Large language models (LLMs) often fail in planning when essential preconditions are not specified. They tend to generate non-existent facts or plans that violate hard constraints.

SQ-BCP: A New Approach

Self-Querying Bidirectional Categorical Planning (SQ-BCP) has been introduced. This method explicitly represents the status of preconditions (Sat/Viol/Unk) and resolves unknowns via:

  • Targeted queries to the user.
  • Bridging hypotheses that establish the missing condition through an additional action.

SQ-BCP performs bidirectional search and uses a pullback-based verifier as a categorical certificate of goal compatibility, using distance-based scores only for ranking and pruning.

Results

Across WikiHow and RecipeNLG tasks with withheld preconditions, SQ-BCP reduced resource-violation rates to 14.9% and 5.8%, compared to 26.0% and 15.7% for the best baseline.