Reachability Is Not Generalization: Understanding Verb--Noun Decomposition in Assembly Action Recognition
The paper investigates why decomposing assembly actions into verbs and nouns allows for predicting unseen actions but fails to generalize well. While decomposition makes unseen compositions reachable, success depends heavily on training data structure, specifically vocabulary asymmetry and component entanglement. It suggests current methods rely more on reusing observed co-occurrences than true compositional reasoning.