Researchers have developed fourteen foundation models, ranging from 135 million to 14 billion parameters, trained on the Psych-101 dataset comprising 10.7 million trial-level choices from 160 experiments. The study found that while model scale had minimal impact on in-distribution performance, larger models demonstrated superior generalization to novel task structures out-of-distribution. Diagnostic tests revealed that models heavily rely on task instructions, experimental stimuli, and outcome feedback, with performance dropping significantly when these channels were masked. AI
IMPACT This research provides insights into the scaling properties and generalization capabilities of foundation models in cognitive science, potentially informing future model development.
RANK_REASON The cluster contains an academic paper detailing the training and evaluation of foundation models on a psychological dataset. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
