arXiv:2510.05141v2 Announce Type: replace Abstract: When we read, we make predictions about upcoming words; these predictions influence our reading behavior. The success of large language models (LLMs), which, like humans, make predictions about upcoming words, has motivated their use as models of human linguistic prediction. Surprisingly, in the last few years, as LLMs' ability to predict the next word has improved, their ability to explain reading behavior has declined. We argue this is because current LLMs can predict upcoming words much better than human readers can. This 'superhumanness'
The paper highlights a growing divergence between advanced LLM capabilities and their utility as direct models for human cognition, specifically linguistic prediction.
This research provides a nuanced perspective on LLM development, suggesting that pure predictive power might not equate to human-like intelligence or effective cognitive modeling, which is critical for AI alignment and interpretability.
The focus for some AI research may shift from merely improving next-word prediction accuracy towards developing LLMs that deliberately mimic human cognitive limitations to be better scientific models.
- · Cognitive science researchers
- · AI safety and alignment teams
- · Developers of more human-centric AI models
- · LLM developers focused solely on 'superhuman' prediction accuracy
- · Philosophers who equate LLM-level prediction to human understanding
There will be increased research into deliberately 'handicapping' LLMs to make them more human-like for cognitive modeling purposes.
This could lead to new architectures or training paradigms for LLMs that prioritize explainability and cognitive fidelity over raw performance.
A deeper understanding of human linguistic processing, informed by appropriately constrained AI models, might emerge, influencing future human-computer interaction designs.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.CL