![[2602.15874] P-RAG: Prompt-Enhanced Parametric RAG with LoRA and Selective CoT for Biomedical and Multi-Hop QA](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.15874] P-RAG: Prompt-Enhanced Parametric RAG with LoRA and Selective CoT for Biomedical and Multi-Hop QA
Summary
The paper introduces P-RAG, a novel hybrid architecture that enhances Retrieval-Augmented Generation (RAG) for biomedical question answering, demonstrating significant performance improvements over existing methods.
Why It Matters
As large language models (LLMs) face limitations due to static training data, P-RAG offers a promising solution by integrating parametric knowledge and retrieval methods, which could enhance the accuracy and adaptability of AI in biomedical applications. This research is crucial for advancing AI's role in healthcare and improving multi-hop reasoning capabilities.
Key Takeaways
- P-RAG significantly outperforms Standard RAG in biomedical question answering tasks.
- The integration of Chain-of-Thought prompting enhances multi-hop reasoning capabilities.
- LoRA fine-tuning of LLaMA-3.2-1B-Instruct improves model performance on specialized datasets.
- P-RAG achieves state-of-the-art results on PubMedQA and 2WikiMultihopQA benchmarks.
- The study highlights the importance of dynamic knowledge retrieval in enhancing LLM capabilities.
Computer Science > Computation and Language arXiv:2602.15874 (cs) [Submitted on 2 Feb 2026] Title:P-RAG: Prompt-Enhanced Parametric RAG with LoRA and Selective CoT for Biomedical and Multi-Hop QA Authors:Xingda Lyu, Gongfu Lyu, Zitai Yan, Yuxin Jiang View a PDF of the paper titled P-RAG: Prompt-Enhanced Parametric RAG with LoRA and Selective CoT for Biomedical and Multi-Hop QA, by Xingda Lyu and 3 other authors View PDF Abstract:Large Language Models (LLMs) demonstrate remarkable capabilities but remain limited by their reliance on static training data. Retrieval-Augmented Generation (RAG) addresses this constraint by retrieving external knowledge during inference, though it still depends heavily on knowledge base quality. To explore potential improvements, we evaluated three RAG variants-Standard RAG, DA-RAG, and our proposed Prompt-Enhanced Parametric RAG (P-RAG), a hybrid architecture that integrates parametric knowledge within the LLM and retrieved evidence, guided by Chain-of-Thought (CoT) prompting and Low-Rank Adaptation (LoRA) fine-tuning-on both general and biomedical datasets. Using LLaMA-3.2-1B-Instruct fine-tuned via LoRA, we evaluate on PubMedQA and 2WikiMultihopQA. P-RAG outperforms Standard RAG on PubMedQA by 10.47 percentage points in F1 (93.33% vs. 82.86%; 12.64% relative). On 2WikiMultihopQA, P-RAG nearly doubles the overall score vs. Standard RAG (33.44% vs. 17.83%) and achieves 44.03% on the Compare subset (with 42.74% Bridge, 21.84% Inference, 8.60% Co...
