$\text{R}^2\text{R}$: A Route-to-Rerank Post-Training Framework for Multi-Domain Decoder-Only Rerankers

TL;DR

R2R is a flexible, domain-aware reranking framework that improves multi-domain retrieval tasks by combining expert routing and entity abstraction to enhance relevance understanding across high-stakes fields.

Contribution

The paper introduces R2R, a novel framework integrating dynamic expert routing and EAG to improve domain-specific reranking without overfitting or forgetting.

Findings

R2R outperforms generalist and single-domain models across legal, medical, and financial domains.

EAG effectively prevents overfitting by masking surface cues, promoting domain-invariant learning.

The Latent Semantic Router efficiently activates domain experts, enhancing model robustness.

Abstract

Decoder-only rerankers are central to Retrieval-Augmented Generation (RAG). However, generalist models miss domain-specific nuances in high-stakes fields like finance and law, and naive fine-tuning causes surface-form overfitting and catastrophic forgetting. To address this challenge, we introduce R2R, a domain-aware framework that combines dynamic expert routing with a two-stage training strategy, Entity Abstraction for Generalization (EAG). EAG introduces a counter-shortcut mechanism by masking the most predictive surface cues, forcing the reranker to learn domain-invariant relevance patterns rather than memorizing dataset-specific entities. To efficiently activate domain experts, R2R employs a lightweight Latent Semantic Router that probes internal representations from the frozen backbone decoder to select the optimal LoRA expert per query. Extensive experiments across different reranker backbones and diverse domains (legal, medical, and financial) demonstrate that R2R consistently surpasses generalist and single-domain fine-tuned baselines. Our results confirm that R2R is a model-agnostic and modular approach to domain specialization with strong cross-domain robustness.