Lists
Lists
Rajpoot, Abha Kiran, Virdee, Bal Singh and Khanna, Ashish (2025) Retrieval augmented neural adapters for domain specific customization of large language models. International Journal of Applied Mathematics, 38 (5s). pp. 536-549. ISSN 1314-8060
Large language models (LLMs) excel at general‑purpose text generation but struggle to reflect specialized domain knowledge and to adapt quickly without expensive fine‑tuning. This paper introduces Retrieval‑Augmented Neural Adapters (RANA) a scalable framework that augments a frozen LLM with a lightweight neural adapter and a retrieval module. The adapter integrates retrieved domain evidence into the model’s hidden states using low‑rank transformations, while prompt construction stitches together the query, retrieved evidence and adapter signal. The research addresses two objectives: (1) to develop a general yet domain‑aware framework that enhances LLMs with domain‑specific knowledge for specialized applications; and (2) to eliminate the need for full fine‑tuning by dynamically integrating knowledge through retrieval and prompt engineering. Unlike existing knowledge‑injection approaches, RANA keeps the LLM parameters frozen and applies adapter conditioning at inference time, enabling fast deployment across domains. We evaluate RANA on entity typing, relation classification, open‑domain question answering, factual probing and biomedical question answering. Comparative experiments against BERT, RoBERTa, knowledge‑infused baselines and K‑Adapter show that RANA consistently improves micro‑F1, exact‑match and precision‑at‑1 metrics by 2–4 % and achieves state‑of‑the‑art results on BioASQ. These findings demonstrate that retrieval‑augmented adapters can serve as a versatile bridge between static LLMs and dynamic domain requirements. The proposed methodology opens new directions for zero‑shot domain customization and underscores the role of retrieval in aligning language models with specialized expertise.
 |
View Item |