The Core Problem: Planning Fragility

Large language models often struggle with long-horizon decision-making, frequently generating infeasible or logically incorrect plans. This failure stems from the model's inability to consistently track complex task constraints and maintain semantic consistency over extended sequences. The authors argue that standard prompting is insufficient for these tasks, necessitating a more structured, feedback-driven approach.

The Symbolic Feedback-Driven Framework

The proposed framework improves reliability by integrating symbolic reasoning with natural language generation through three primary components:

• Symbolic-to-Natural Language Mapping: The system maps logical symbols into descriptive natural language. This allows the LLM to process task constraints using its native linguistic strengths while maintaining the rigor of symbolic logic.
• Symbolic Verifier: This component acts as a guardrail, identifying logical errors in the generated plan. Instead of simply rejecting the plan, it converts these errors into specific, corrective instructions that the LLM can interpret and act upon.
• Plan Recognizer: This module evaluates goal reachability, providing the LLM with a signal on whether its current trajectory is likely to succeed. This facilitates more effective guidance, allowing the model to pivot or refine its strategy before reaching a dead end.

Iterative Self-Refinement

The framework operates as an iterative loop. By combining the verifier's error detection with the plan recognizer's reachability assessment, the model undergoes multiple rounds of self-correction. Empirical results indicate that this approach significantly increases both the feasibility and correctness of plans in long-horizon scenarios, moving LLMs closer to being reliable agents for complex, multi-step tasks.