一句话：提出 CoSTAR：用轻量工作流级控制组件维护下游 GRPO 的起始状态，并与下游 GRPO 交替优化，提升工业 LLM 工作流后训练适配。
工业软件工作流日益将 LLM 作为组件，但新工作流数据有限、任务级反馈稀疏，部署后适配困难。GRPO 后训练虽实用，却依赖适配起始状态；若初始行为与目标工作流对齐不佳，早期样本奖励信号弱，会浪费适配预算并延迟上线。为此提出 CoSTAR：用轻量工作流级控制组件维护下游 GRPO 的起始状态，仅对该控制组件做 GRPO 式优化而固定任务模型，并与下游 GRPO 交替执行，使控制状态跨阶段持续适配，而非一次性初始化。在公司内部工业工作流自动化任务上，CoSTAR 显著提升任务特定后训练，exact match 达 94.00%，复杂案例性能升至 80.56%。结果表明，维护下游 GRPO 起始状态可在有限反馈下改善适配。
Large language models (LLMs) are increasingly used as components in industrial software workflows, where they are expected to support reliable automation. However, newly introduced workflows often provide only limited data and sparse task-level feedback, making post-deployment adaptation difficult. GRPO-based post-training provides a practical adaptation way in this setting, but its effectiveness depends on the state from which adaptation begins. When the initial behavior of an LLM workflow is poorly aligned with the target workflow, early samples may provide weak reward signals, wasting adaptation budget and delaying workflow onboarding. We propose CoSTAR, an industrial post-deployment adaptation framework that maintains the starting state for downstream GRPO through a lightweight workflow-level control component. CoSTAR performs GRPO-style optimization on this control component while keeping the task model fixed, and alternates this maintenance with downstream GRPO, so that the control state can continue adapting across stages rather than remain a one-shot initialization. Experiments on a company-internal industrial workflow automation task show that CoSTAR substantially improves task-specific post-training, achieving 94.00% exact match and raising performance on complex cases to 80.56%. Overall, our results show that maintaining the starting state of downstream GRPO improves adaptation under limited workflow feedback and provides a more effective path for industrial LLM workflow maintenance.
@inproceedings{paperbot12418,
title = {Maintaining Control States for Initialization-Aware GRPO in Industrial LLM Workflows},
author = {Chen Liu and Zhenzhao Yuan and Fengyuan Fan and Limei Che and Qiao Qian and Qingyuan Liang and Zeyu Sun and Lu Zhang},
booktitle = {ASE},
year = {2026},
doi = {10.1145/3832783.3834459}
}