Structured Prompt Optimization Meets Reinforcement Learning for Global and Local Interpretability over Complex Text

A new text classifier called eXTC aims to resolve a longstanding trade-off in language-model classification by pairing fast inference with both local reasoning traces and a global rulebook, according to research submitted on 27 May 2026 [1][2]. The work, posted to arXiv, addresses a divide in how large language models (LLMs) are applied to text classification. Supervised fine-tuning using only labels is scalable but offers limited reasoning on complex text and lacks broader model transparency, while discrete prompt optimization yields human-readable instructions yet struggles with performance and scalability [1][2]. LLMs, which are neural networks trained on vast text corpora for tasks such as generation, summarization, and analysis, have become foundational to modern chatbots and other generative AI tools [3][4]. eXTC — short for eXplainable Text Classifier — is built around three progressive stages. First, a Standard Operating Procedure, or rulebook, is learned in natural language through a new Structured Prompt Optimization algorithm. Second, reasoning grounded in that SOP is distilled from a large teacher LLM into a compact language model. Third, the system’s reasoning capabilities are expanded beyond the initial SOP using reinforcement learning [1][2]. The design choices target two practical demands. A compact language model enables fast inference, while the architecture supplies inference-time local reasoning traces and a global, modular explanation of the domain rules it has acquired. The authors report that eXTC significantly outperforms existing paradigms across diverse benchmarks in both classification performance and explanation quality, with measurable gains at each stage [1][2]. Generative AI, which encompasses LLMs and related transformer-based models, has seen rapid adoption since the AI boom of the 2020s, spreading into software development, healthcare, finance, and customer service [4]. As deployment widens, concerns about transparency, bias from training data, and factual reliability have drawn scrutiny from researchers and regulators [3][4]. The eXTC framework enters this landscape as an attempt to make classification decisions more auditable without sacrificing the speed that production systems require. The paper was submitted under the Computation and Language category on arXiv and is associated with arXivLabs, a framework that lets community collaborators develop and share experimental features on the platform [1].