The core problem is the model ' s hallucination—it invented a factual detail—in a context (news reporting) where factual accuracy is non-negotiable. To correct a factual error in a generative summary, the model must be constrained to speak only based on verifiable facts from a reliable source.
The most effective technique to combat hallucinations and ensure factual adherence is Grounding (D). Grounding connects the Large Language Model ' s (LLM ' s) output to a specific, trusted, and verifiable source of information. This is often implemented using Retrieval-Augmented Generation (RAG). In this scenario, grounding the summary model on the original source articles ensures that every generated statement is directly entailed by the provided facts (the source article content).
Option B, fine-tuning, is expensive and only updates the model ' s general knowledge and style; it does not prevent the model from guessing or fabricating details when retrieving information. Option C, increasing temperature, would make the output less consistent and more diverse, likely increasing the chance of hallucination, which is the opposite of the desired effect. Option A is unrelated to factual accuracy. Therefore, Grounding is the necessary step to anchor the model ' s responses to the true content of the source articles.
(Reference: Google Cloud documentation on RAG/Grounding emphasizes that its primary purpose is to address the “knowledge cutoff” and hallucination issues of LLMs by retrieving relevant, up-to- date information from external knowledge sources and using this retrieved information to ground the LLM ' s generation, ensuring factual accuracy.)
===========
