Thanks! As a preview of some ongoing work, we have since replicated this result in a simple visual working memory with item-memory only (i.e., no binding needed). See attached figures. This new experiment is essentially a follow up of He et al. (Exp. 1, 2026, JEP:LMC): osf.io/preprints/os...
![R code and output showing the new functionality:
``` r
## pak::pkg_install("quentingronau/bridgesampling#44")
## see: https://cran.r-project.org/web/packages/bridgesampling/vignettes/bridgesampling_example_stan.html
library(bridgesampling)
### generate data ###
set.seed(12345)
mu <- 0
tau2 <- 0.5
sigma2 <- 1
n <- 20
theta <- rnorm(n, mu, sqrt(tau2))
y <- rnorm(n, theta, sqrt(sigma2))
### set prior parameters ###
mu0 <- 0
tau20 <- 1
alpha <- 1
beta <- 1
stancodeH0 <- 'data {
int<lower=1> n; // number of observations
vector[n] y; // observations
real<lower=0> alpha;
real<lower=0> beta;
real<lower=0> sigma2;
}
parameters {
real<lower=0> tau2; // group-level variance
vector[n] theta; // participant effects
}
model {
target += inv_gamma_lpdf(tau2 | alpha, beta);
target += normal_lpdf(theta | 0, sqrt(tau2));
target += normal_lpdf(y | theta, sqrt(sigma2));
}
'
tf <- withr::local_tempfile(fileext = ".stan")
writeLines(stancodeH0, tf)
mod <- cmdstanr::cmdstan_model(tf, quiet = TRUE, force_recompile = TRUE)
fitH0 <- mod$sample(
data = list(y = y, n = n,
alpha = alpha,
beta = beta,
sigma2 = sigma2),
seed = 202,
chains = 4,
parallel_chains = 4,
iter_warmup = 1000,
iter_sampling = 50000,
refresh = 0
)
#> Running MCMC with 4 parallel chains...
#>
#> Chain 3 finished in 0.8 seconds.
#> Chain 2 finished in 0.8 seconds.
#> Chain 4 finished in 0.8 seconds.
#> Chain 1 finished in 1.1 seconds.
#>
#> All 4 chains finished successfully.
#> Mean chain execution time: 0.9 seconds.
#> Total execution time: 1.2 seconds.
H0.bridge <- bridge_sampler(fitH0, silent = TRUE)
print(H0.bridge)
#> Bridge sampling estimate of the log marginal likelihood: -37.73301
#> Estimate obtained in 8 iteration(s) via method "normal".
#### Expected output:
## Bridge sampling estimate of the log marginal likelihood: -37.53183
## Estimate obtained in 5 iteration(s) via method "normal".
```](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:lgxrkw556dtxkplwrj5rzlli/bafkreicf6zdyyjmaxrdfyq2sekokprym3c2o5atca6i2cczcx5zec4q4hm)
![Abstract
Establishing a unified theory of cognition has been a major goal of psychology [1, 2]. While there have been previous attempts to instantiate such theories by building computational models [1, 2], we currently do not have one model that captures the human mind in its entirety. Here we introduce Centaur, a computational model that can predict and simulate human behavior in any experiment expressible in natural language. We derived Centaur by finetuning a state-of-theart language model on a novel, large-scale data set called Psych-101. Psych-101 reaches an unprecedented scale, covering trial-by-trial data from over 60,000 participants performing over 10,000,000 choices in 160 experiments. Centaur not only captures the behavior of held-out participants better than existing cognitive models, but also generalizes to new cover stories, structural task modifications, and entirely new domains. Furthermore, we find that the model’s internal representations become more aligned with human neural activity after finetuning. Taken together, Centaur is the first real candidate for a unified model of human cognition. We anticipate that it will have a disruptive impact on the cognitive sciences, challenging the existing paradigm for developing computational models.
Keywords: cognitive science, cognitive modeling, unified theory of cognition, large language models](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:lgxrkw556dtxkplwrj5rzlli/bafkreifeef77udzg5p64o26brzpym43ykjuxqotv5twwkcre3qsl53uncu)
