NeuroSymbolic Knowledge-Grounded Planning and Reasoning in Artificial Intelligence Systems

Why it matters

• Decision-support systems in domains such as health care require multistep reasoning, constraint enforcement, and regulatory compliance.
• LLMs generate coherent language but struggle with structured search, logical verification, and protocol adherence.
• Implicit knowledge representations prevent explicit state tracking, rule enforcement, and safety validation.
• High-stakes environments require interpretable, constraint-aware, and dynamically adaptable planning, which standalone LLMs do not guarantee.

What we did

• Proposed a neurosymbolic framework integrating domain-adapted LLMs with knowledge graphs (KGs), symbolic reasoning, and constraint-aware planning.
• Used the LLM to generate initial candidate plans in natural language aligned to domain context.
• Transformed LLM outputs into structured representations compatible with KG-backed reasoning modules.
• Applied deductive reasoning for logical consistency and abductive reasoning to resolve incomplete or conflicting constraints.
• Illustrated the approach through a health-care MTSS use case involving a 16-year-old student subject to protocol and budget constraints.

Figure 1 shows how the LLM, symbolic modules, and knowledge graphs interact across grounding, reasoning, and planning layers.

How it works

• Domain-adapted LLM interprets user queries and generates candidate plans in natural language.
• Preprocessing interface converts textual outputs into structured actions and conditions aligned with KG entities.
• Constraint manager and safety checker validate compliance with domain rules and flag violations.
• Symbolic inference engine applies deductive reasoning for verification and abductive reasoning for resolving gaps.
• Temporal and resource reasoner + optimizer ensure feasibility and iteratively refine plans using past history.

Key contributions

• Introduces a layered neurosymbolic architecture organized around knowledge grounding, reasoning integration, and dynamic planning.
• Formalizes the integration of LLM-generated plans with KG-backed constraint checking and logical verification.
• Demonstrates deductive and abductive reasoning over structured domain knowledge to resolve conflicts (for example, protocol versus budget).
• Illustrates an iterative feedback loop between LLM and symbolic modules to ensure policy-compliant, executable plans.