The ability to track non-adjacent dependencies (the relationship between ai and bi in an aiXbi string) has been hypothesized to support detection of morpho-syntactic dependencies in natural languages (‘The princess reluctantly kiss the frog’). But tracking such dependencies in natural languages entails being able to generalize dependencies to novel contexts (‘The general angrily berat his troops’), and also tracking co-occurrence patterns between functional morphemes like and (a class of elements that often lack perceptual salience). We use the Headturn Preference Procedure to investigate (i) whether infants are capable of generalizing dependencies to novel contexts, and (ii) whether they can track dependencies between perceptually non-salient elements in an artificial grammar aXb. Results suggest that 18-month-olds extract abstract knowledge of a_b dependencies between non-salient a and b elements and use this knowledge to subsequently re-familiarize themselves with specific ai_bi co)