New Study Compares Human Reading to LLM Word Prediction
TL;DR. A new study by NYU and UMass Amherst researchers found similarities and differences in how humans and LLMs process text during reading. - Humans and AI use next-word prediction during initial reading, as seen in eye-tracking data. - Human processing diverges from AI in complex passages, especially when rereading is required. - LLMs explain simple word recognition but not the more complex integration of words into context.
- Humans and LLMs both rely on next-word prediction for early language processing during reading.
- Human reading diverges from LLM processing when passages become complex or require rereading.
- LLMs can account for simple word recognition but not for the intricate cognitive processes behind deeper comprehension and integration of words into context.
Sources
- Why do we labor when reading some words but not others? AI offers a partial answer — techxplore.com
- viterbischool.usc.edu — viterbischool.usc.edu