DUPLEX: Agentic Dual-System Planning via LLM-Driven Information Extraction
Summary: arXiv:2603.23909v1 Announce Type: new
Abstract
While Large Language Models (LLMs) provide semantic flexibility for robotic task planning, their susceptibility to hallucination and logical inconsistency limits their reliability in long-horizon domains. To bridge the gap between unstructured environments and rigorous plan synthesis, we propose DUPLEX, an agentic dual-system neuro-symbolic architecture that strictly confines the LLM to schema-guided information extraction rather than end-to-end planning or code generation.
Overview of the DUPLEX Framework
The DUPLEX framework incorporates two distinct systems to enhance the planning capabilities of LLMs:
- Fast System: This feed-forward system utilizes a lightweight LLM to extract relevant entities, relationships, and other critical information from natural language inputs. The extracted information is then deterministically mapped into a Planning Domain Definition Language (PDDL) problem file, which is compatible with classical symbolic planners.
- Slow System: This system is activated exclusively upon planning failures encountered by the Fast System. It leverages solver diagnostics to drive a high-capacity LLM through iterative reflection and repair processes, effectively resolving complex or underspecified scenarios.
Performance Evaluation
Extensive evaluations of the DUPLEX framework across 12 classical and household planning domains have shown promising results. The findings indicate that DUPLEX significantly outperforms existing end-to-end and hybrid LLM baselines in both success rates and reliability. This performance is attributed to the architectural design, which effectively separates the strengths of the LLM in structured semantic grounding from the logical plan synthesis handled by a symbolic planner.
Key Insights
The core insight of the DUPLEX framework is that the key is not to improve the LLM’s planning abilities but to restrict its role to areas where it excels. By confining the LLM to structured semantic extraction, the framework allows for more reliable and efficient planning outcomes. This shift in focus opens new avenues for integrating LLM capabilities in robotic task planning while addressing the challenges of hallucination and logical inconsistencies.
Conclusion
DUPLEX represents a significant advancement in the field of robotic task planning by combining the strengths of LLMs with traditional symbolic planning techniques. The dual-system architecture not only enhances the reliability of planning processes but also sets the stage for more sophisticated and adaptable robotic systems in the future. As research continues, the implications of DUPLEX could extend beyond robotics, potentially influencing various applications in artificial intelligence and automation.