Pairwise Ranking Excels in Offline LLM Explanation Selection.

Explainable recommendation systems that rely on Large Language Models (LLMs) often face high serving costs and latency due to on-demand generation. To mitigate this, a new approach proposes pre-generating a pool of explanations and then using a lightweight selector at request time. This decouples the expensive generation step from the real-time selection, allowing for faster, more cost-effective responses without requiring GPU resources during serving. Researchers benchmarked various offline selectors, including pairwise learning-to-rank algorithms like LambdaRank and single-action reinforcement learning (RL) methods such as PPO, GRPO, and DPO. The evaluation used a Google Local subset and a MovieLens-1M split, measuring performance with BERTScore-F1. The findings clearly indicate that pairwise learning-to-rank, particularly LambdaRank, achieved superior F1 scores, outperforming single-action RL methods. The study attributes this success to pairwise methods utilizing more labeled candidates per rollout compared to single-action RL, which often leaves many labels unused. Additionally, a separate analysis of generator pools showed that while the absolute F1 score depends on the generator, the ranking of selectors remains consistent. This suggests that selector and generator components can be evaluated independently, offering flexibility in system design.