Rats employ a task general strategy to report calibrated confidence during learning

Humans can flexibly monitor their own data processing, and unlike modern AI algorithms appropriately adjust their confidence estimates for decisions in new contexts. Here, we develop a behavioral task to evaluate whether rats, like humans, can calibrate their confidence reports while learning new stimuli-response associations. We compare rat behavior to a self-reflective (metacognitive) model that uses statistical inference to produce calibrated confidence reports. Consistent with model predictions, the variance and calibration of rat confidence increases as their behavioral performance improves. These results suggested rats might employ a flexible, task general second stage of processing – e.g. metacognitive to generate confidence reports. Consistent with this conclusion, Orbitofrontal Cortex (OFC) confidence-encoding neurons were stable across novel stimulus learning. These behavioral and neural correlates of confidence reports support the conclusion that the sense of confidence in rats is a separable metacognitive module that can be flexibly applied in novel situations.