Transformer Fundamentals for NCA-GENL: The 30% Domain [2026]

Core Machine Learning and AI Knowledge is 30% of NCA-GENL, and at a minute per question it is the domain where fuzzy understanding costs the most: you either retrieve the concept instantly or you donate the point. The good news is that the domain is smaller than it looks. Nearly every question resolves to one of five mental models: what a transformer does with a sequence, what embeddings are, what tokenization implies, what the decoding parameters control, and which adaptation technique fits which problem. This guide builds each one at exactly exam altitude: deep enough to answer scenario questions, without the graduate-course detours the exam never asks about.

The transformer, in the exam's terms

A transformer processes a sequence of tokens in parallel, and attention is the mechanism that lets every position weigh every other position when building its representation: the word "bank" attends to "river" or "loan" and means accordingly. That parallelism (versus the step-by-step processing of older recurrent networks) is why transformers train efficiently on GPUs, which is the historical question the exam actually asks. Generative LLMs are typically decoder-style: they produce output one token at a time, each new token conditioned on everything before it. Two graded consequences: generation is inherently sequential even though training parallelizes, and everything the model "knows" during a conversation must fit in its context window (the bounded token budget for prompt plus response combined).

Embeddings and tokenization

Embeddings are vectors that represent meaning: similar things land near each other, which is what makes semantic search and retrieval augmentation work, and "compare embeddings" is the graded answer whenever a stem wants similarity beyond keyword matching. Tokenization splits text into subword units before any of this happens, and its exam-relevant implications are practical: token counts drive context limits and costs, rare words split into more tokens, and a model reasons over tokens rather than characters, which explains classic failure modes on spelling-level tasks.

Decoding: the knobs and what they trade

Sampling parameters control how the next token gets picked, and the exam tests them as behavior, never as formulas. Temperature scales randomness: low values make outputs focused and repeatable (extraction, classification), higher values diversify (brainstorming), and the trap options apply high temperature to tasks needing determinism. Top-k and top-p bound the candidate pool by count or cumulative probability. Max output length caps generation, and a truncated answer points at it. When a stem shows inconsistent outputs on a task that demands consistency, the graded reflex is lowering the sampling randomness before touching anything else.

The adaptation ladder

The domain's judgment questions live here: four ways to make a general model fit your task, ordered by cost, and the skill is matching the problem to the cheapest rung that solves it.

The two traps worth naming: fine-tuning as a knowledge fix (knowledge freshness is retrieval's job; fine-tuning bakes yesterday in) and prompting dismissed as unserious (the exam treats it as the professional first move). Hallucination stems connect here too: a model inventing facts is grounds for grounding (retrieval plus instructions to cite), never for more temperature.

Worked scenario: the inconsistent extractor

A team extracts invoice fields with an LLM. The same invoice, run twice, produces differently worded and occasionally differently valued outputs. What is the first fix?

Lower the sampling randomness (temperature toward the deterministic end) for a task that has one right answer, and the variance in wording and values drops immediately at zero cost. Fine-tuning is the over-engineered distractor, RAG is the wrong-problem distractor (nothing here is a knowledge gap), and a longer prompt may help quality but does not address the stochastic knob the stem is pointing at.

Worked scenario: the outdated assistant

A support bot fine-tuned eight months ago answers confidently with discontinued product details. The catalog changes weekly. What does the exam want?

Retrieval augmentation over the live catalog: fresh knowledge belongs in a store the system searches at question time, so weekly changes propagate without retraining. Re-fine-tuning weekly is the cost trap; prompting alone cannot inject facts the model was never given; and raising temperature adds variety to wrongness. The stem's tell is "changes weekly": update frequency picks the rung.

Next steps

Drill it at pace: the practice tests ask this domain at 30% weight under the 60-minute clock, and the mistakes article shows where its points usually leak.

Sources:

• NVIDIA Generative AI LLMs certification
• NVIDIA: What is a transformer model?
• NVIDIA NeMo documentation