DeepMind Paper: LLMs Lack the "Jump" for Scientific Discovery, Need Multimodal World Models

maier_ak · x · 2026-08-07

This article provides an in-depth analysis of the 2026 Google DeepMind position paper titled "LLMs can't jump." It highlights that while LLMs exhibit staggering abilities in induction (pattern recognition) and deduction (theorem proving), they face a structural barrier that prevents the creative leap necessary for true scientific breakthroughs.

The paper conceptualizes the process of discovery into three stages, arguing that current language-based architectures hit a substrate-based limit. To achieve genuine scientific leaps, AI must combine language mastery with consistent multimodal world models that run counterfactual experiments, effectively grounding symbols in sensation.

Related event: DeepMind Paper: LLMs Lack Intuitive Leap for Scientific Breakthroughs(4 posts)→

Original post →

More from AGI Musings

AGI Musings channel →