Position: LLMs can't jump — part 3/4

Cleaned-text code points [29512, 43382). Bibliography omitted here; source.md includes it.

LLMs can’t jump
offered no satisfying explanation for his imagined observa-
tion. He faced a silence in the space of language—a lack of
prior symbolic representation:
”The words or the language, as they are written
or spoken, do not seem to play any role in my
mechanism of thought.”
– Albert Einstein
To fill this void, he relied on a physical prior. Because
the simulated sensory experience of acceleration was indis-
tinguishable from the remembered sensory experience of
gravity, Einstein abducted that they must be the same phe-
nomenon. The field inside the box was not a fake inertial
effect; it was, by definition, a genuine gravitational field.
From Chinese Rooms to World Models. This cognitive
process—anchoring abstract symbols in tangible physical
simulations—is known as manipulative abduction (Magnani
et al., 2009). This stands in sharp contrast to the operational
mechanics of LLMs.
While LLMs excel at Induction (finding patterns in data),
they lack the sensory agency required to ground these sym-
bols in physical reality. They operate as high-dimensional
”Chinese Rooms” (Harnad, 1990), manipulating the lan-
guage of physics without access to the physical referents
that give that language meaning. This limitation prevents
the AI from making the Abductive Jump (E→A ). While
Einstein could ground his axioms in the physical experi-
ence of a falling body, an LLM is confined to the logical
deduction of existing texts.
It is worth addressing whether emerging techniques such as
prompt engineering, in-context learning, or iterative context
refinement might unlock abductive capabilities in LLMs.
While these techniques significantly improve reasoning per-
formance within an established symbolic space—enabling
the model to navigate and recombine existing concepts
more efficiently—they are fundamentally constrained by
the boundaries of that symbolic space. Prompting cannot
provide the sensory grounding required for manipulative ab-
duction, which entails generating axioms without symbolic
precedent. It cannot bridge the gap between abstract tokens
and raw, un-symbolized physical experience.
This deficit in physical grounding is central to recent cri-
tiques of AI. Experts contend that despite linguistic mastery,
current systems lack the spatial intelligence(Li, 2025) and in-
ternal world models(LeCun, 2022) required to reason about
physical reality. Without the ability to perceive or interact
with the world, LLMs struggle with spatial reasoning tasks
that are trivial for toddlers .
The emergence of World Models offers a pathway to bridge
this divide, but a critical distinction must be drawn be-
tween visual prediction and interactive simulation. Cur-
rent video generation models like Veo exhibit intuitive
physics(Hassabis, 2025) primarily as a byproduct of sta-
tistical correlation; they correctly generate a falling apple
not because they model gravity, but because falling is the
dominant continuation of unsupported object in their train-
ing distribution.
However, recent architectures like Genie (Bruce et al.,
2024) mark a fundamental shift by introducing action-
controllability into generative world models. Unlike passive
video generators, Genie learns an action space that allows
for agentic intervention—a prerequisite for Manipulative
Abduction (thinking by doing). To replicate Einstein’s eleva-
tor thought experiment, an AI cannot merely watch a video
of an elevator; it must possess the capacity for counterfac-
tual intervention (Pearl & Mackenzie, 2018). It must be able
to essentially take control of the simulation to conceptually
cut the cable.
We argue that physically consistent, action-controllable
world models are not just useful, but are theoretically suffi-
cient to enable abductive leaps. An interactive world model
functions as a ”synthetic laboratory” where an agent can
run counterfactual simulations and test reasoning that goes
beyond standard linguistic training data. Rather than merely
predicting the next token or frame based on statistical cor-
relation, the world model allows the agent to manipulate
the environment, observe the consequences of novel actions,
and abduce underlying rules. The sensory and interactive
feedback from this synthetic laboratory serves as the ground-
ing mechanism necessary to propose new axioms where no
symbolic precedents exist in the language data. We propose
that future iterations of such interactive environments, oper-
ating on a consistent latent physics manifold rather than just
pixels, will provide the substrate necessary to transform the
Abductive Jump from a mystical insight into a reproducible
algorithmic process.
Lastly, its important to note that Einstein relied on his Phys-
ical Prior, using the sensation of gravity to prune the search
space of possible axioms. However, manipulative abduction
extends beyond physics. Historical scientific revolutions are
often driven by strong, pre-symbolic intuitions—whether
Kepler’s Neoplatonic belief in the centrality of the Sun or
the ’objective anger’ that drove Marx’s modeling of capital.
To automate invention, we may need systems that do not
just simulate the world, but hold strong beliefs or priors
about how that world should be structured, using simulation
to test those specific intuitions.
5. Conclusion
In this paper, we posed a fundamental question: Could a
modern Artificial Intelligence, given the knowledge avail-
7

LLMs can’t jump
able to Einstein, invent General Relativity? Our investiga-
tion suggests that for current Large Language Models, the
answer is no. While the field has successfully mechanized
Induction (via statistical compression) and Deduction (via
formal verification), these mechanisms alone are insufficient
to sustain the cycle of scientific invention.
The prevailing Creativity as Compression hypothesis fails to
account for this discovery because it presumes the existence
of a pervasive error signal. Yet, the Newtonian paradigm
faced no such crisis, and the data required to validate Gen-
eral Relativity did not exist until years after its formulation.
Furthermore, while the deduction paradigm offers a path to
derive field equations once axioms are set, it is ultimately a
downstream process—a verification step within the inven-
tion loop, rather than the mechanism of invention.
This limitation is visible even in the apex of today’s auto-
mated discovery systems. Agents such as Sakana’s AI Scien-
tist (Lu et al., 2024) and Google DeepMind’s AlphaEvolve
(Novikov et al., 2025) demonstrate the immense power of
mechanizing scientific loops and evolutionary optimization.
However, they highlight the very abductive gap we identify.
The AI Scientist recombines existing symbolic concepts
to optimize metrics—a sophisticated ”Chinese Room” op-
eration that lacks the sensory grounding to invent axioms
without symbolic precedent. Similarly, while AlphaEvolve
excels at optimization within a fixed framework, it relies on
a gradient; Einstein, by contrast, had no error signal from
Newtonian mechanics to drive his discovery. These systems
lack the embodied world model required to perform the
counterfactual physical simulations that drive the abductive
Jump to entirely new paradigms.
Our analysis suggests that the critical bottleneck is this
intuitive Jump from sensory experience to formal axioms
(E→A ). We frame this work not as a definitive proof, but
as establishing a theoretical framework to guide future em-
pirical testing of abductive capabilities in generative systems.
Einstein did not discover General Relativity by searching
over symbols; he discovered it by simulating the sensual
experience of a falling observer. The formulation of the
Equivalence Principle was a self-contained act of physi-
cal abduction, where the premises were established solely
through internal simulation, independent of immediate ex-
ternal verification.
To build an AI capable of true invention, we must there-
fore move beyond systems that merely read scientific liter-
ature to systems that can perceive the physical world. The
emergence of physically consistent World Models offers a
pathway to a synthetic laboratory. By enabling agents to
run counterfactual simulations—to experience the physi-
cal consequences of a thought experiment—we may finally
mechanize the feedback loop between intuition and logic.
Finally, we emphasize that this proposal is specifically tai-
lored to the physical sciences, where the object of study is
external material reality. In abstract domains such as Mathe-
matics or Computer Science, the Sense Experience (E) may
be grounded in high-dimensional topology or have other
goals such as generality or minimality. While the necessity
of the Abductive Jump remains universal, the nature of the
simulation must be adapted to the ontology of the discipline:
for physics, the substrate is the world; for mathematics, it is
the abstract landscape of formal systems.
References
Achim, T., Best, A., Bietti, A., Der, K., F´ed´erico, M., Gukov,
S., Halpern-Leistner, D., Henningsgard, K., Kudryashov,
Y ., Meiburg, A., Michelsen, M., Patterson, R., Rodriguez,
E., Scharff, L., Shanker, V ., Sicca, V ., Sowrirajan, H.,
Swope, A., Tamas, M., Tenev, V ., Thomm, J., Williams,
H., and Wu, L. Aristotle: Imo-level automated theorem
proving, 2025. URL https://arxiv.org/abs/
2510.01346.
Bruce, J., Dennis, M., Edwards, A., Parker-Holder, J.,
Shi, Y ., Hughes, E., Lai, M., Mavalankar, A., Steiger-
wald, R., Apps, C., Aytar, Y ., Bechtle, S., Behbahani,
F., Chan, S., Heess, N., Gonzalez, L., Osindero, S.,
Ozair, S., Reed, S., Zhang, J., Zolna, K., Clune, J.,
de Freitas, N., Singh, S., and Rockt ¨aschel, T. Ge-
nie: Generative interactive environments, 2024. URL
https://arxiv.org/abs/2402.15391.
Chollet, F., Knoop, M., Kamradt, G., and Landers, B. Arc
prize 2024: Technical report, 2025. URL https://
arxiv.org/abs/2412.04604.
Gambardella, A., Iwasawa, Y ., and Matsuo, Y . Language
models do hard arithmetic tasks easily and hardly do easy
arithmetic tasks, 2024. URL https://arxiv.org/
abs/2406.02356.
Harnad, S. The symbol grounding problem.Physica D:
Nonlinear Phenomena, 42(1-3):335–346, 1990.
Hassabis, D. Demis hassabis: Future of ai, simulating
reality, physics and video games — lex fridman pod-
cast 475, 2025. URL https://lexfridman.com/
demis-hassabis-2-transcript/.
Hubert, T., Mehta, R., Sartran, L., Horv ´ath, M. Z., ˇZuˇzi´c,
G., Wieser, E., Huang, A., Schrittwieser, J., Schroecker,
Y ., Masoom, H., Bertolli, O., Zahavy, T., Mandhane, A.,
Yung, J., Beloshapka, I., Ibarz, B., Veeriah, V ., Yu, L.,
Nash, O., Lezeau, P., Mercuri, S., S¨onne, C., Mehta, B.,
Davies, A., Zheng, D., Pedregosa, F., Li, Y ., von Glehn,
I., Rowland, M., Albanie, S., Velingker, A., Schmitt,
8

LLMs can’t jump
S., Lockhart, E., Michalewski, H., Sonnerat, N., Hass-
abis, D., Kohli, P., and Silver, D. Olympiad-level for-
mal mathematical reasoning with reinforcement learning.
Nature, 2025. URL https://www.nature.com/
articles/s41586-025-09833-y.
LeCun, Y . A path towards autonomous machine intelligence
version 0.9. 2, 2022-06-27.Open Review, 62(1):1–62,
2022.
Li, F. F. From words to worlds: Spatial in-
telligence is ai’s next frontier, 2025. URL
https://drfeifei.substack.com/p/
from-words-to-worlds-spatial-intelligence .
Lu, C., Lu, C., Lange, R. T., Foerster, J., Clune, J., and
Ha, D. The ai scientist: Towards fully automated
open-ended scientific discovery, 2024. URL https:
//arxiv.org/abs/2408.06292.
Magnani, L. et al.Abductive cognition: The epistemological
and eco-cognitive dimensions of hypothetical reasoning,
volume 3. Springer, 2009.
mathlib Community, T. The lean mathematical library.
CoRR, abs/1910.09336, 2019. URL http://arxiv.
org/abs/1910.09336.
Norton, J. D. Einstein’s pathway to general relativity, 2020.
URL https://sites.pitt.edu/˜jdnorton/
teaching/HPS_0410/chapters/general_
relativity_pathway/index.html.
Novikov, A., V˜u, N., Eisenberger, M., Dupont, E., Huang,
P.-S., Wagner, A. Z., Shirobokov, S., Kozlovskii, B., Ruiz,
F. J. R., Mehrabian, A., Kumar, M. P., See, A., Chaudhuri,
S., Holland, G., Davies, A., Nowozin, S., Kohli, P., and
Balog, M. Alphaevolve: A coding agent for scientific and
algorithmic discovery, 2025. URL https://arxiv.
org/abs/2506.13131.
Pearl, J. and Mackenzie, D.The Book of Why: The New
Science of Cause and Effect. Basic Books, 2018.
Peirce, C. S.Collected papers of charles sanders peirce.
Harvard University Press., 1934.
Schaeffer, H. Learning partial differential equations via data
discovery and sparse optimization.Proceedings of the
Royal Society A: Mathematical, Physical and Engineer-
ing Sciences, 473(2197):20160446, 2017.
Schmidhuber, J. Driven by compression progress: A simple
principle explains essential aspects of subjective beauty,
novelty, surprise, interestingness, attention, curiosity, cre-
ativity, art, science, music, jokes. InWorkshop on antici-
patory behavior in adaptive learning systems, pp. 48–76.
Springer, 2008.
Vafa, K., Chang, P. G., Rambachan, A., and Mullainathan, S.
What has a foundation model found? using inductive bias
to probe for world models. InProceedings of the 42nd
International Conference on Machine Learning, volume
267, Vancouver, Canada, 2025. PMLR.
Wu, T. and Tegmark, M. Toward an artificial intelligence
physicist for unsupervised learning.Physical Review E,
100(3), September 2019. ISSN 2470-0053. doi: 10.1103/
physreve.100.033311. URL http://dx.doi.org/
10.1103/PhysRevE.100.033311.
Yang, Z., Ding, M., Lv, Q., Jiang, Z., He, Z., Guo, Y .,
Bai, J., and Tang, J. GPT can solve mathematical
problems without a calculator, 2024. URL https:
//openreview.net/forum?id=LojXXo2xaf.
Zahavy, T., Veeriah, V ., Hou, S., Waugh, K., Lai, M.,
Leurent, E., Tomasev, N., Schut, L., Hassabis, D., and
Singh, S. Diversifying ai: Towards creative chess with
alphazero, 2024. URL https://arxiv.org/abs/
2308.09175.
9