MARS: Multi-Agent Rescheduling Framework with Large Language Models for Human-Centric Manufacturing

Authors: Jimin Park, Karen Seojin Kim, Zuhdi Faturrahman, and Hyun-Jung Kim Venue: Journal of Manufacturing Systems (JMS), 2026 — Accepted (volume/DOI to follow)

TL;DR

On the shop floor, operators constantly tweak algorithm-generated schedules using tacit knowledge — but doing it through Gantt charts and spreadsheets is time-cconsuming and labor-intensive. MARS lets an operator just type what they want in natural language; it edits the schedule with code and explains what changed and why, using lightweight open source LLMs that run on-premise.

The problem: two hidden gaps

Manually adjusting a schedule runs into two gaps:

• Intent–execution gap — one high-level intent (“what if J3 on M3 is swapped with J6 on M4?”) has to be broken down into many low-level edits: find the jobs, drag them, fix affected jobs, re-run a repair.
• Change–comprehension gap — after the edit, a single change cascades through the schedule, and the operator must manually compare the old and new schedules to understand the downstream effects.

What MARS does

MARS is a multi-agent framework with two cooperating LLM agents:

• Schedule Modifier — turns the natural-language request into executable Python that directly updates the scheduling parameters or the schedule itself. Because it edits parameters/schedules rather than rewriting the problem as solver constraints, it stays compatible with any scheduling algorithm. Internally it uses a scheduler re-run classifier, a modification type classifier, and a modification code generator, built on a parameter-efficient multi-adapter (QLoRA) architecture with a level-wise asymmetric in-context-learning (ICL) strategy.
• Explainer — builds a tree-based causal chain by comparing the original and modified schedules, then verbalizes a faithful, step-by-step natural language explanation of the cascading changes.

Why it’s different

• One natural language interface that bridges both gaps at once.
• Lightweight & deployable — runs on small open-source models, avoiding the cost, latency, and data-privacy problems of large closed-source APIs in computationally constrained factory environments.
• Scheduling algorithm-agnostic — modifies parameters/schedules directly instead of forcing changes into rigid solver constraints.

Results

Evaluated on the nrelated parallel machine scheduling problem (UPMSP) with comprehensive constraints, MARS accurately executes user requests, produces transparent explanations of schedule changes, and generalizes to larger problem sizes.