The paper proposes a framework that combines the advantages of large language models (LLMs) and satisfiability modulo theory (SMT) solvers to tackle the complex travel planning problem.
The key components of the framework are:
NL-JSON Translation: The LLM translates the natural language input into a JSON format that describes the travel planning problem.
JSON-Step Generation: The LLM generates steps to formulate the travel planning problem as an SMT problem based on the JSON description.
Step-Code Generation: The LLM generates the corresponding Python code to encode the problem and call the SMT solver.
SMT Solver: The framework executes the generated code to encode the problem and call the SMT solver. The SMT solver guarantees to find a solution if the constraints are satisfiable, and provides unsatisfiable reasons if the constraints cannot be satisfied.
Interactive Plan Repair: If the initial query is not satisfiable, the framework leverages the LLM's reasoning capabilities to analyze the unsatisfiable reasons, collect more information, and provide suggestions to the user to modify the constraints. The framework then updates the code based on the user's feedback and calls the SMT solver again to find a feasible plan.
The framework is evaluated on two datasets: the TravelPlanner benchmark and the authors' own UnsatChristmas dataset, which introduces new constraint types not present in TravelPlanner. The results show that the framework can reliably handle diverse human inputs, deliver formally-verified plans, and interactively modify unsatisfiable queries based on user preferences, outperforming baseline methods. The framework also demonstrates the ability to generalize to unseen constraint types.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Yilun Hao,Yo... at arxiv.org 04-19-2024
https://arxiv.org/pdf/2404.11891.pdfDeeper Inquiries