“The same goes for essentially everything an LLMs talks about; there is a degree of generalization to similar examples, but distribution shift is hard (the crux of my own work going back to 1998), and nearly any generalization can be easily broken.”
Our reading, the sentence that was checked:
For essentially everything an LLM talks about, there is a degree of generalization to similar examples, but distribution shift is hard and nearly any generalization can be easily broken.
Assessment: Unresolved Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Caricaturing Noam Chomsky, paragraph 44.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-caricaturing-noam-chomsky-e0cd11-a2
We recorded 5 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2023-03-12. 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
- essentially everything an LLM talks about (scope, stated by the speaker)
- nearly any generalization (scope, stated by the speaker)
From the speaker
- essentially everything an LLM talks about
- a degree of generalization to similar examples
- distribution shift is hard
- nearly any generalization can be easily broken
Our interpretation
- Plain reading: across nearly all domains, LLMs of the time generalize to inputs similar to their training data, but small or moderate changes to inputs (distribution shift) easily break that generalization.
- Target period: LLMs as studied from 2022-01 to the post date (2023-03-12). This is my choice, because the claim describes a present state without naming a period.
Counts as supported if: Robustness and distribution-shift studies of contemporary LLMs show, across many domains, that performance collapses under small or modest input changes, with few robust exceptions.
Counts as contradicted if: Studies of contemporary LLMs show robust generalization under distribution shift in a broad range of domains, so 'nearly any' generalization does not break easily.
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
The claim says that for essentially everything LLMs talk about, nearly any generalization can be easily broken under distribution shift. The period assessed runs from January 2022 to the post date, 12 March 2023. Shi et al. (first posted 31 January 2023) found that LLM accuracy on grade-school math 'dramatically decreased' when irrelevant information was added. The same study reported ways to reduce this drop. Razeghi et al. (first posted 15 February 2022) found that accuracy on numerical tasks depended on how often terms appeared in pretraining. Lewkowycz et al. (29 June 2022) reported that Minerva 540B scored over 80% on 10-digit addition and over 20% on 18-digit addition. That is a partial counterexample on one arithmetic task. All three studies concern math or numerical tasks. None tests the claim's cross-domain scope, and 'easily broken' has no set threshold. The cited evidence neither confirms nor refutes the claim.
Limitations
- All the evidence comes from math and numerical tasks, not the range of domains the claim covers.
- The contract notes that 'easily broken' does not say how large an input change counts.
- The copy of e1 that was observed is a version revised in June 2023, after the post. The finding is attributed to the first version from January 2023.
- The Minerva figure covers addition only and does not test other kinds of distribution shift.
Evidence (3 items)
Supports the claim: arXiv (Shi et al., Google; ICML 2023),
find that the model performance is dramatically decreased when irrelevant information is included.
Why it matters here: Shows a generalization (grade-school math) broken by a small, irrelevant change to the input. It supports brittleness in one domain, and the authors also report ways to reduce it.
- Where
- Abstract
- Publisher
- arXiv (Shi et al., Google; ICML 2023)
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv (Razeghi et al.),
our results raise the question of how much models actually generalize beyond pretraining data
Why it matters here: Consistent with generalization limited to cases like the training data, but limited to numerical tasks in GPT-style models trained on the Pile.
- Where
- Abstract
- Publisher
- arXiv (Razeghi et al.)
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Goes against the claim: arXiv (Lewkowycz et al., Google),
The model achieves over 80% accuracy on 10-digit addition and over 20% accuracy on 18-digit addition.
Why it matters here: Minerva held over 80% accuracy on 10-digit addition, far from typical training examples. That is generalization that did not break easily, though it fell off at 18 digits.
- Publisher
- arXiv (Lewkowycz et al., Google)
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
A robustness evaluation of LLMs from 2022 to early 2023, covering many domains, that measured how often their generalizations failed under small input changes.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Caricaturing Noam Chomsky
The same goes for essentially everything an LLMs talks about; there is a degree of generalization to similar examples, but distribution shift is hard (the crux of my own work going back to 1998), and nearly any generalization can be easily broken.
- Where
- paragraph 44
- Published
- 2023-03-12
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-caricaturing-noam-chomsky-e0cd11-a2: 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-caricaturing-noam-chomsky-e0cd11-a2.