Joint LinUCB Algorithm for Progressive LLM Content Refinement

This method introduces a contextual bandit algorithm using Expectation-Maximization to model reward decay during iterative LLM refinement. By treating prompts as arms in a joint learning framework, it prevents over-exploitation and addresses performance saturation on GSM8K and Sentiment Reversal benchmarks.