“LLMs, like their predecessors generalize well to similar items but struggle with reliability in unfamiliar regimes, even minor variations sometimes break them.”
Our reading, the sentence that was checked:
LLMs generalize well to similar items but are unreliable in unfamiliar regimes, and minor variations sometimes break them.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Why I don’t share Sam Altman’s confidence that AGI is basically a solved problem, paragraph 9.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-sam-altman-thinks-that-agi-is-basically-95ab84-a1
We recorded 16 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 2025-01-06. 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
- in unfamiliar regimes (qualifier, stated by the speaker)
From the speaker
- generalize well to similar items
- struggle with reliability in unfamiliar regimes
- even minor variations sometimes break them
Our interpretation
- The ordinary reading is a tendency: models can perform well on familiar task formats while making recurring errors on altered versions. Sometimes does not mean all models or all alterations.
- Break means producing an incorrect answer, not crashing or losing all capability.
- The target is the 2024 research cycle through January 6, 2025, inferred from the surrounding references to recent reasoning evaluations.
- Benchmark familiarity is a proxy for similarity to training tasks; high benchmark accuracy alone does not prove uncontaminated generalization.
Counts as supported if: Controlled comparisons show competent baseline performance alongside reduced reliability on unfamiliar variants, including small changes that sometimes cause incorrect answers.
Counts as contradicted if: The relevant comparisons show preserved reliability under unfamiliar and minor variants, and the reported failures are explained by invalid questions or evaluation 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
The controlled GSM-Plus evaluation directly supports the stated pattern: models can answer original questions yet struggle with variants. Its findings vary by model and perturbation, which fits the word sometimes. The evidence supports the practical reliability observation without establishing a universal inability to reason.
Limitations
- The cited evidence concerns mathematical word problems, not every application of LLMs.
- Not every tested variant is a minor change; some alter the required operation or remove necessary information.
- Baseline accuracy cannot fully distinguish learned generalization from training-data familiarity.
- The comparison with earlier generations of neural models was not separately quantified.
Evidence (1 item)
Supports the claim: arXiv; Qintong Li and colleagues,
struggling with answering the variations
Why it matters here: The study compares twenty-five models and four prompting techniques on original questions and controlled variants, finding competent original-task performance alongside failures on altered tasks. It also reports relative resilience to some numerical and wording changes, consistent with Marcus's qualification that small changes cause failures only sometimes.
- Publisher
- arXiv; Qintong Li 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, Why I don’t share Sam Altman’s confidence that AGI is basically a solved problem
LLMs, like their predecessors generalize well to similar items but struggle with reliability in unfamiliar regimes, even minor variations sometimes break them.
- Where
- paragraph 9
- Published
- 2025-01-06
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-sam-altman-thinks-that-agi-is-basically-95ab84-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-sam-altman-thinks-that-agi-is-basically-95ab84-a1.