LLM-Guided Task-Semantic Field Factorization for Industrial Process Forecasting

Short summary

Researchers propose Task-Semantic Field Factorization (TSF), a framework that uses LLMs offline to extract semantic relationships between process variables and prediction targets from industrial documentation, then injects that semantics into conventional time-series backbones during training and inference. TSF reduces mean absolute error by 6.4% on average across multiple industrial forecasting and soft-sensing benchmarks, with up to 25.5% reduction in best cases, while adding only ~1.8–3.0k parameters and negligible inference overhead. The approach enables existing process documents to improve forecasting without costly pipeline rebuilds across changing operating regimes.

• •TSF uses LLMs offline to build task-semantic fields from process documents and variable tables
• •Reduces MAE by 6.4% average (up to 25.5%) across industrial forecasting tasks with minimal parameter and latency overhead
• •Semantics are activated per numerical window, enabling adaptation to shifting operating regimes and different prediction targets

Generated with AI, which can make mistakes.