LLMs Adapt Industrial Specialist Models to New Scenarios Without Retraining.

Process industries frequently encounter issues where their validated specialist models degrade due to factors like sensor drift or feedstock variations, requiring costly retraining or leading to persistent biases. Traditional adaptation methods often demand significant labeled data or parameter modifications, making rapid responses difficult. This research introduces ROAM (Reasoning-Driven Open Adaptation for Specialist Models), a novel framework designed to address these challenges. ROAM leverages the world knowledge and reasoning capabilities of Large Language Models (LLMs) to adapt pre-existing, frozen specialist models to new, unseen scenarios. Crucially, it avoids retraining the original model by applying all corrections within a low-dimensional, semantically interpretable latent space. The framework fuses LLM-generated scenario judgments with online observations under a probabilistic model, incorporating a risk-constrained mechanism to suppress unreliable corrections. This approach significantly reduces prediction errors in major shift settings, as demonstrated on mineral thickening and penicillin fermentation datasets, with minimal additional parameters and negligible online inference overhead.