RARE: Retrieval-Augmented Reasoning Modeling

TL;DR

RARE tackles the challenge of domain-specific intelligence by decoupling knowledge storage from reasoning. It externalizes domain knowledge to retrievable sources and trains a reasoning-centric model that learns contextualized reasoning patterns via retrieval-augmented prompts and masked losses, enabling lightweight models to outperform much larger baselines. Across medical, legal, financial, and multimodal benchmarks, RARE achieves state-of-the-art accuracy with improved efficiency and robustness, even rivaling trillion-parameter systems in some cases. The approach is data- and parameter-efficient, and it generalizes well under multi-task and RL settings, signaling a scalable path toward domain-specific, reasoning-focused AI systems.

Abstract

Domain-specific intelligence demands specialized knowledge and sophisticated reasoning for problem-solving, posing significant challenges for large language models (LLMs) that struggle with knowledge hallucination and inadequate reasoning capabilities under constrained parameter budgets. Inspired by Bloom's Taxonomy in educational theory, we propose Retrieval-Augmented Reasoning Modeling (RARE), a novel paradigm that decouples knowledge storage from reasoning optimization. RARE externalizes domain knowledge to retrievable sources and internalizes domain-specific reasoning patterns during training. Specifically, by injecting retrieved knowledge into training prompts with masked losses, RARE transforms learning objectives from rote memorization to contextualized reasoning. It enables models to bypass parameter-intensive memorization and prioritize the development of higher-order cognitive processes. Extensive experiments demonstrate that lightweight RARE-trained models (e.g., Llama-3.1-8B) could achieve state-of-the-art performance, surpassing retrieval-augmented GPT-4 and DeepSeek-R1 up to approximately 20\% accuracy. RARE establishes a paradigm shift where maintainable external knowledge bases synergize with compact, reasoning-optimized models, collectively driving more scalable domain-specific intelligence.

Figures (7)

• Figure 1: Performance of RARE versus baselines on benchmarks.
• Figure 2: Motivation of RARE. Left: A pyramid-shaped Bloom’s Taxonomy, illustrating the cognitive hierarchy from basic "Remember" to advanced "Evaluate" and "Create" levels. Right: The correspondence between Domain Knowledge and Domain Thinking with Bloom’s cognitive hierarchy (example related to government bond yields). In contrast to domain knowledge, domain thinking corresponds to the higher-order cognitive process—although relatively rare, it plays a crucial role.
• Figure 3: Preliminary experiments of RARE on PubHealth (medical), CaseHOLD (legal), and FinFact (financial). It reveals the dynamics: injecting retrieved context into training prompts transforms learning objectives from memorization into knowledge integration and contextualized reasoning.
• Figure 4: Comparing RARE with RAFT and SFT+RAG across various benchmarks and diverse backbones. RARE not only achieves better accuracy, but also exhibits better training robustness.
• Figure 5: Performance of RARE versus baselines on benchmarks (extended version).