“models still don’t reason reliably (as anticipated in What to Expect When Your Expecting GPT-4)”
Our reading, the sentence that was checked:
LLMs still do not reason reliably as of April 2025.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Scaling is over, the bubble may be deflating, LLMs still can’t reason, and you can’t trust Sam, paragraph 12.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-scaling-is-over-the-bubble-may-be-719549-a1
We recorded 8 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 2025-04-06. 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)
From the speaker
- models still don’t reason reliably
- LLMs
- mathematical reasoning
- challenging Olympiad-level problems
Our interpretation
- Reliable reasoning means dependable reasoning across relevant problems, including unfamiliar tasks and logically valid intermediate steps.
- The claim does not mean models never reason successfully, or that any single error establishes general unreliability.
- Recurring substantive failures across models and problem types support the ordinary reading.
- The post's mathematical examples make rigorous proof generation particularly relevant, but elementary reasoning evaluations are also considered.
- For this ongoing-state observation, the target period is the early-2025 evaluation cycle through April 6. This window is an evaluator choice because model research has no standard reporting period.
Counts as supported if: Evaluations of current frontier LLMs document recurring substantive reasoning failures, including systematic errors or invalid proofs, beyond isolated mistakes.
Counts as contradicted if: Broad, rigorous evaluations establish dependable reasoning across relevant novel problems and controlled variations, and show that purported recurring failures are evaluation artifacts or unrepresentative exceptions.
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
Contemporaneous studies document recurring reasoning failures rather than isolated wrong answers. Expert grading found very poor proof performance on newly released USAMO problems. A separate benchmark-cleaning study found systematic failures even on elementary tasks after removing evaluation errors. These findings support the claim of unreliable reasoning. They coexist with substantial reasoning gains and do not establish that LLMs never reason successfully.
Limitations
- The cited versions are preprints.
- USAMO results concern six unusually difficult problems and should not be generalized into an everyday-task failure rate.
- The tested models do not exhaust the April 6 frontier.
- Reliability remains task-dependent, and Marcus supplied no numerical threshold.
Evidence (3 items)
Supports the claim: arXiv; Petrov and colleagues,
less than 5% on average
Why it matters here: Experts graded complete solutions to six newly released USAMO problems from six reasoning models, including o1-pro, o3-mini, Claude 3.7, and DeepSeek R1. The reported percentage concerns graded performance, not a general accuracy rate for all reasoning tasks.
- Where
- Abstract
- Publisher
- arXiv; Petrov and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv; Vendrow and colleagues,
consistently struggle
Why it matters here: After cleaning benchmark errors, the researchers still found systematic failures on elementary tasks, including mathematical word problems and event ordering. This extends the reliability concern beyond exceptionally difficult Olympiad proofs.
- Publisher
- arXiv; Vendrow and colleagues
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv; Mahdavi and colleagues,
enhanced capabilities in reasoning
Why it matters here: The study explicitly acknowledges reasoning progress and high scores on established math benchmarks while documenting weaknesses in proof generation and verification. This weighs against interpreting Marcus as establishing that LLMs cannot reason at all.
- Publisher
- arXiv; Mahdavi 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, Scaling is over, the bubble may be deflating, LLMs still can’t reason, and you can’t trust Sam
models still don’t reason reliably (as anticipated in What to Expect When Your Expecting GPT-4)
- Where
- paragraph 12
- Published
- 2025-04-06
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-scaling-is-over-the-bubble-may-be-719549-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-scaling-is-over-the-bubble-may-be-719549-a1.