“June, 2025: the Apple reasoning paper confirmed that even with “reasoning”, LLMs still can’t solve distribution shift, the core Achille’s heel in neural networks that I have been writing about for nearly 30 years.”
Our reading, the sentence that was checked:
Even with reasoning, LLMs still cannot handle distribution shift, as the June 2025 Apple paper confirmed.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Game over. AGI is not imminent, and LLMs are not the royal road to getting there., paragraph 2.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-the-last-few-months-have-been-devastating-8b9119-a1
We recorded 2 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 2025-10-18. 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)
Conditions
- even with reasoning (qualifier, stated by the speaker)
From the speaker
- June, 2025
- the Apple reasoning paper confirmed
- even with “reasoning”
- LLMs still can’t solve distribution shift
Our interpretation
- The plain reading is that reasoning models still lacked reliable generalization beyond their training distribution, and Apple's paper established that limitation.
- This does not mean that models must fail every unfamiliar example, or that future models can never improve.
- The target is the June 2025 finding expressly identified in the claim. Later experiments are context unless they directly reassess those June results.
- Increasing puzzle complexity is relevant to generalization, but identifying failures specifically caused by distribution shift requires information about the training distribution.
Counts as supported if: The June paper and reliable reassessments establish failures of reasoning models under identified distribution shifts, sufficient to substantiate the claimed general limitation.
Counts as contradicted if: Controlled evidence establishes reliable generalization across the relevant shifts, or shows that the attributed June finding results from experimental constraints rather than the claimed limitation.
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
Apple reported substantial failures despite reasoning, so Marcus's concern has experimental support. But its complexity experiments do not establish the broader distribution-shift conclusion as stated. A June critique disputes the interpretation, and the August follow-up studies a different controlled setting. These records establish a narrower limitation without settling the full claim.
Limitations
- The competing papers are evidence of a scientific dispute, not equally decisive demonstrations.
- Interpreting the claim solely as failure on Apple's tested puzzles would make it substantially easier to support.
- Later model capabilities cannot retroactively settle the June observation.
Evidence (3 items)
Supports the claim: Apple researchers via arXiv,
complete accuracy collapse
Why it matters here: Supports the narrower finding that reasoning did not prevent failure on increasingly complex puzzles. The experiments vary complexity; their discussion treats limited training exposure as a possible explanation rather than establishing the training distribution.
- Where
- Abstract
- Publisher
- Apple researchers via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: A. Lawsen via arXiv,
experimental design limitations
Why it matters here: Challenges the claim that Apple's results confirmed a general limitation. The critique raises output-length and scoring concerns, but does not itself establish robust generalization by reasoning models.
- Publisher
- A. Lawsen via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Chengshuai Zhao and coauthors via arXiv,
beyond training distributions
Why it matters here: This follow-up directly studies distribution shifts using models trained in a controlled environment. It supports concern about generalization, but concerns different experiments outside the June target period.
- Where
- Abstract
- Publisher
- Chengshuai Zhao and coauthors via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A controlled replication addressing output and scoring concerns, with identified training and test distributions and a justified scope for the claimed generalization limitation.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Game over. AGI is not imminent, and LLMs are not the royal road to getting there.
June, 2025: the Apple reasoning paper confirmed that even with “reasoning”, LLMs still can’t solve distribution shift, the core Achille’s heel in neural networks that I have been writing about for nearly 30 years.
- Where
- paragraph 2
- Published
- 2025-10-18
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-the-last-few-months-have-been-devastating-8b9119-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-the-last-few-months-have-been-devastating-8b9119-a1.