OptiSet: Unified Optimizing Set Selection and Ranking for Retrieval-Augmented Generation

TL;DR

Retrieval-Augmented Generation often relies on static top-k retrieval that can waste context on redundant passages and miss combinatorial gains from diverse evidence. OptiSet introduces an Expand-then-Refine paradigm to build compact, diverse evidence sets and a self-synthesized, set-list wise training framework that jointly optimizes set selection and set-level ranking. Utility signals based on perplexity and delta-entropy guide data labeling without strong supervision, with an objective $L_{all} = L_{CE} + \lambda L_{KL}$ to balance selection and ranking. Across HotpotQA, Bamboogle, MuSiQue, and TriviaQA, OptiSet achieves improved QA performance with smaller, more efficient evidence sets and demonstrates robustness to different retrievers and K settings, offering practical efficiency gains for real-world RAG systems. The approach advances combinatorial set optimization in RAG by coupling diversity, redundancy reduction, and learning-driven ranking into a single framework, enabling scalable improvements in complex reasoning tasks.

Abstract

Retrieval-Augmented Generation (RAG) improves generation quality by incorporating evidence retrieved from large external corpora. However, most existing methods rely on statically selecting top-k passages based on individual relevance, which fails to exploit combinatorial gains among passages and often introduces substantial redundancy. To address this limitation, we propose OptiSet, a set-centric framework that unifies set selection and set-level ranking for RAG. OptiSet adopts an "Expand-then-Refine" paradigm: it first expands a query into multiple perspectives to enable a diverse candidate pool and then refines the candidate pool via re-selection to form a compact evidence set. We then devise a self-synthesis strategy without strong LLM supervision to derive preference labels from the set conditional utility changes of the generator, thereby identifying complementary and redundant evidence. Finally, we introduce a set-list wise training strategy that jointly optimizes set selection and set-level ranking, enabling the model to favor compact, high-gain evidence sets. Extensive experiments demonstrate that OptiSet improves performance on complex combinatorial problems and makes generation more efficient. The source code is publicly available.

Figures (5)

• Figure 1: Illustration of the difference in objectives: Set-list wise modeling unifies the modeling of set selection to capture combinatorial gains and reduce redundancy.
• Figure 2: The framework diagram of OptiSet is shown. The upper part represents the "Expand-then-Refine" paradigm, which can serve as both a training-free framework and a training data synthesis pipeline. The lower part represents set-list wise training. It utilizes multiple sets generated by the framework, performs partial ordering and signal labeling, and trains the selector.
• Figure 3: Illustration of the distribution of the number of documents dynamically selected from a set.
• Figure 4: Illustration of gain. Changes in hit of the gold answer and QA performance using OptiSet. The left side compares to RankZephyr, and the right side compares to Rank1. Avg(EM, F1) is the metric.
• Figure 5: Illustration showing how generation performance changes as the candidate sizes changes or different retrievers are replaced. The left part is HotpotQA, and the right part is MuSiQue. For fairness, Avg(EM, F1) is uesd as the evaluation metric.