“Consequently, any “reasoning” or “planning” in LLMs is hit or miss, viable if the particulars of a given situation are close enough to what is in the training set, but highly vulnerable if they are not.”
Our reading, the sentence that was checked:
Marcus says any reasoning or planning in LLMs is hit or miss: viable if a situation's particulars are close enough to the training set, but highly vulnerable if they are not.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, AlphaProof, AlphaGeometry, ChatGPT, and why the future of AI is neurosymbolic, paragraph 38.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-alphaproof-alphageometry-chatgpt-19c5a2-a1
We recorded 17 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 2024-07-28. 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
- if the particulars are close enough to the training set (scope, stated by the speaker)
From the speaker
- any “reasoning” or “planning” in LLMs
- hit or miss
- viable if the particulars of a given situation are close enough to what is in the training set
- highly vulnerable if they are not
Our interpretation
- He describes an ongoing state with no stated period. I use the GPT-4-era LLMs current at the post: from GPT-4's release (2023-03-14) to the post date (2024-07-28).
- Plain reading: on controlled tests, LLM reasoning accuracy drops a lot when tasks are changed to versions that are less common in training data. It need not fall to zero.
Counts as supported if: Controlled studies of then-current LLMs show large, consistent drops in accuracy on unfamiliar or low-probability variants of tasks they handle well in their standard forms.
Counts as contradicted if: Controlled studies show then-current LLMs perform about as well on novel or counterfactual variants as on familiar ones.
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
Two controlled studies of GPT-4-era models back the claim. Wu et al. (NAACL 2024) tested 11 tasks and found accuracy fell substantially and consistently on counterfactual variants that depart from familiar defaults. McCoy et al. found GPT-4 decoded a simple cipher 51% of the time when the output was common text, but only 13% when it was uncommon. That is the pattern Marcus describes: workable when a case resembles training data, fragile when it does not. Wu et al. also found nontrivial performance on the variants, which fits 'hit or miss' rather than total failure.
Limitations
- Reasoning models released after July 2024 are not assessed.
- 'Any reasoning' is broad; the studies cover selected tasks.
Evidence (2 items)
Supports the claim: Wu et al., MIT/Boston University (arXiv; NAACL 2024),
Across a suite of 11 tasks, we observe nontrivial performance on the counterfactual variants, but nevertheless find that performance substantially and consistently degrades compared to the default conditions.
Why it matters here: Reasoning accuracy falls substantially and consistently on unfamiliar variants of the same tasks. Performance is still nontrivial, which fits 'hit or miss' rather than total failure.
- Where
- Abstract
- Publisher
- Wu et al., MIT/Boston University (arXiv; NAACL 2024)
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Supports the claim: McCoy et al., Princeton (arXiv),
For instance, GPT-4's accuracy at decoding a simple cipher is 51% when the output is a high-probability word sequence but only 13% when it is low-probability.
Why it matters here: GPT-4's success on a deterministic task depends heavily on how familiar the content is, which matches his claim of being viable near the training data and fragile away from it.
- Where
- Abstract
- Publisher
- McCoy et al., Princeton (arXiv)
- 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, AlphaProof, AlphaGeometry, ChatGPT, and why the future of AI is neurosymbolic
Consequently, any “reasoning” or “planning” in LLMs is hit or miss, viable if the particulars of a given situation are close enough to what is in the training set, but highly vulnerable if they are not.
- Where
- paragraph 38
- Published
- 2024-07-28
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-alphaproof-alphageometry-chatgpt-19c5a2-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-alphaproof-alphageometry-chatgpt-19c5a2-a1.