Dreams are more "predictable" than you think.

Area of Science:

• Computational linguistics
• Dream research
• Artificial intelligence

Background:

• Machine learning and AI tools are increasingly used to analyze textual data, including dream reports.
• Concerns exist that AI models trained on web text may struggle with the unique nature of dream reports.

Purpose of the Study:

• To evaluate the suitability of large language models (LLMs) for encoding and predicting dream reports.
• To assess whether LLMs can capture known group differences within dream report data.

Main Methods:

• Employed a set of LLMs to encode dream reports from DreamBank and Wikipedia.
• Utilized perplexity as a metric to quantify how well LLMs model and predict textual sequences.

Main Results:

• Dream reports exhibited significantly lower perplexity scores compared to Wikipedia articles, indicating they are easier for LLMs to model.
• Perplexity scores differed significantly between genders, between blind and sighted individuals, and between clinical and healthy subjects.

Conclusions:

• LLMs are effective tools for analyzing dream reports, demonstrating a better capacity to model them than standard web text.
• LLMs implicitly encode demographic and clinical differences present in dream reports, aligning with existing research findings.