The Challenge of Long-Horizon Planning

Large Language Models often struggle with long-horizon decision-making, frequently generating infeasible or logically incorrect plans. These failures stem from the model's inability to maintain strict adherence to task constraints and semantics over extended sequences. The proposed framework addresses this by moving away from pure natural language generation toward a structured, feedback-driven refinement loop.

Symbolic Feedback and Iterative Refinement

The core of the framework is a three-part mechanism designed to enforce reliability:

• Symbolic Mapping: The system maps logical symbols into natural language descriptions. This bridges the gap between formal constraints and the model's linguistic capabilities, ensuring the LLM understands the underlying task requirements.
• Symbolic Verifier: Instead of relying on the LLM to self-correct based on vague intuition, a dedicated verifier checks the generated plan against formal constraints. When errors are detected, the verifier translates these failures into specific, actionable corrective instructions that the LLM can interpret and execute.
• Plan Recognizer: To prevent the model from drifting, a plan recognizer evaluates goal reachability. This provides a signal that guides the refinement process toward valid, achievable end-states.

By combining these components, the framework enables an iterative self-refinement loop where the LLM continuously updates its plan based on concrete, symbolic feedback rather than probabilistic guessing. This approach consistently improves both the feasibility and correctness of plans in complex, multi-step tasks.