Towards Reliable and Robust LLM Planning: Symbolic Feedback-Driven Iterative Self-Refinement Framework
06:00 · June 29, 2026 · arXiv cs.AI RSS
Large language models (LLMs) have attracted widespread attention from academia and industry, yet their deployment raises critical security concerns regarding robustness and reliability. Planning, a core component of intelligent behavior, remains challenging for LLMs, which often produce infeasible or incorrect solutions in long-horizon decision-making tasks due to inherent complexity. In this paper, we propose a symbolic feedback-driven iterative self-refinement framework to enhance the robustness and reliability of LLMs in long-horizon planning. Specifically, a natural language prompting mechanism is introduced to map logical symbols into natural language descriptions, enabling LLMs to better capture task constraints and semantics. We further design a symbolic verifier that identifies errors and converts them into corrective instructions interpretable by the LLM, thereby guiding self-refinement. In addition, we leverage a plan recognizer to infer goal reachability, facilitating more effective guidance toward desired goals. Empirical results demonstrate that the proposed framework consistently improves both feasibility and correctness in long-horizon planning tasks. This highlights its effectiveness in enhancing the reliability of LLM-based planning and potential to enable more trustworthy AI systems.
Summary
Large language models continue to struggle with long-horizon planning, where sequences of interdependent actions must be constructed over many steps while respecting constraints and avoiding the accumulation of early errors. The paper attributes these shortcomings to the models’ tendency to produce locally plausible but globally infeasible or incorrect plans, a limitation that raises reliability concerns for applications requiring verifiable decision sequences.
To address this gap, the authors present a symbolic feedback-driven iterative self-refinement framework that integrates structured symbolic reasoning with the language capabilities of LLMs. A natural-language prompting layer first translates PDDL-style logical symbols into readable descriptions, allowing the model to interpret task constraints and semantics without direct exposure to formal syntax. A symbolic verifier then evaluates candidate plans for violations such as action conflicts or unmet preconditions, converting detected errors into corrective natural-language instructions that the LLM can use to revise its output. Complementing this loop, a plan recognizer assesses whether the current trajectory can still reach the intended goal, supplying an additional signal that steers refinement toward reachable states.
The resulting iterative process repeatedly feeds symbolic feedback back into the LLM until the plan satisfies both feasibility and goal conditions. Experiments across standard planning domains indicate that the framework consistently raises both the feasibility and correctness of generated plans relative to unassisted LLM baselines, demonstrating a practical route to more reliable LLM-based planning without requiring fully manual symbolic encodings.
Why it matters
This research directly supports the Dutch and EU focus on trustworthy and reliable AI by addressing the critical robustness concerns of LLM deployments. The proposed symbolic verification framework offers advanced researchers actionable methodologies to build more transparent and dependable AI planning systems.
