Learning to Decompose: Human-Like Subgoal Preferences Emerge in Neural Networks Learning Graph Traversal

Cognitive scientists have discovered normative and heuristic principles that capture human subgoal preferences when partitioning problems into smaller ones. However, it remains unclear where such preferences come from and why they tend to be both effective and efficient. In this work, we study the processes through which these preferences may be implicitly encoded over learning as learners improve towards optimal traversals. We build on the graph-based environments from prior work and use neural networks as model learners to test if learning shortest-path traversal can lead to human-like path decomposition. We find that simple transformer models develop a preference for paths containing nodes that occur frequently on the shortest paths, consistent with human subgoal preferences found in prior work. This preference is observed when models solve shortest path traversals for unseen problems in both known graphs and new graphs, demonstrating that human-like subgoal preferences can arise without requiring explicit preference computation or exhaustively searching over all possible paths. The same preference does not emerge when models learn to perform random or Hamiltonian traversals. Our findings are robust across several transformer variants as well as recurrent neural networks, suggesting they depend more on the data distribution than the network architecture.