LLM+P is a framework that combines the natural language understanding capabilities of large language models with the formal guarantees of classical AI planners.¹⁾ Introduced by Liu et al., 2023, the approach uses an LLM to translate natural language problem descriptions into Planning Domain Definition Language (PDDL), which is then solved by an established planner such as Fast Downward. This hybrid architecture leverages the strengths of both paradigms: LLMs handle ambiguous natural language input while classical planners provide optimal and correct solutions for well-defined planning problems.

The core LLM+P pipeline operates in three stages:

1. Problem Translation: The LLM receives a natural language task description along with a PDDL domain template. It generates the PDDL problem file specifying the initial state, goal conditions, and relevant objects.
2. Classical Planning: An off-the-shelf planner (e.g., Fast Downward, LAMA, FF) solves the PDDL problem, producing a guaranteed-valid action sequence.
3. Plan Translation: The solution is converted back to natural language or executable actions for the user or downstream agent.

This separation of concerns ensures that the LLM handles what it excels at (language understanding, common sense, disambiguation) while the planner handles what it excels at (combinatorial search with correctness guarantees).

A 2025 survey by Tantakoun, Muise, and Zhu (ACL Findings 2025) reframes LLMs not as planners themselves but as planning formalizers that construct and iteratively refine PDDL models. Key contributions:

• LLMs generate initial PDDL domain and problem specifications from natural language
• Feedback from the planner (e.g., unsolvable problems, invalid actions) is used to iteratively correct the PDDL formulation
• This iterative loop addresses the brittleness of one-shot PDDL generation
• The approach scales to long-horizon problems where pure LLM planning degrades

Common planners used in LLM+P architectures:

• Fast Downward (Helmert, 2006): Supports multiple heuristic search algorithms; the most widely used in LLM+P research²⁾
• LAMA (Richter & Westphal, 2010): Landmark-based planner optimized for satisficing planning
• FF (Hoffmann & Nebel, 2001): Fast-forward planner using relaxed plan heuristics

The 2025 International Planning Competition evaluation tested frontier LLMs (DeepSeek R1, Gemini 2.5 Pro, GPT-5) directly against LAMA on standard IPC domains. While GPT-5 was competitive on standard domains, all LLMs degraded significantly on obfuscated variants where semantic cues were removed, confirming that pure LLM planning relies heavily on pattern matching rather than formal reasoning.

• Correctness Guarantees: Classical planners produce provably valid plans when given well-formed PDDL
• Optimality: Planners can find optimal or near-optimal solutions; LLMs tend to generate satisficing but suboptimal plans
• Scalability: Classical planners handle large state spaces through efficient heuristic search
• Interpretability: PDDL plans are human-readable formal specifications
• Robustness: Plans don't suffer from hallucination or reasoning errors once correctly formalized

The LLM+P paradigm has inspired several extensions:

• SayCan (Ahn et al., 2022, Google): Combines LLM semantic planning with affordance-based scoring to ground plans in physical robot capabilities.³⁾ The LLM proposes actions while a value function evaluates their feasibility.
• PaLM-E (Driess et al., 2023, Google): A 562B parameter embodied language model that processes visual and sensor inputs alongside text, enabling multimodal planning for robotics.⁴⁾
• Inner Monologue (Huang et al., 2022, Google): Uses LLM self-dialogue with environmental feedback for embodied task planning, incorporating success detection and scene description.⁵⁾
• Code as Policies (Liang et al., 2022): Generates executable Python code as plans, leveraging code interpreters as “planners” for spatial reasoning and control.⁶⁾
• MIT Optimization-LLM Integration (2025): Teaches LLMs optimization algorithms directly, enabling them to solve planning problems that combine discrete and continuous variables.

• PDDL Coverage: Many real-world problems resist clean PDDL formulation (partial observability, continuous dynamics, stochastic effects)
• Translation Errors: LLMs may generate syntactically valid but semantically incorrect PDDL, requiring iterative correction
• Domain Engineering: Creating PDDL domain templates still requires planning expertise
• Scalability of Translation: As problem complexity grows, accurate PDDL generation becomes harder for LLMs
• Obfuscation Sensitivity: LLMs struggle when domain descriptions are unfamiliar or abstracted away from natural patterns

Active research directions include automated PDDL domain learning, end-to-end differentiable planning, and integration with reinforcement learning for problems that resist pure symbolic formulation.

• RAP: Reasoning via Planning with LLM as World Model
• Memory Management for LLM Agents
• Agent-Based Task Management with LLMs
• AutoGPT
• Multimodal LLM