A Graph-Constrained Agent Framework Lifts Retail Warehouse Requirement Success From 72-76% to 79%+
Retail supply chains run coupled decision modules that must adapt as requirements change, and existing LLM methods target individual optimization models rather than heterogeneous pipelines where one requirement admits several intervention paths with different downstream effects. The authors formulate requirement-driven adaptation as jointly selecting an intervention route and an admissible module-level change, with domain agents exposing reformulation interfaces and a central processor searching bounded intervention paths, validating candidates against downstream KPIs. Evaluated with a large retail partner on 100 warehouse requirements elicited from practitioner interviews, it improves correctness and raises end-to-end success from 72-76% to 79%+ across GPT, Qwen and DeepSeek backbones.
↳ Follow the thread