LLM-as-MMR: Stepwise Diversification for Redundancy-Aware Document Reranking

ACL ARR 2026 January Submission6031 Authors

Keywords: Large Language Model, Rerank, Retrieval-Augmented Generation

Abstract:

Modern retrieval pipelines often rely on relevance-only ranking, which can return highly redundant passages and fail to cover complementary evidence needed to answer ambiguous queries.
We propose an LLM-driven diversification reranker inspired by Maximal Marginal Relevance (MMR).
Given a query and a fixed pool of candidate documents, our method constructs the ranked list iteratively: at each step, the LLM selects the next document by jointly considering (i) relevance to the query and (ii) marginal novelty with respect to the already selected set, explicitly penalizing information overlap while rewarding complementary coverage.
This “LLM-as-MMR” formulation leverages the model’s semantic understanding to estimate redundancy beyond surface-level similarity, without relying on hand-crafted similarity functions.
Experiments on ambiguous question answering style reranking benchmarks show that our approach improves answer coverage and ranking quality while producing more diverse, less redundant top-$k$ results compared to similarity-based ranking and heuristic diversification baselines.

Paper Type: Long

Research Area: Retrieval-Augmented Language Models

Research Area Keywords: Information Retrieval and Text Mining,Question Answering

Languages Studied: English

Submission Number: 6031