ROAM Adapts Industrial Models to New Scenarios Without Retraining

Process industries often rely on validated specialist models, but these models can degrade over time due to sensor drift, feedstock variations, or regime switching in new operational scenarios. Retraining models with new labeled data is costly and time-consuming, while continuing with outdated models leads to persistent bias. Existing adaptation methods typically require modifying model parameters with sufficient labeled data, making rapid responses difficult for deployed systems. This article introduces Reasoning-Driven Open Adaptation for Specialist Models (ROAM), a novel framework that leverages Large Language Model (LLM) world knowledge and reasoning capabilities to adapt frozen specialist models to previously unseen scenarios without the need for retraining. ROAM restricts all corrections to a low-dimensional, semantically interpretable latent space. The framework integrates LLM-generated scenario judgments with online observations within a unified probabilistic framework. A risk-constrained mechanism is included to suppress corrections when LLM evidence is unreliable or during abrupt scenario shifts, allowing the system to revert to the original frozen model if evidence is insufficient. Experiments on a mineral thickening process and a public penicillin fermentation dataset demonstrate that ROAM reduces Mean Absolute Error (MAE) by over 20% in major shift settings, using minimal additional parameters and negligible per-step overhead. This indicates that LLM reasoning can provide a conservative yet effective adaptation signal for industrial models already in service.