Accounting for Centaur’s pretraining in marginal-likelihood comparisons

Determine how Centaur’s pretraining should be accounted for when computing or comparing its marginal likelihood with the COntextual Retrieval and Encoding (CORE) model and other models.

Background

The paper compares CORE with Centaur, a foundation model of human cognition, using predictive negative log-likelihood. However, the authors do not compute Centaur’s marginal likelihood because the appropriate treatment of its pretraining is unresolved. This creates an open methodological issue for fair marginal-likelihood comparisons involving pretrained cognitive foundation models.

References

We did not compute the marginal likelihood for Centaur, because it is unclear how its pretraining should be accounted for.

— A small cognitive core explains human behavior at scale  (2610.13603 - Binz et al., 9 Oct 2026) in Materials and Methods, “Model comparison”