Systematic comparison of prompting-based vs. RL-based query augmentation

Determine the relative strengths, weaknesses, and applicability in real-world settings of prompting-based query augmentation methods (e.g., pseudo-document generation) versus reinforcement learning–based query rewriting formulations through a systematic, controlled comparison.

Background

The paper highlights two dominant paradigms for improving queries in information retrieval: prompting-based pseudo-document/query expansion and reinforcement learning–based query rewriting. Although each has advantages and limitations, the literature has lacked a rigorous, apples-to-apples comparison to clarify when and why one approach outperforms the other.

Establishing a systematic comparison is important for practitioners who must choose between these approaches for diverse retrieval settings (sparse and dense retrievers) and applications, but prior work left this question unresolved, motivating the present paper.