DUPLEX: Agentic Dual-System Planning via LLM-Driven Information Extraction

Researchers propose DUPLEX, a dual-system architecture that restricts LLMs to information extraction rather than end-to-end planning, using symbolic planners for logical synthesis. The system demonstrated superior performance across 12 planning domains by leveraging LLMs for semantic grounding while avoiding their hallucination tendencies in complex reasoning tasks.

• →DUPLEX confines LLMs to schema-guided information extraction rather than full planning or code generation to avoid hallucinations.
• →The system uses a fast lightweight LLM for entity extraction and a slow high-capacity LLM for complex scenario resolution.
• →Classical symbolic planners handle logical plan synthesis while LLMs focus on structured semantic grounding.
• →Extensive testing across 12 domains showed significant outperformance over existing end-to-end LLM baselines.
• →The research demonstrates that restricting LLMs to their strengths produces more reliable results than end-to-end approaches.