An Adaptive Prompt Optimization Framework for Domain-Specific Large Language Models

Domain-specific deployment of large language models (LLMs) remains constrained by prompt brittleness, inference cost, and uneven generalization across task subtypes. This paper presents APOF (Adaptive Prompt Optimization Framework), a closed-loop framework that jointly opti- mizes prompt structure, retrieval context, and inference-time control signals for domain-specific LLM applications. APOF combines three elements: (i) a policy-guided prompt composer that dynamically allocates instruction budget across task facets, (ii) a critic model that estimates prompt-task fitness before expensive decoding, and (iii) an online adaptation module that up- dates prompt policies using delayed feedback from production outcomes. We instantiate APOF in three high-stakes domains—clinical note summarization, legal clause risk classification, and materials-science question answering—using a shared 13B parameter base model and domain adapters.Our experiments include 162,000 annotated instances across public and institutionally cu- rated corpora, 12 baseline methods, and controlled ablation studies. APOF improves macro-F1 by up to 6.8 points over the strongest static prompt baseline, while reducing median latency by 18.4% through pre-decoding prompt pruning and adaptive generation parameters. The frame- work also improves calibration (ECE reduction of 0.041) and demonstrates higher robustness under distribution shift (average relative performance drop 12.3% vs. 21.7% for static meth- ods). We provide mathematical formulations, complexity analysis, and practical deployment recommendations. Results suggest that adaptive prompting, when treated as a structured op- timization problem rather than manual engineering, is a viable path to reliable domain-specific LLM systems.