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

2025-11-25

AI Summary

To address poor generalization and catastrophic forgetting of decoder-specific re-rankers in multi-domain RAG, this paper proposes a domain-aware dynamic expert re-ranking framework. Methodologically, it introduces: (1) an Entity Abstraction for Generalization (EAG) strategy that masks superficial cues via entity abstraction to mitigate overfitting; (2) a lightweight latent semantic router that dynamically activates LoRA-based expert modules, enabling model-agnostic, modular cross-domain adaptation; and (3) a two-stage training protocol—freezing the backbone decoder’s representation learning while jointly optimizing the router and experts. Evaluated across legal, financial, and medical domains, the method significantly outperforms both generic and single-domain fine-tuned baselines. It demonstrates strong cross-domain robustness and plug-and-play compatibility, achieving effective zero-shot transfer without architectural modification or domain-specific retraining.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Search and Optimization: Metareasoning and Metaheuristics
• Knowledge Representation and Reasoning: Argumentation

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• Decoder-only rerankers miss domain-specific nuances in high-stakes fields
• Naive fine-tuning causes surface-form overfitting and catastrophic forgetting
• Generalist models lack cross-domain robustness for specialized domains

Innovation

• Dynamic expert routing with two-stage training strategy
• Counter-shortcut mechanism masking predictive surface cues
• Latent Semantic Router selecting optimal LoRA expert