Explain the performance gap between LLMs and humans in interpreting Jabberwockified English

Determine the principal factors that best explain why large language models outperform human readers at recovering meaning from Jabberwockified English, specifically whether the advantage arises from LLMs learning more complex or abstract morphosyntactic patterns due to vastly greater training exposure, or from more effective use of largely similar patterns learned by both humans and LLMs.

Background

The authors argue that both humans and LLMs rely on pattern-matching to interpret language, and show that LLMs can recover meaning from heavily degraded, Jabberwockified texts to a degree that appears to exceed typical human capability.

They explicitly state that the reason for this difference is presently unknown, proposing two candidate explanations: LLMs may have learned more complex or abstract morphosyntactic patterns due to vast training data, or they may be leveraging similar patterns more effectively than humans.

References

We do not yet know what best explains this difference in ability.

— The unreasonable effectiveness of pattern matching  (2601.11432 - Lupyan et al., 16 Jan 2026) in Section 3: Language as a set of patterns; near the discussion of human costs and limits in reading scrambled or degraded text

First, as discussed above, the training-data account cannot be ruled out: the observed alignment gap may reflect differences in training-data coverage as much as differences in sublexical processing, and replication in higher-resource languages would help disentangle the two.

— Pseudowords as probes: Large Language Models show little of the sublexical sensitivity that governs human pseudoword processing  (2610.07936 - Chen et al., 6 Oct 2026) in Limitations section