“The unfairly dismissed Apple reasoning paper laid bare the generalization issues; another paper called “Is Chain-of-Thought Reasoning of LLMs Mirage? A Data Distribution Lens” put a further nail in the LLM reasoning and generalization coffin.”
Our reading, the sentence that was checked:
Research papers (Apple's reasoning paper and the chain-of-thought mirage paper) demonstrated LLM reasoning and generalization failures.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, A trillion dollars is a terrible thing to waste, paragraph 6.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-a-trillion-dollars-is-a-terrible-01726b-a1
We recorded 14 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 2025-11-27. 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
- Apple reasoning paper
- laid bare the generalization issues
- Is Chain-of-Thought Reasoning of LLMs Mirage? A Data Distribution Lens
- LLM reasoning and generalization
Our interpretation
- Assess the checkable proposition that both papers demonstrated reasoning or generalization failures in the models and tasks they tested.
- The ordinary reading of demonstrated failures does not require proof that all LLMs are incapable of reasoning or that future improvements are impossible.
- The judgment that Apple was unfairly dismissed is evaluative and is not separately scored.
- The target period covers the cited research available through publication. The source supplies no lower date boundary.
Counts as supported if: Both papers contain empirical demonstrations of reasoning or generalization failures, and methodological criticism does not invalidate every relevant finding.
Counts as contradicted if: The papers do not demonstrate the claimed failures, or convincing checks establish that their relevant findings entirely result from invalid tests or measurement errors.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
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
Both papers demonstrate failures within their tested settings. Apple reports accuracy collapse on increasingly complex puzzles; the other paper reports poor transfer under controlled distribution shifts. Independent replication challenges part of Apple's setup while finding other failures remain. This supports the screened proposition about demonstrated failures, but not the stronger suggestion that LLM reasoning is definitively impossible.
Limitations
- The result depends on reading the claim as empirical task failures rather than universal incapacity.
- Synthetic tasks and selected models limit generalization.
- The fairness of criticism directed at Apple is not adjudicated.
Evidence (3 items)
Supports the claim: arXiv; Apple research authors,
complete accuracy collapse
Why it matters here: Apple reports failures as puzzle complexity increases. Its controlled puzzle experiments support the existence of failures in tested reasoning models, with a narrower scope than universal inability to reason.
- Publisher
- arXiv; Apple research authors
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Zhao and colleagues,
distribution discrepancy increases
Why it matters here: The controlled experiments report failures when transformations or their combinations differ from training. These results demonstrate generalization problems within the study's synthetic environment.
- Publisher
- arXiv; Zhao and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Dellibarda Varela and colleagues,
testing unsolvable configurations
Why it matters here: The replication challenges the interpretation of River Crossing failures and reports success on solvable instances. It also reports remaining Tower of Hanoi failures, so it qualifies rather than eliminates the evidence of limitations.
- Where
- Abstract
- Publisher
- arXiv; Dellibarda Varela and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, A trillion dollars is a terrible thing to waste
The unfairly dismissed Apple reasoning paper laid bare the generalization issues; another paper called “Is Chain-of-Thought Reasoning of LLMs Mirage? A Data Distribution Lens” put a further nail in the LLM reasoning and generalization coffin.
- Where
- paragraph 6
- Published
- 2025-11-27
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-a-trillion-dollars-is-a-terrible-01726b-a1: Current. Outcome: Supported. 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-a-trillion-dollars-is-a-terrible-01726b-a1.