Can LLMs Propose New Theories Like Einstein? The Scientific Creativity Debate in “LLMs can’t jump”

In his personal position paper “LLMs can’t jump”, Google DeepMind researcher Tom Zahavy asks: if an AI were given only the knowledge Einstein possessed at the time, could it propose the new physical principles required for general relativity?

This question is far more specific than whether AI can make scientific discoveries in general. Zahavy contributed to AlphaProof and acknowledges that systems built on large language models are already generating genuine discoveries. The paper argues that current LLMs lack the mechanism to formulate new theoretical premises from physical experience. This leaves a concrete question: in the absence of a large body of anomalous data pointing the way, how can a system take this step?

A model can compute predictions based on an established theory, or search for empirical patterns in data. In proposing the equivalence principle, however, Einstein did something more: he envisioned what an observer would experience in free fall, and elevated that thought experiment into a foundational principle about the physical world. The paper terms this conceptual step a “jump.”

Three Modes of Reasoning in Scientific Discovery and a Common Misconception

Deriving results from existing rules, extracting rules from examples, and proposing explanations for phenomena constitute three distinct tasks. C. S. Peirce’s distinction among modes of inference can be illustrated using the motion of falling bodies:

1. Deduction: Rules and initial conditions → Predictions. Given the laws of gravitation and an object’s initial state, calculate how it will fall. Only when the premises are true and the derivation is valid can the conclusion be reliable. Formal theorem provers can verify whether a derivation adheres to the rules, but whether the physical premises faithfully describe reality still requires observational testing.
2. Induction: Observed cases → General regularities. Comparing observational records of various falling bodies reveals that, neglecting air resistance, they fall with identical acceleration at the same location. The number of observed cases may be large or small; how far the summarized rule generalizes requires continuous verification.
3. Abduction: Background knowledge and phenomena → Candidate explanations. Observing the consistent fall of objects alongside the local weightlessness of a freely falling observer prompts the question: do these phenomena reflect a single, unified physical principle? Proposing an explanation does not guarantee its correctness; it merely offers a new starting point for subsequent mathematical derivation and empirical testing.

When asked what a person feels inside a falling elevator, an LLM might provide an accurate answer. However, training corpora may well already contain the elevator thought experiment, meaning a correct response does not demonstrate that the model could independently invent the equivalence principle. By the same token, one cannot conclude purely from a correct answer that no internal reasoning occurred. To evaluate independent discovery, one must rigorously control the knowledge and pre-existing answers accessible to the system.

The Theoretical Dilemma of “Data Compression as Creativity”

One prominent perspective, championed by computer scientist Jürgen Schmidhuber and others, conceptualizes scientific discovery as compression: explaining a broader body of observations with simpler rules. For instance, rather than recording trajectories point-by-point, it is preferable to discover a compact set of laws capable of generating those trajectories.

Zahavy challenges this account by pointing out that when general relativity was formulated, there was no massive influx of anomalous data explicitly calling for curved spacetime. Newtonian gravity accounted for an overwhelming majority of observations, despite an unexplained anomaly in the orbit of Mercury. Some researchers attempted to account for this anomaly by hypothesizing an undiscovered planet named “Vulcan” near the Sun; a single anomaly does not uniquely point toward one specific new theory.

Conceptual tensions also existed: how could Newtonian gravity’s action-at-a-distance be reconciled with the emerging pictures of special relativity and field theory? Zahavy concedes that AI might well detect such internal inconsistencies. Scarcity of data, therefore, should not be equated with an absence of problems, lack of feedback, or an algorithm devoid of direction.

What the paper fundamentally questions is this: if the objective is primarily to reduce prediction error on existing data, what would compel a system to abandon a well-performing classical framework and embrace an entirely novel conception of spacetime? This challenge targets research objectives; the paper does not train a historical model, measure an empirical loss near zero, and observe it failing to break through.

A simpler final theory does not imply that every intermediate step of finding it was simpler. Unfamiliar mathematical tools, erroneous field equations, and abandoned hypotheses can initially expand the candidate space and increase complexity. Whether rewarding short-term empirical fit or compression improvements alone can support this type of exploration is precisely the question raised by the author.

Empirical verification also encompasses both old and new data: Mercury’s orbital anomaly was an existing observation that later served as an important test; the gravitational deflection of starlight during a solar eclipse was subjected to fresh observational tests only after the theory had been formulated. Scientists can first formulate an explanatory hypothesis, and only then seek out discriminating evidence capable of distinguishing the new theory from the old.

Zahavy’s Argument: Bodily Simulation, Embodied Jumps, and Cognitive Blind Spots

In a letter to Maurice Solovine, Einstein depicted scientific creation as a recurring cycle: Experience → Formulation of theoretical premises → Derivation of testable consequences → Experimental testing → New experience. The paper focuses squarely on the leap from experience to new premises. The cycle diagram in the paper is an AI-redrawn illustration, and the elevator graphics were likewise generated by AI; they explain conceptual relationships rather than reporting model capability benchmarks.

Theoretical premises here refer to the fundamental assumptions or physical requirements tentatively accepted for subsequent mathematical deduction. The paper identifies seven core requirements across Einstein’s investigation, including local equivalence, energy and momentum acting as gravitational sources, and the recovery of Newtonian results under weak fields and low velocities. Zahavy judges that modern AI could potentially carry out equation searching, formal derivation, and error checking if provided with these premises; the paper does not conduct an empirical rediscovery test under historical knowledge constraints.

How, then, does the proposal of premises occur? The elevator thought experiment sets up two paired comparisons. Within a sufficiently small region and short time interval, neglecting tidal effects:

• A freely falling cabin near Earth and an inertial cabin drifting in deep space both induce weightlessness. In the former, the observer, the cabin walls, and a released ball fall together; the ball appears to hover relative to the observer. Gravity has not vanished; rather, local motion within the cabin mimics a gravity-free environment.
• A stationary cabin on the ground and a cabin accelerating upward in deep space both produce a sensation of weight. In the accelerating cabin, the floor continuously catches up with the released ball; choosing an acceleration of approximately 9.8 m/s² roughly mirrors the terrestrial falling scenario.

Einstein generalized this correspondence far beyond falling balls: within a sufficiently small, freely falling laboratory, neglecting tidal effects, local non-gravitational physical experiments obey the same laws as those in a gravity-free inertial laboratory. This claim no longer merely explains why a person feels weightless; it demands that optical signals, clocks, and other physical processes be consistently described under local equivalence. Zahavy terms this process of generating explanations via manipulable mental simulations “manipulative abduction”: altering conditions, comparing consequences, and subsequently advancing a more general hypothesis.

Replacing the ball with a beam of light shows how a new premise generates fresh questions. Imagine a deep-space rocket, initially at rest, that accelerates upward; a light beam emitted horizontally from the left wall travels in a straight line from the perspective of an external inertial observer. During the transit of the light, the cabin shifts upward, causing an internal observer to see the light path bend toward the floor. The local equivalence between acceleration and gravity prompts a deeper question: how should light propagation be influenced within a gravitational field? The Albert Einstein Institute’s explainer on the deflection of light outlines the scope and limits of this reasoning: local equivalence provides predictive clues, but the full bending angle around the Sun requires the global geometry of spacetime.

Clocks likewise yield testable consequences. In a static gravitational field near Earth, identical clocks placed at different altitudes can be compared by having the lower clock transmit a light pulse to the higher clock on every tick. The higher observer, comparing the incoming signals against a local clock, finds the arrival interval longer than local ticks, concluding that the lower clock runs slower. Continuous light climbing upward correspondingly shifts to lower frequencies, a phenomenon known as gravitational redshift. Every observer finds local clocks running normally; discrepancies appear through comparisons between different locations. Such consequences transform an intuitive principle into a calculable, measurable claim.

The author suggests that operating purely on physical text may lack the experiential foundation needed to cultivate such intuitions. Yet this remains a mechanistic argument; one cannot infer from a visually realistic video of falling objects whether a generative model internally maintains reliable physical representations. The paper acknowledges that few-shot grid benchmarks like ARC-AGI already evaluate explanatory inference for uncovering hidden transformation rules; it questions whether this spans the active formulation of physical situations, the acquisition of synthetic experience, and the formation of new physical principles.

Consequently, the paper advocates exploring world models that can be actively intervened upon while preserving physical consistency. A system must be capable of altering conditions—such as cutting an elevator’s suspension cables—and comparing downstream consequences, subsequently abstracting relationships across varied scenarios into testable propositions for formal derivation and experimental scrutiny. While physical priors help constrain the search space, simulators can learn incorrect dynamics, and the capacity to manipulate an environment does not automatically translate into the ability to invent theoretical frameworks. Controlled environments like Genie inspire this research direction, but their existence does not demonstrate that this challenge has been resolved.

There is also the active agency of deciding what to investigate: human researchers generate variations of problems themselves, deciding which consequences warrant investigation. Citing test-time reinforcement learning as an adjacent capability, the paper acknowledges that systems can generate and solve new chess positions or mathematical problems; it presses on whether they can move beyond variations within fixed rule systems to propose new physical premises. The author situates this world-model trajectory primarily within physical science; exploration in mathematics and computer science operates over abstract structures and formal systems without needing to replicate embodied physical sensation.

Intuition Is Not Theory: The Strict Physical Boundaries of the Equivalence Principle

The elevator thought experiment established a productive principle, but that principle alone cannot uniquely determine a complete gravitational theory. The principle itself is subject to explicit scale limitations.

1. Why Confine to Local Regions and Ignore Tidal Effects?

Real gravitational fields near Earth are non-uniform. Consider two balls separated horizontally inside the same freely falling cabin, initially at rest relative to the walls. Because both balls fall toward the center of the Earth, they gradually converge laterally in addition to their common downward plunge. An observer inside the cabin, having cancelled out the shared downward motion, can still detect this subtle relative movement.

This phenomenon is a manifestation of tidal effects: differences in gravitational pull across distinct spatial locations. In contrast, inside an ideal deep-space inertial cabin, neglecting external gravity and mutual gravitational attraction, two initially stationary balls maintain a constant separation.

The smaller the spatial extent and the shorter the duration of observation, the more negligible these differences become. The Albert Einstein Institute’s explainer on the equivalence principle clarifies that an “infinitesimal spacetime region” represents a mathematical limit; physical experiments need only choose regions and durations sufficiently small relative to the required precision. Expanding the domain or extending the duration exposes tidal effects, demonstrating that local approximations cannot be casually generalized to arbitrarily large scales.

2. The Long Trek from Intuitive Axioms to Field Equations

Why introduce geometry at this juncture? Reconsider the two falling balls: falling along with the cabin cancels their common downward acceleration, but it cannot simultaneously cancel the relative motion between different freely falling bodies across an extended region. General relativity organizes these kinematic relations through spacetime geometry: objects free from non-gravitational forces travel along natural trajectories in curved spacetime known as geodesics; neighboring geodesics can naturally converge or diverge. Tidal effects represent the physical expression of spacetime curvature in relative motion. While this provides intuitive motivation for a geometric framework, an elevator scenario does not uniquely determine the complete field equations.

Deriving the complete equations required identifying mathematical tools capable of describing spacetime curvature while simultaneously satisfying rigorous physical constraints. Einstein collaborated with mathematician Marcel Grossmann to address this challenge.

These constraints narrow the space of viable candidates. For example, equations describing how matter and energy shape spacetime must be compatible with energy-momentum conservation; under weak-field and low-velocity limits, they must recover the motions that Newtonian theory already explained well. Intuition developed inside an elevator does not directly supply the mathematical machinery needed to satisfy these requirements.

The paper’s appendix notes an additional requirement: the gravitational field itself carries energy, meaning gravity acts as its own source. Consequently, one cannot simply add up separate gravitational fields via linear superposition to obtain a total field; the equations must accommodate this self-interaction, providing a physical reason why general relativity requires a non-linear mathematical structure. Encoding this property mathematically marks another major transition from physical intuition to a computable theory.

Drawing on historical records in the Zurich Notebook, the paper illustrates how Einstein and Grossmann approached the correct curvature tensors, only to discard them due to mistaken assumptions regarding static fields and the Newtonian limit, subsequently wandering through the flawed Entwurf (“Draft”) theory before correcting their errors. The hurdle lay in misjudging physical constraints and premises.

Throughout this gradual convergence on the final field equations, the calculation of Mercury’s perihelion shift provided critical reinforcement; it served as an active check during development, rather than first appearing only after publication of the final version. The resulting relationship can be summarized as: matter and energy tell spacetime how to curve, while the geometry of spacetime tells freely falling matter how to move. Between physical intuition and such a computable theory lies an extensive process of mathematical construction, debugging, and observational reconciliation.

Counterarguments and the Real Divide: Is There Only One Path to Relativity?

In a response entitled “Hot Take: LLM can ‘jump’”, Yong Zheng-Xin raises a counter-question: even if an AI model cannot replicate Einstein’s specific cognitive process, might it still reach the same theoretical destination through an alternative route?

Yong references the Feynman-style derivation, which begins with special relativity and the quantum field theory description of particle interactions, combining them with known experimental properties of gravity to constrain candidate models. The technical starting point is a massless spin-2 field. For a general reader, the operational logic is clear: select a candidate field and demand that it both satisfy established physical rules and describe interactions with matter consistently. Mutually incompatible candidates are systematically eliminated; applying these interaction-consistency requirements can recover the structure of general relativity. This pathway converts knowledge established elsewhere into filtering criteria, leaning heavily on inter-theoretic consistency and mathematical deduction without needing to reenact the elevator thought experiment.

This counterargument changes the conditions regarding available knowledge. Quantum field theory was developed long after general relativity; Feynman was already aware of the established relativistic framework. Reconstructing a known theory in retrospect is easier than discovering it for the first time. Yong acknowledges that this route does not prove Feynman would have discovered general relativity from scratch, let alone that an AI possessing only Einstein’s historical knowledge could do so today.

The value of this perspective lies in highlighting that a historical human path of discovery is not necessarily an obligatory bottleneck for all intelligent systems. Yong argues that when an AI possesses a sufficiently rich knowledge graph, it might propose candidate explanations by reconciling contradictions across disparate fields, subsequently verifying them through code, formal provers, or experimental data. How far cross-disciplinary knowledge synthesis can reach remains an open empirical question requiring systematic validation.

The two sides focus on distinct questions: Zahavy examines how novel physical premises emerge from physical experience; Yong examines whether an extensive network of interdisciplinary knowledge can provide an alternative discovery route. Explaining how humans made a historical discovery does not preclude alternative methods for future machines.

How Far Does the Paper’s Argument Actually Reach?

The contribution of LLMs can’t jump lies in deconstructing scientific discovery into distinct operational components: extracting patterns from data, deriving consequences from premises, proposing new explanatory hypotheses, and confronting theoretical predictions with empirical reality. Successes in mathematical theorem proving by AlphaProof and algorithmic discovery by AlphaEvolve demonstrate meaningful progress within their respective formulations, but they do not resolve how novel physical premises originate.

The operational pipeline of AlphaEvolve is concrete: a language model generates candidate programs, an automated evaluator executes and scores them, and an evolutionary loop selects promising candidates for iterative refinement. While it relies on verifiable feedback, this architecture cannot be characterized merely as raw gradient descent; nor can valid, novel algorithms be dismissed as mere regurgitation of training data.

Zahavy’s skepticism regarding the origins of new premises does not amount to a formal proof that existing architectures are fundamentally incapable of generating them. This position paper presents neither controlled training experiments with historical knowledge cutoffs, nor comparative rediscovery benchmarks across models, nor a mathematical impossibility theorem. Zahavy has clarified that he acknowledges genuine scientific discoveries assisted by LLMs, and concedes that scaling current paradigms may ultimately prove sufficient.

Determining whether a machine has taken such a leap requires distinguishing what it has seen, what novel hypotheses it generates, and how those hypotheses are subjected to independent testing. Fluidly explaining established relativity cannot substitute for such empirical proof; nor does Einstein’s individual discovery path rule out machines uncovering viable alternative routes. How to generate worthwhile, testable theoretical premises when data remains scarce stands as the concrete research challenge articulated by the paper.