Learning the order of things: Modeling encoding- and retrieval-based strategies for transitive inference
Transitive inference (TI) is a fundamental form of reasoning whereby, after learning a set of premises (e.g., A < B, B < C), people infer the relationship between novel pairs of items (e.g., A < C). Existing computational models of TI differ on how premises are combined to support novel inferences: According to encoding-based models, people form a unified cognitive map of the hierarchy (e.g., A < B < C < D …) during training and directly compare items’ positions during inference, with faster, more accurate judgments for items that are more distant. Under retrieval-based models people retrieve and integrate premises at the time of test, but because distant inferences require the retrieval of more intervening premises, these models predict slower, less accurate judgments for more distant inferences. Previous studies have examined either encoding- or retrieval-based models, while little existing work has considered how the reliance on these strategies might differ across individuals, training conditions, or even for different judgments within the same task. The present study examined how the use of encoding- and retrieval-based TI depends on the difficulty of training, with more difficult training expected to interfere with the construction of a unified cognitive map and increase the reliance on retrieval-based inference. While there was little evidence of pure retrieval-based inference, more difficult training conditions were associated with increased use of a hybrid strategy such that people relied on an unified map for distant inferences, while resorting to more effortful premise retrieval for inferences about nearby items whose positions in the hierarchy were more uncertain. I present a novel approach for identifying this hybrid inference strategy using Bayesian hierarchical modeling, with models fit to both choices and RTs during inference. These findings suggest that individuals adaptively recruit direct premise memory to complement inferences supported by unified cognitive maps.