2026-08-17
DeepMind position paper: LLMs have mechanized induction and deduction, but Einstein's leap to new axioms needs abduction, and the path there runs through interactive world models.
A popular view in ML, systematized by Schmidhuber in 2008, holds that scientific discovery is data compression: find the short program that explains the observations. The corollary is that discovery decomposes into induction (statistical pattern finding) plus deduction (formal proof), and that an LLM given enough compute should in principle be able to invent the next general relativity.
Tom Zahavy, a researcher on the AlphaProof team at Google DeepMind, argues against that corollary in an ICML 2026 position paper. His case study is the sketch Einstein drew for Maurice Solovine: between sense experience (E) and the axioms (A) sits a jump (J) with no logical road across it. The question is concrete: handed everything Einstein knew, could a modern LLM reach the field equations on its own? The answer argued here is no. The missing mechanism has a name, abduction.
The framework comes from Peirce's triad: deduction runs rule plus case to result, induction runs case plus result to rule, and abduction runs rule plus result to case, inventing a new explanation for a surprising outcome.
Against that triad the paper walks through the three phases of general relativity's birth from 1907 to 1915 and eliminates the first two readings:
What remains is the positive claim: the step from E to A is abduction, specifically what Magnani calls manipulative abduction, grounded in embodied simulation. Einstein's self-described happiest thought is the specimen: an observer falling from a roof feels no gravitational field. The observation is first manufactured in simulation, then explained, and that is not symbol shuffling. ARC-AGI sits halfway there. With only two to five grid pairs per task, too sparse for induction and too underspecified for deduction, it forces an abductive leap, but it tests only the explanation half and carries no embodied component.
A position paper has no benchmarks. What can be checked is the chain of cited facts:
| Claim | Basis |
| Induction lacked an error signal | Mass equivalence verified to 10^-9; the Mercury anomaly absorbed by the Vulcan hypothesis |
| Deduction is already mechanized | AlphaProof IMO silver 2024; several systems gold in 2025; Aristotle on open problems |
| The 1913 answer was within reach | The Riemann tensor was found and discarded over one wrong assumption, costing two years |
| Abduction is the gap | LLMs as high-dimensional Chinese Rooms, manipulating physics language with no physical referents |
The verdict on current systems follows: Sakana's AI Scientist recombines existing symbolic concepts to optimize metrics, and AlphaEvolve optimizes inside a fixed framework where a gradient exists. Neither crosses the step of generating axioms with no error signal to follow.
For AI4Science practitioners the ceiling gets drawn in a useful place: verification and optimization inside an existing paradigm still have large headroom for automation, while expecting a system to invent a new paradigm runs into a missing mechanism. For model builders the paper names a concrete direction: interactive, physically consistent world models that support counterfactual intervention. It is careful to separate video generators such as Veo, where a falling apple reflects the dominant continuation in the training distribution rather than a model of gravity, from action-controllable world models such as Genie. Replaying the elevator thought experiment requires the capacity to cut the cable. The landing zone overlaps with LeCun's world model agenda and Fei-Fei Li's spatial intelligence argument.
The appendix discloses that the author's physics training ended at the undergraduate level, that the historical reading relies on secondary sources, and that Gemini was used to polish the prose. It is a single-author position paper, not an experimental study.
The larger issue is falsifiability. The chain from induction needing an error signal and deduction being unable to generate premises to a third mechanism existing is philosophy, and the claim that LLMs are structurally incapable of abduction is neither tested nor easy to test. One counterpath goes untreated: Einstein's own inputs were also finite, and if vast multimodal data plus the right inductive bias could yield an equivalent capability, structurally impossible weakens to unlikely for current architectures. Whether the language-has-no-prior-representation argument still holds for multimodal models is left open as well. Read it as a directional argument, not a theorem.