Prepare and test your skills
Prepare and test your skills
Free sample question, answered and explained below. The Generative AI Leader practice bank has 2,191 more like it. Try 10 free: instant score, no account.
An enterprise is developing an internal generative AI assistant to help customer support agents answer questions about corporate policies and benefits. During testing, the team observes that the base model occasionally generates convincing but completely inaccurate statements and fails to incorporate policy guidelines updated earlier this week.
The engineering team needs to tether the model's responses to authoritative corporate documents, enable auditability through citations, and reduce inaccurate generations.
Which technique should the team implement?
In generative AI, grounding is the process of connecting a model's output to verifiable, authoritative sources of information. Rather than forcing a model to rely solely on the static knowledge embedded in its parameters during pre-training, grounding tethers the generated responses directly to specified data repositories, such as enterprise file stores, internal databases, or live search indexes.

Grounding is specifically designed to solve the twin challenges of hallucinations and outdated static knowledge. By supplying factual reference material directly to the model during inference, it guarantees reliable, auditable, and current answers.
Foundation model pre-training is the initial, computationally massive phase of building an AI model from scratch. During pre-training, a deep neural network processes hundreds of billions or trillions of tokens across diverse text, code, and multimodal datasets to learn fundamental grammar, reasoning patterns, and general world knowledge.
Pre-training is not a mechanism for connecting a model to verifiable operational documents. Building a foundation model requires millions of dollars in compute infrastructure, specialized machine learning teams, and months of training time. Furthermore, pre-training results in a static snapshot of data with an explicit knowledge cutoff date, meaning it cannot reflect recent policy changes and remains prone to producing statistical hallucinations.
The company were a frontier AI laboratory developing an entirely new foundational model architecture from raw web-scale corpora.
Enterprise solutions rarely train foundational models from scratch. Instead, they consume existing foundation models and apply grounding to supply proprietary context securely.
Parameter-efficient fine-tuning (PEFT) is an adaptation method that customizes a pre-trained foundation model for specific tasks or styles by updating only a small percentage of its weights or training small modular adapter layers (such as Low-Rank Adaptation or LoRA), leaving the core model weights frozen.
While PEFT effectively trains a model to adopt a specific tone, formatting style, or domain vocabulary, it does not anchor responses to verifiable reference sources in real time. Model weights cannot be updated every day or week as company policies change. More importantly, fine-tuning modifies internal parametric memory, which does not eliminate hallucinations and cannot dynamically provide auditable citations linking back to original corporate documents.
The organization wanted to train the assistant to follow a highly specialized conversational style, adhere to a strict structured output schema (like JSON), or perform classification without needing external fact retrieval.
A common industry best practice is to pair fine-tuning for style and syntax with grounding for factual accuracy and verification.
Model quantization is a post-training compression technique used to make large language models smaller and faster. It converts model weights and activation tensors from high-precision numerical formats (such as 16-bit or 32-bit floating-point numbers) into lower-precision representations (such as 8-bit or 4-bit integers).
Quantization addresses computational efficiency and hardware constraints, not factual accuracy or source verification. Compressing weight parameters does not provide the model with access to recent policy updates, does not supply external reference documents, and does not provide an audit trail of citations. In fact, if applied too aggressively, quantization can slightly degrade reasoning fidelity without mitigating hallucinations.
The team needed to run the virtual assistant on low-cost hardware or edge devices with restricted memory capacity while minimizing latency.
Quantization is an infrastructure and deployment optimization strategy, whereas grounding is an architectural technique for factual accuracy and verification.
Answer them yourself and get your score the moment you finish. No account needed.
Full access includes
Keep the momentum going with these hand-picked practice scenarios
Last updated