CoRe: A Collaborative and Reflective Language Model Framework for Flexible Job Shop Scheduling

Abstract

Effective production scheduling is pivotal to intelligent manufacturing, where Mixed-Integer Linear Programming (MILP)-based exact methods remain attractive for their guarantee of global optimality. Nevertheless, their deployment is constrained by the specialized expertise required for model formulation, particularly in Flexible Job-shop Scheduling Problems (FJSP), where distinct production lines operate under unique constraints that necessitate manual model reconstruction for each scenario. Large Language Models (LLMs) offer a promising avenue to automate this process, yet their industrial adoption faces two coupled challenges: transmitting sensitive production parameters to cloud-based LLMs raises data exposure concerns, while locally deployable Small Language Models (SLMs) lack the reasoning capacity for complex constraint modeling; moreover, LLM-generated formulations are prone to implicit errors that execute successfully yet yield physically infeasible schedules. To address these challenges, this paper proposes CoRe, a collaborative and reflective framework built upon an “SLM-foundational, LLM-augmented” paradigm. Specifically, a local SLM is employed to abstract constraint logic and accomplish the majority of modeling tasks, while a confidence-based routing mechanism selectively delegates only symbolically abstracted, mathematically complex components to an auxiliary cloud LLM, thereby balancing privacy preservation with reasoning capability. Building upon this, a deterministic programmatic verification module coupled with a hierarchical reflective repair strategy is developed to autonomously detect and rectify these implicit errors, ensuring physical feasibility prior to shop-floor execution. Extensive experiments on 140 heterogeneous FJSP instances demonstrate that CoRe attains an overall accuracy of 81.43%, outperforming baseline architectures by more than 17 percentage points, while cloud LLM tokens account for only 15.1% of per-instance token usage among LLM-invoked cases. A case study on aircraft skin fabrication validates its practical applicability. Note to Practitioners—This work is motivated by a recurring difficulty on modern shop floors: whenever a production line changes its products, machines, or operating rules, the underlying scheduling model must be rebuilt by an operations research specialist, a process that is slow, costly, and hard to scale across heterogeneous lines. The proposed CoRe framework allows engineers to describe a workshop and its constraints in plain text, and then automatically returns an executable production schedule, without requiring expertise in mathematical modeling. To make this practical in real factories, most of the modeling is carried out by a lightweight model running on a local workstation, so that sensitive production data need not leave the plant, while an external service is consulted only for a small, carefully isolated portion of the task. Every generated schedule is automatically checked against real-world feasibility rules and corrected if necessary before being released to the floor. CoRe is therefore suitable for small- and medium-sized manufacturers operating heterogeneous or frequently reconfigured lines, where hiring dedicated optimization specialists for each scenario is not economically viable.