“neural networks of various kinds can generalize within a training distribution of data they are exposed to, but their generalizations tend to break down outside that distribution.”
Our reading, the sentence that was checked:
Neural networks of various kinds can generalize within the training distribution of data they are exposed to, but their generalizations tend to break down outside that distribution.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, A knockout blow for LLMs?, paragraph 8.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-a-knockout-blow-for-llms-68a208-a1
We recorded 19 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 2025-06-07. 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
- neural networks of various kinds (scope, stated by the speaker)
- tend to (qualifier, stated by the speaker)
From the speaker
- neural networks of various kinds
- can generalize within a training distribution
- generalizations tend to break down outside that distribution
Our interpretation
- The plain reading is a recurring empirical weakness: networks that perform well on held-out examples resembling their training data often perform substantially worse after a distribution shift.
- The words 'tend to' allow successful exceptions. 'Break down' means a substantial loss of successful generalization, without requiring accuracy to fall to zero.
- The surrounding passage explicitly describes a longstanding pattern dating from 1998 through the post. I assess that historical research record through 2025-06-07. This scientific claim has no quarterly reporting cycle.
- A new example from the same distribution is distinguished from a test that changes combinations, sequence lengths, domains, or other relevant data properties.
Counts as supported if: Controlled studies across several neural-network variants and tasks show successful within-distribution generalization alongside substantial performance losses under specified distribution shifts.
Counts as contradicted if: Broad, controlled comparisons across varied neural networks and meaningful distribution shifts show that within-distribution performance generally carries over, making substantial breakdown an exceptional rather than recurring pattern.
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 ordinary reading is supported. WILDS finds substantial performance losses across ten distribution-shift datasets, and SCAN finds failures on systematic generalization despite success on easier held-out commands. These are direct experimental examples of the recurring pattern Marcus describes. Successful meta-learning results limit the conclusion: neural networks can generalize systematically under suitable training. They do not overturn a claim expressly qualified by 'tend to'.
Limitations
- Support concerns a recurring empirical weakness, not an impossibility theorem or a measured failure rate for all neural networks.
- The selected studies do not establish the behavior of every LLM or explain every reasoning failure.
- Successful systematic generalization depends on the training procedure and the test being used.
Evidence (3 items)
Supports the claim: Proceedings of Machine Learning Research,
On each dataset, we show that standard training yields substantially lower out-of-distribution than in-distribution performance.
Why it matters here: WILDS compares within-distribution and shifted performance across ten datasets covering varied applications. The consistent substantial gaps support the claimed recurring weakness.
- Where
- Abstract, sentence 4
- Publisher
- Proceedings of Machine Learning Research
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Supports the claim: Proceedings of Machine Learning Research,
However, when generalization requires systematic compositional skills (as in the "dax" example above), RNNs fail spectacularly.
Why it matters here: The study tests several recurrent-network variants and reports successful generalization when training and test commands differ little, followed by failures on systematic compositional tests. It supports the distinction Marcus describes in a controlled setting.
- Where
- Abstract, sentence 6
- Publisher
- Proceedings of Machine Learning Research
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Context only: Nature,
Our results show how a standard neural network architecture, optimized for its compositional skills, can mimic human systematic generalization in a head-to-head comparison.
Why it matters here: Meta-learning produces successful systematic generalization and improves performance on several compositional benchmarks. This limits any categorical reading of neural-network incapacity, but a successful exception does not refute the stated tendency.
- Publisher
- Nature
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, A knockout blow for LLMs?
neural networks of various kinds can generalize within a training distribution of data they are exposed to, but their generalizations tend to break down outside that distribution.
- Where
- paragraph 8
- Published
- 2025-06-07
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-a-knockout-blow-for-llms-68a208-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.