A neurosymbolic pipeline that delegates systematic reasoning about actions from a large language model to a symbolic action-language reasoner. The LLM acts only as a multi-stage translator -- generating a program signature, extracting commonsense and domain knowledge, converting it into rules of the action language BC+, and revising them from the reasoner's feedback -- while the BC+ reasoner (compiling to answer set programming) performs the search over states, indirect effects, and defaults that standalone LLMs cannot do reliably. On McCarthy's Missionaries-and-Cannibals elaboration-tolerance benchmark it reaches correct answers with a few human corrections where leading LLMs fail even with feedback.
A neurosymbolic method, LLM+AL, for complex reasoning about actions -- task planning, temporal prediction (what follows from a sequence of actions), and postdiction (inferring an earlier state) -- th…