LLM-Guided Contextual Action Evaluation for Operational Decisions in Industrial Processes
August 25, 2026·
,,
Youcheng Zong
Runda Jia
Dakuo He
LCAE overview: offline semantic construction, history-conditioned relation gains, and action-effect composition.Abstract
LLM-Guided Contextual Action Evaluation (LCAE) uses industrial documents to construct a semantic representation of actions for operational decisions. Before training, an LLM extracts a fixed basis describing which observations an action affects, the effect direction, and the response delay. Recent numerical action-response history adjusts relation strengths, and candidate actions are represented through a state-conditioned action-effect field. The critic evaluates this representation, while the actor uses the same relation gains to generate actions. Training and deployment require no online LLM or embedding-model calls. The approach investigates whether documented action relations, conditioned on recent process history, provide a more useful decision bias than raw action coordinates.
Type
Publication
arXiv preprint arXiv:2608.24156
Authors
PhD Candidate
Youcheng Zong is a PhD candidate in Control Science and Engineering at Northeastern University, China. He develops LLM-based methods for industrial forecasting, measurement correction, model adaptation, and decision support. He also studies structure-aware visual analysis for medical imaging.
