{
  "id": 19216314,
  "title": "Insight about how 'nature makes no leaps' promises better robot predictions",
  "url": "https://urgent.news/2026/11/09/insight-about-how-nature-makes-no-leaps-promises-better-robot-19216314",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-09T17:00:00.000Z",
  "source": {
    "name": "The Register",
    "slug": "the-register",
    "url": "https://www.theregister.com/ai-and-ml/2026/11/09/insight-about-how-nature-makes-no-leaps-promises-better-robot-predictions/5308301"
  },
  "original_language": "en",
  "account": "Recent advancements in machine learning have led to techniques that allow world models to make more accurate predictions about how objects will behave in the real world. World models, which are used to train physical AI systems like robots, predict the consequences of actions based on observed data. However, these models do so by basing their predictions on the outcomes that have already been observed, rather than inferring possible actions without prior training.\n\nResearchers from Brown University, AMI Labs, NYU, AMI Labs, and Turing Award winner Yann LeCun have developed a technique that could allow world models to make inferences about possible actions without being trained for every potential outcome. This approach is described in a pre-print paper titled \"Counterfactual World Models: Natura Non Facit Saltus,\" which translates to \"Nature makes no leaps.\"\n\nThe technique, known as Counterfactual World Models, uses the mismatch between a prediction and the real world to determine where the prediction went wrong. By understanding the source of the discrepancy, the model can make better predictions, aid in debugging, and enable reasoning and future predictions. These models can be applied to any world model that predicts the average next state under any action, but they are particularly well-suited to Joint Embedding Predictive Architecture (JEPA) models.\n\nJEPA models, developed by Yann LeCun and colleagues in 2023, take an action as input and ask what the world would be like with that action. However, JEPA models produce an average of outcomes and cannot anticipate counterfactual scenarios, which are events that did not occur. The researchers' technique can be applied to any world model, but JEPA models are particularly well-suited to being adapted.\n\nThe counterfactual world model predicts alternative scenarios by inferring possibilities from what actually happened. This method involves three steps: learning something about hidden factors from observed data, changing the action, and predicting again with those hidden factors kept constant. The counterfactual world model performs all three steps with a single subtraction.\n\nAn example from the paper involves a boat crossing a river with a hidden current. Both the JEPA world model and the counterfactual world model predicted where the boat would end up after steering left, but the counterfactual model also predicted the boat's position if the operator had steered right. This difference between the actual outcome and the counterfactual prediction is attributed to the hidden current.",
  "summary": "Leibniz lives!",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}