Human Causal Reasoning as Finite-State Dynamics: Implications for Neurosymbolic AI

Artificial systems continue to struggle where human cognition excels: in causal inference, learning from sparse data, and flexible problem-solving. Human causal reasoning integrates symbolic constraint satisfaction with graded probabilistic inference, making human cognition a natural architectural model for neurosymbolic AI. To investigate this, 40 participants solved two instruction-less puzzle benchmarks with five levels each. Verbal reasoning traces were elicited using a think-aloud protocol; keystrokes, timestamps, and game states were logged as participants navigated the puzzles. Through grounded theory analysis of behavioral observations, verbal markers were organized into a taxonomy of knowledge acquisition using natural language processing for feature extraction and classification. Findings reveal the structured inferential processes of human learning, challenging the sufficiency of the traditional explore-exploit framework and motivating neurosymbolic architectures capable of genuine causal learning in sparse-reward, exploration-driven environments. We introduce novel sparse-reward puzzle benchmarks, a validated dataset of corresponding human solutions, and labeled transition systems derived from these data, using theory-based causal induction as a semi-formalized framework amenable to neurosymbolic implementation.