DIVERGE: Diversity-Enhanced RAG for Open-Ended Information Seeking

TL;DR

Diverge tackles the tendency of retrieval augmented generation to produce homogenized open-ended answers by explicitly modeling diverse viewpoints through reflection-guided generation and memory-augmented iterative refinement. The framework enables viewpoint-conditioned retrieval and groundings in external evidence without relying on token-level logits, achieving state-of-the-art diversity while preserving quality on the real-world Infinity-Chat dataset. The authors introduce novel diversity–quality metrics, including semantic and viewpoint diversity, and report substantial improvements in diversity with negligible quality loss, highlighting a systemic limitation in current LLM-based information seeking. This work broadens access to multiple plausible perspectives and offers a practical, plug-and-play approach for open-ended information seeking in real-world deployments.

Abstract

Existing retrieval-augmented generation (RAG) systems are primarily designed under the assumption that each query has a single correct answer. This overlooks common information-seeking scenarios with multiple plausible answers, where diversity is essential to avoid collapsing to a single dominant response, thereby constraining creativity and compromising fair and inclusive information access. Our analysis reveals a commonly overlooked limitation of standard RAG systems: they underutilize retrieved context diversity, such that increasing retrieval diversity alone does not yield diverse generations. To address this limitation, we propose DIVERGE, a plug-and-play agentic RAG framework with novel reflection-guided generation and memory-augmented iterative refinement, which promotes diverse viewpoints while preserving answer quality. We introduce novel metrics tailored to evaluating the diversity-quality trade-off in open-ended questions, and show that they correlate well with human judgments. We demonstrate that DIVERGE achieves the best diversity-quality trade-off compared to competitive baselines and previous state-of-the-art methods on the real-world Infinity-Chat dataset, substantially improving diversity while maintaining quality. More broadly, our results reveal a systematic limitation of current LLM-based systems for open-ended information-seeking and show that explicitly modeling diversity can mitigate it. Our code is available at: https://github.com/au-clan/Diverge

Figures (8)

• Figure 1: Left: Illustrative example of an open-ended information-seeking query. (a) LLMs exhibit homogenized (blue) outputs, (b) standard RAG still produces repetitive responses even when the retrieved contexts contain diverse evidence. In contrast, (c) Diverge generates diverse outputs while maintaining high answer quality. Right: Diversity–quality trade-off of different methods. Upper-Right indicates better. Diverge achieves the best performance among all methods. See Appendix for additional baselines and metrics.
• Figure 2: Overview of Diverge, a plug-and-play agentic RAG framework for open-ended settings that promotes diverse viewpoints via reflection-guided viewpoint generation and viewpoint-conditioned retrieval and generation, with broad LLM compatibility.
• Figure 3: Ablation study of Diverge showing the impact of removing search grounding and result refinement on the performance of Diversity, Quality, and Unified Score on the GPT-5-mini model, highlighting the contributions of these components in Diverge.
• Figure 4: Responses to one query are projected into two dimensions using PCA over sentence embeddings. In this case, all responses are plausible. The visualization reveals three prominent clusters: homogeneous responses from direct LLM prompting; a separate but tightly grouped cluster from RAG and its variants, indicating that they differ from the LLM yet remain highly similar to each other; and a more diverse cluster corresponding to Diverge.
• Figure 5: Responses with higher viewpoint (vs. semantic) diversity contain more distinct claims, while higher semantic (vs. viewpoint) diversity corresponds to fewer claims.