcomes.
The new method uses the gap between a model's prediction and what actually happened to infer hidden factors, then estimate what might have happened after a different action.
In the researchers' example, a boat ends up farther downstream than predicted because of a hidden current, and that difference helps estimate where it would go if steered another way.
The method could improve reasoning and future predictions in world models, and the researchers say it can be added to existing models with limited changes.
Researchers have developed a method intended to help AI world models predict how objects would behave under actions that did not occur.
The approach uses differences between a model’s predictions and what actually happens to infer hidden factors affecting an event.
The researchers say their method can generate counterfactual predictions with a single subtraction.
The technique can be applied to world models that predict an average next state, and the researchers say it is particularly suited to JEPA models.
The work is described in a preprint titled “Counterfactual World Models: Natura Non Facit Saltus.”
- Who
- Randall Balestriero, Kerda Varaku and Yann LeCun.
- What
- They describe a method for counterfactual predictions in AI world models. It aims to help models infer outcomes for actions that were not taken.
- When
- The Register published its report on 2026-11-09. The preprint’s publication date is not stated.
- Where
- Not stated.
- Why
- World models could better guide robots if they could infer possible outcomes without being trained on every outcome.
This story does not have two clearly opposing sides.
JEPA models were developed by LeCun and colleagues.
The Register published a report on the researchers’ counterfactual world model approach.
- Researchers
- Randall Balestriero, Kerda Varaku and Yann LeCun
- Preprint
- Counterfactual World Models: Natura Non Facit Saltus
- Earlier models
- JEPA models were developed in 2023
- Encoders used
- DINOv2 and MAE
- Training described
- Off-the-shelf encoders were used without fine-tuning, with a small predictor trained on top.
What people are saying
the mismatch between your prediction and the real world to figure out what did you get wrong, and from that you can better understand what happened for better future predictions, debugging, reasoning, all at test time.
Our counterfactual world model does all three steps with one subtraction.