I recently came across an interesting position paper titled “LLMs Can’t Jump,” by Google DeepMind researcher Tom Zahavy. It asks a difficult question: can today’s AI systems genuinely invent a new scientific theory, or are they mainly good at extending ideas that already exist?
The author distinguishes three kinds of reasoning:
- Induction: learning patterns from examples.
- Deduction: deriving conclusions from established rules.
- Abduction: proposing a new explanation when existing concepts are not enough.
His argument is that AI has made impressive progress in induction and deduction, but still struggles with abduction. He calls this missing act the “Jump.”
From experience to ideas
Einstein once described scientific discovery as a movement from sense experience to a system of axioms:
Experience → (Jump) → Axioms
Once the axioms exist, logical reasoning can derive their consequences. The harder part is deciding what those axioms should be in the first place.
This is the paper’s central distinction: proving results within a theory is not the same as inventing the theory.
Why General Relativity matters
The author uses Einstein’s development of General Relativity as his main example.
At the time, Newtonian gravity was still extremely successful. Mercury’s unusual orbit was one of the few known problems, and many scientists believed it could be explained by an undiscovered planet called Vulcan. There was no large collection of failed observations clearly demanding a new theory.
A system focused only on reducing prediction error might have preferred a small correction to Newton’s model. Einstein instead reconsidered gravity, space and time themselves.
This is why the paper questions the idea that scientific creativity is simply a form of data compression. General Relativity was elegant in its final form, but the path toward it was uncertain, complicated and full of mistakes.
Einstein’s “happiest thought”
The key example is Einstein’s elevator thought experiment.
Imagine a person inside a sealed elevator accelerating upward in empty space. An object released inside appears to fall toward the floor. To the observer, this experience could be indistinguishable from gravity.
Einstein’s leap was to suggest that, locally, acceleration and gravity are physically equivalent.
This idea did not come from thousands of examples, nor was it deduced from an existing theory. It came from constructing an imagined experience and asking what it meant.
The paper describes this as manipulative abduction: creating a scenario, changing its conditions and using the result to form a new explanatory principle.
Why language may not be enough
A language model can explain Einstein’s elevator experiment because the concept already exists in human writing.
The harder question is whether it could have invented the equivalence principle before anyone had described it.
The author is doubtful. He argues that LLMs manipulate symbols without directly experiencing the physical realities those symbols refer to. A model may understand how words such as gravity, falling and acceleration are related, but it has never felt weight or free fall.
The paper therefore proposes interactive world models: systems that can simulate environments, intervene in them and run counterfactual experiments. Such a system would not merely watch an elevator. It could accelerate it, cut the cable, release objects and compare what different observers experience.
I am not fully convinced that physical grounding alone would be enough.
Still, the paper makes an important point.
An AI may become excellent at proving, predicting and optimizing without learning how to create the foundational ideas that make those activities possible.
The real challenge may not be building a machine that follows an existing path faster.
It may be building one that recognizes when the path itself needs to change.
Reference: Tom Zahavy, “Position: LLMs Can’t Jump,” Google DeepMind, presented at ICML 2026. Paper (OpenReview) · Author’s project page