Larger LLMs Show Diminishing Returns in Predicting Human Cognitive Data

A new study published on arXiv suggests that larger language models (LLMs) do not necessarily improve predictions of human reading time and fMRI data when controlling for dimensionality expansion. The research found that the predictive power of trained LLMs over untrained counterparts diminishes significantly as model size increases, dropping to zero around a few billion parameters for most datasets. This challenges the notion that larger, more accurate LMs inherently better reflect human sentence processing architecture. AI

IMPACT Suggests that simply scaling up LLMs may not lead to better models of human cognition, potentially shifting research focus.

RANK_REASON Research paper published on arXiv detailing findings on LLM predictive capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Human Reading Time
• Language Models
• LLM Vector Predictors
• Trained LMs
• Untrained LMs
• Yi-Chien Lin

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →