Title: Integrating Structured Prompt Optimization with Reinforcement Learning to Achieve Global and Local Interpretability in Complex Text Analysis

Abstract:

While Large Language Models (LLMs) have significantly progressed text classification tasks, current methodologies are constrained by a fundamental compromise. Supervised fine-tuning, which relies solely on labels, scales efficiently but provides limited reasoning capacity for intricate texts and offers little transparency regarding the model’s inner workings. Conversely, discrete prompt optimization delivers human-readable directives but often fails to match the performance and scalability required for robust applications. To address these limitations, we present eXTC (eXplainable Text Classifier), a framework built upon three sequential phases. First, we employ a novel Structured Prompt Optimization algorithm to acquire a Standard Operating Procedure (SOP)—essentially a natural language rulebook. Second, we distill reasoning grounded in this SOP from a large teacher LLM into a more compact language model. Finally, we utilize reinforcement learning to extend the model’s reasoning capabilities beyond the boundaries of the initial SOP. This architecture allows eXTC to deliver rapid inference through a compact model while providing both local reasoning traces during inference and a global, modular breakdown of its learned domain rules. Empirical results demonstrate that eXTC significantly surpasses existing approaches across various benchmarks, achieving superior outcomes in both classification accuracy and explanation quality, with measurable improvements at each stage of the process.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC