“Importantly, the neurosymbolic method used by DeepMind’s AlphaProof and AlphaGeometry systems (which we discussed recently) which (more or less) achieved a silver-medal level performance on the 2024 International Math Olympiad, is immune to this problem.”
Our reading, the sentence that was checked:
DeepMind's neurosymbolic AlphaProof and AlphaGeometry are immune to producing confidently incorrect proofs.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Reports of LLMs mastering math have been greatly exaggerated, paragraph 5.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-reports-of-llms-mastering-math-have-dd731d-a1
We recorded 11 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-04-05. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
From the speaker
- the neurosymbolic method
- AlphaProof and AlphaGeometry systems
- 2024 International Math Olympiad
- immune to this problem
- cannot generate an incorrect proof
Our interpretation
- The contextual reference identifies the systems used for the 2024 IMO demonstration, including AlphaGeometry 2.
- The ordinary reading concerns proofs delivered as successfully established, rather than unsuccessful candidates generated during search.
- Formal correctness is relative to the supplied formal statement and the verifier's assumptions. The claim is not read as a guarantee that natural-language problems are always translated correctly.
- A demonstrated gap in an accepted proof would be material. Merely hypothesizing software bugs would not refute the claim.
Counts as supported if: Technical descriptions establish that both relevant systems accept successful proofs only through checking that guarantees the validity of the complete proof under the stated formal assumptions.
Counts as contradicted if: A relevant system demonstrably presents an invalid or incomplete proof as successfully established, or its documented acceptance procedure permits that behavior.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
AlphaProof's final Lean checking supports its part of the claim. AlphaGeometry 2 uses a symbolic deduction engine, but that alone does not establish the same complete guarantee. A later paper identifies a gap in an original AlphaGeometry proof, making broad immunity doubtful. Because the post's IMO reference points to AlphaGeometry 2, that older-system example cannot fairly settle the claim against the relevant version. The combined assertion remains unresolved.
Limitations
- The result depends materially on distinguishing AlphaGeometry from AlphaGeometry 2.
- The counterexample is a reported proof gap, not a demonstrated false theorem accepted by AlphaGeometry 2.
- Formal verification establishes consequences of formal assumptions; it does not itself establish faithful translation.
Evidence (3 items)
Supports the claim: Nature; Google DeepMind authors,
Lean kernel
Why it matters here: The paper describes the AlphaProof system used for the 2024 IMO and explicitly grounds final proof validity in Lean kernel checking. This supports the AlphaProof part of the claim, conditional on the formal statement and trusted verifier.
- Publisher
- Nature; Google DeepMind authors
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: Chervonyi et al., Google DeepMind; arXiv,
DDAR2
Why it matters here: The paper documents AlphaGeometry 2's symbolic deduction engine and distinguishes it from its predecessor. The architecture supports stronger checking than unrestricted text generation, but this passage and the inspected methods do not establish an unconditional guarantee for every delivered proof.
- Publisher
- Chervonyi et al., Google DeepMind; arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Duan et al.; arXiv,
step 9
Why it matters here: The authors identify an unproved collinearity assumption in a proof published with the original AlphaGeometry paper in 2024. This is material counterevidence to a broad claim of immunity for the whole system family, but it is not a demonstrated failure of AlphaGeometry 2.
- Publisher
- Duan et al.; arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A documented soundness and final-proof validation account for the IMO 2024 AlphaGeometry 2 system, or a reproducible invalid proof accepted by that version. Clarification that Marcus meant original AlphaGeometry would also change which counterevidence is decisive.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Reports of LLMs mastering math have been greatly exaggerated
Importantly, the neurosymbolic method used by DeepMind’s AlphaProof and AlphaGeometry systems (which we discussed recently) which (more or less) achieved a silver-medal level performance on the 2024 International Math Olympiad, is immune to this problem.
- Where
- paragraph 5
- Published
- 2025-04-05
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-reports-of-llms-mastering-math-have-dd731d-a1: Current. Outcome: Unresolved. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment marcus-reports-of-llms-mastering-math-have-dd731d-a1.