Jessica Golding¹, Tianye Ma², Weiwei Zhang³; ¹University of California Riverside

Schurgin et al. (2020) developed the Target Confusability Competition (TCC) model, a uni-dimensional signal detection model for short-term memory (STM) recall. Formal model comparisons, using goodness-of-fit indices like Bayesian Information Criterion (BIC), favored the TCC model over capacity-limited models. such as those incorporating a pure-guess component. Nevertheless, the validity of this computational model selection method in inferring the capacity limit in STM is unclear. To address this, the current study evaluated the validity of model selection by replicating Schurgin et al.’s analysis on simulated data from a process model explicitly incorporating a capacity limit. Simulations were conducted using a revised version of the Binding Pool model (Swan & Wyble, 2013), with population code decoding as the response mechanism. The TCC model was compared to two capacity-limited models: the Slot model and the TCC with capacity (TCCk) model, both of which were included in the original analysis by Schurgin et al. (2020). Results indicated that parameters from the TCC and Slot models, but not the TCCk model, successfully tracked parameter changes in the process model. However, the TCC model consistently outperformed the other models across a wide range of parameters in the Binding Pool simulations. Importantly, model selection failed to detect the capacity limit embedded in the process model, suggesting that computational model selection alone is insufficient for inferring underlying memory mechanisms. These findings highlight the risk of over-interpreting good model fits without careful theoretical validation.