In 1905, sitting at his desk in the Swiss Patent Office, Albert Einstein had not found a contradiction in existing physical data. Newtonian mechanics was still enjoying tremendous success, and no signal from experiments was saying "we need a new framework." What Einstein used was not data, but imagination—picturing a person in free fall inside a sealed elevator. That simulation of bodily sensation gave birth to a new axiom, the equivalence principle, which connects gravity and acceleration.
Today's large language models, if shown the equations Einstein derived, can brilliantly carry out the subsequent mathematical deduction. But they cannot generate that axiom on their own. In 2026, Tom Zahavy, a research scientist at Google DeepMind, attempted to explain this asymmetry from the structure of reasoning itself in a position paper titled "LLMs Can't Jump," submitted to ICML (International Conference on Machine Learning).
The Map of Science Einstein Drew in 1952
Before diving into the heart of the matter, we need to confirm how Einstein himself understood scientific discovery.
On March 30, 1952, Einstein drew a small diagram in a letter to his friend, the philosopher Maurice Solovine. It depicted the path along which science proceeds. The starting point is E (Experience)—raw sensory data. From there, the scientist reaches A (Axioms) through an intuitive leap, deduces verifiable consequences from the axioms through deduction, and checks them against experience E once again. It is a circulating loop, and only the "E→A" segment lies outside logic. Einstein labeled this not as a logical step but as "J (Jump)."
The philosopher Charles Sanders Peirce (1839-1914) named this leap "abduction," stating that "abduction is the only logical operation which introduces any new idea." While induction reads trends from observation and deduction derives results from existing rules, abduction makes the leap of "if this hypothesis were true, it would explain that observation" when observational data is scarce. It is not proof, but a bet on the best explanation.
Zahavy's paper overlays this tripartite diagram onto a mapping of LLM capabilities, attempting to show where current AI stands on the map of science.
Notably, Einstein himself strongly distrusted induction. He explicitly stated that "there is no logical path from observational data to physical laws," and believed that axioms were "free inventions of the human mind." This epistemological stance serves as the starting point for Zahavy's paper.
Two Domains AI Has Conquered, and One It Cannot Reach
There is little dispute that current AI excels at induction. Statistical pattern matching across billions of tokens is essentially the industrialization of induction.
Deduction, too, has advanced rapidly in mastery over the past few years. AlphaProof solved 4 out of 6 problems at the 2024 International Mathematical Olympiad (IMO), achieving a score of 28/42—equivalent to a silver medal. This marked the first time AI reached medal-level performance at the IMO. The following year, in 2025, Gemini (equipped with Deep Think) solved 5 problems, reaching 35/42, a gold-medal-level score. The proposition that "given axioms, LLMs can reliably derive logical consequences from them" is now barely in question.
However, Zahavy argues that the situation is fundamentally different for the third mode of reasoning: abduction.
The issue is not the quantity of knowledge. When Einstein formulated general relativity, observational data was not commanding him to "create new physics." Classical mechanics was functioning perfectly well experimentally. The breakthrough came from outside the data. A thought experiment rooted in bodily intuition—a person falling inside a sealed elevator—gave birth to the axiom that "gravity and acceleration are essentially the same thing." The order was to gather observations and develop deductions toward this axiom afterward.
LLMs know only text. They have no sensation of the body floating inside a sealed elevator. As a result, there is no route to construct causal models from physical intuition. Even by expanding data centers or pushing computational scale into the stratosphere, this constraint cannot be resolved through parameter count.
Below is a comparison table showing the three modes of reasoning and current AI capabilities at a glance.
| Reasoning Mode | Definition | Current AI Status | Representative Example |
|---|---|---|---|
| Induction | Reading probabilistic patterns from data | Mastered | GPT's linguistic interpolation, image recognition |
| Deduction | Deriving logical consequences from axioms | Highly mastered | AlphaProof (IMO2024 Silver), Gemini (IMO2025 Gold) |
| Abduction | Generating new hypotheses under data scarcity | Structural constraints (per Zahavy's paper) | Einstein's discovery of the equivalence principle |
The Counterargument: "Creativity Is Data Compression"
There is a counterargument that Zahavy confronts head-on. It is a view proposed by AI researcher Jürgen Schmidhuber and widely accepted in some circles: "Scientific discovery is nothing more than advanced data compression. If you give an LLM enough text, scientific breakthroughs will naturally emerge as a result of compression."
Zahavy attempts to refute this using Einstein's example. During the period when general relativity was being formed, observational data was not demanding a new framework for physics. There was no "mass of contradictory data" that needed to be compressed. The equivalence principle was not the result of compressing data, but an axiom that was conceptualized beforehand through a bodily thought experiment. From there began the loop of deduction and observational verification.
"Even if an algorithm compresses data, it cannot generate axioms in domains where that data does not exist"—this is the paper's core argument.
However, Zahavy himself added this careful caveat in a post on X: "This is not DeepMind's official position but my own personal position paper. There is a high likelihood that merely improving current systems could lead to many discoveries in the near future, and I do not rule out the possibility that scaling up could lead to new inventions in physics as well." He also stated, "As a key contributor to AlphaProof, I know firsthand that remarkable discoveries are being made using LLMs." It is a peer-reviewed but ongoing paper, and needs to be read as one perspective.
The debate ignited by the paper spread more widely than Zahavy had initially anticipated. On X, voices of both support and caution rose from the AI research community, with some emphasizing a pragmatic line that "steadily improving current systems can lead to many practical discoveries." A representative example of this is the observation that even if Einstein could not have generated a new axiom, problems within the scope of Newtonian mechanics could still have been solved. This resonates with Kuhn's philosophy of science, which divides scientific discovery into "paradigm shifts" and "puzzle-solving," and it cannot be denied that LLMs are pushing the efficiency of puzzle-solving to the limit.
Embodiment: The Missing Variable
So how might we implement this "leap" in machines? The direction Zahavy proposes in the paper is a physically consistent, multimodal world model.
This involves building systems that go beyond processing text and images to execute embodied simulations within virtual environments that obey physical laws. The idea is that if AI could actually simulate the experience of "what happens when you fall inside a sealed elevator," the intuition necessary for the leap from sensory experience to axioms could be formed.
How much support this direction currently holds on the research frontier remains an open question. In replies on X, some pointed out that "the current generation of LLMs can already handle multimodal input (images and video)." Whether the gap between physical simulation and video viewing is sufficient to produce abductive capability remains experimentally unverified.
Questions also remain about the logical structure of the argument itself. From the two premises that "Einstein's generation of the equivalence principle was abduction" and "LLMs cannot perform abduction," the paper draws the conclusion that "LLMs face an essential constraint in scientific invention"—but a critical reading could point out that discussion of other forms of scientific invention (such as those born inductively from observation) is thin.
On the other hand, what made Zahavy's paper a genuinely valuable contribution for many readers was that it transformed the question of "why can't current AI make scientific inventions" from a vague intuition into a structure that can be engineering-verified. By objectifying "what abduction is, and why current AI cannot do it," it has allowed those designing the next architecture to narrow down their targets.
One concrete approach toward this goal is fusion with "embodied learning" in robotics research. Currently, reinforcement learning models that learn by interacting with physical laws in simulated environments have developed in a context different from language models. If an agent could autonomously experience falls in enclosed spaces or collisions between objects, and extract parameters of physical law from that sensory data, this could be called a primordial form of abduction.
However, it remains uncertain whether trial and error in physical simulation connects seamlessly to the kind of highly abstract leap that Einstein made. While video generation models trained on vast amounts of video data appear to "understand" physical laws, there is persistent criticism that these are still merely statistical pattern interpolation.
If the essence of scientific discovery lies not in compressing data but in leaping outside of it, then AI's evolution must venture into uncharted territory. Mathematical deductive ability rose sharply from an estimated 28 points to a perfect score on the IMO scale between 2024 and 2026. No benchmark for measuring abductive ability yet exists. Designing a benchmark to answer the question "Can LLMs jump?" itself seems likely to become the next research challenge.