new preprint on the use of #LLMs as cognitive models, that is, as models of human cognition:
pretraining Arithmetic-GTP with environmentally plausible distributions of probabilities and values leads to embeddings for expected values that in turn can be used to predict between 65% and 95% of human choices in 4 extant datasets of risky choice and inter-temporal choice
1/2
@cogsci 2/2
these results are interesting both for the current debate on the possibility/value of using LLMs for cognitive modelling and for to the long-standing substantive issue of why human choices deviate systematically from rational prescription in that they fit with Stewart, Chater and Brown's (2006) "decision by sampling" account which invokes the shape of real world distributions of probability and value in conj. with the idea that humans are sensitive (only) to rank order information
Ulrike Hahn