After identifying that a medical diagnostic assistant consistently generates fictitious citations and incorrect drug dosages during testing, a…

The most effective and appropriate solution is to implement Retrieval-Augmented Generation (RAG). This architecture mitigates hallucinations by "grounding" the Large Language Model (LLM) in a specific, curated, and verifiable knowledge base, such as a vector database containing approved medical literature and drug information. Before generating a response, the RAG system retrieves relevant, factual documents from this database and provides them to the LLM as context. This forces the model's output to be based on reliable data rather than its internal, parametric knowledge, which can be prone to fabrication. Furthermore, the retrieval process itself creates a clear, auditable trail, as each generated answer can be linked directly to the source documents used to create it, which is essential for regulatory compliance in a medical context.

A. Adjust the temperature parameter to a value of 1.0: A temperature of 1.0 increases the randomness and "creativity" of the LLM's output, which would likely exacerbate the problem of hallucinations, not solve it. A lower temperature (closer to 0) would make the output more deterministic but would not solve the core issue of the model lacking verifiable, external knowledge.

B. Increase the GPU memory allocation for faster inference: This is a performance optimization. It affects the speed at which the model generates a response but has no bearing on the factual accuracy or truthfulness of the content itself.

C. Deploy a Web Application Firewall (WAF) for API protection: A WAF is a network security control designed to protect web applications from common attacks like SQL injection, Cross-Site Scripting (XSS), and denial-of-service. It operates at the application layer but does not inspect or control the internal logic of the AI model's content generation process.

1. Lewis, P., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. In Advances in Neural Information Processing Systems 33 (NeurIPS 2020). This foundational academic paper introduces the RAG model, demonstrating its effectiveness in reducing factual inaccuracies (hallucinations) by retrieving relevant documents from a knowledge corpus before generation. (See Section 1: Introduction).

2. Stanford University. (2023). CS224N: NLP with Deep Learning, Lecture on Large Language Models and Retrieval. Course materials frequently discuss hallucination as a key failure mode of LLMs and present RAG as a primary mitigation strategy. The lectures explain that RAG provides a mechanism for fact-checking and attribution by linking outputs to source documents.

3. Gao, Y., et al. (2023). Retrieval-Augmented Generation for Large Language Models: A Survey. This survey paper provides a comprehensive overview of RAG architectures, explicitly stating that a key motivation for RAG is "to mitigate issues like hallucination and to provide access to up-to-date information." (See Section 2.1: Motivations). DOI: https://doi.org/10.48550/arXiv.2312.10997