Abstract
The present disclosure provides a system for adaptive multi-format prompt preprocessing using a unified reinforcement learning framework. The system includes a prompt-type classifier (116) configured to identify a format type of an input prompt, wherein the format type includes one of natural language text, code, log, hybrid content, or error trace, the prompt-type classifier (116) being configured in some embodiments to perform deterministic rule-based classification. The system further includes a feature extraction layer (118) configured to produce a multi-modal representation of the input prompt capturing structural, linguistic, and semantic signals. The system also includes a hierarchical reinforcement learning policy engine (120) configured to learn format-specific noise-removal pipelines based on the identified format type and the multi-modal representation. The system additionally includes an action library (122) storing format-aware preprocessing operations selectable by the hierarchical reinforcement learning policy engine (120), a cached pipeline repository (124) configured to store learned preprocessing pipelines, and an output generator (126) configured to produce cleaned prompts for forwarding to a large language model service (108), wherein the learned preprocessing pipelines may be applied deterministically to subsequent prompts to provide predictable behavior with reduced runtime overhead.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kolluru, Katyayani Kiranmayee and Mohan, Shivam, "ADAPTIVE MULTI-FORMAT PROMPT PREPROCESSING USING UNIFIED REINFORCEMENT LEARNING FRAMEWORK", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10798