Pairwise Ranking Outperforms Single-Action RL for Offline Explanation Selection: A Practical Lesson

Tanay Chowdhury, Saeideh Shahrokh Esfahani

Industrial explainable-recommendation systems built on LLMs incur a substantial serving cost: each request triggers an LLM generation, with latency in the hundreds of milliseconds…