NVIDIA NCP-GENL: NVIDIA-Certified Professional: Generative AI LLMs

Last checked October 22, 2026 against the official NVIDIA exam page.

1. Question 1Inference Optimization

   Which technique stores the key and value tensors from earlier tokens so that each new token does not recompute them?

   1. AGradient accumulation
   2. BDropout at inference time
   3. CKV cache
   4. DLabel smoothing
   Show answer and explanationHide answer

   Correct answer: C

   The KV cache keeps attention keys and values for previously processed tokens, so generation only computes the new token's contribution. Dropout, gradient accumulation and label smoothing are training-time techniques.

2. Question 2Core ML and LLM Foundations

   A team is choosing a decoder-only transformer for text generation. Which component lets each token attend only to earlier tokens during training?

   1. AMax pooling over the sequence
   2. BBatch normalization in each block
   3. CBidirectional cross-attention
   4. DCausal (masked) self-attention
   Show answer and explanationHide answer

   Correct answer: D

   A causal mask sets attention scores for future positions to negative infinity, so each token attends only to itself and earlier tokens. Bidirectional cross-attention belongs to encoder-decoder models, and pooling or batch normalization do not restrict attention order.

3. Question 3Retrieval-Augmented Generation

   A support chatbot must answer from a company's policy documents that change weekly. Which architecture keeps answers current without retraining the model?

   1. APlacing the documents in a fixed system prompt with no retrieval layer, loaded once at deployment time
   2. BIncreasing the sampling temperature to 2.0 during decoding so answers vary more across runs
   3. CRetrieval-augmented generation with a vector index that is refreshed as documents change
   4. DFine-tuning once a year on a static snapshot of internal policy documents
   Show answer and explanationHide answer

   Correct answer: C

   RAG retrieves current passages at query time and supplies them as context, so updating the index updates answers without retraining. A yearly snapshot goes stale, temperature changes randomness rather than knowledge, and a fixed prompt cannot hold a changing corpus.

4. Question 4Trustworthy AI

   A deployed assistant occasionally reveals personal data that appeared in its training set. Which combination is the most direct mitigation?

   1. AData governance review, deduplication, and output filtering for personal identifiers, checked by a privacy evaluation
   2. BSwitching to a larger batch size during pretraining to smooth gradient noise across the run
   3. CAdding more transformer layers to the network to increase model depth and total parameter count
   4. DIncreasing the vocabulary size of the byte-pair encoding tokenizer so that rare personal identifiers are split into fewer tokens
   Show answer and explanationHide answer

   Correct answer: A

   Reviewing training data, removing duplicated sensitive content, and filtering outputs address memorization and leakage directly, and a privacy evaluation confirms the result. The other changes affect model size or optimization, not leakage.

5. Question 5Evaluation

   Which two metrics are most useful for judging the retrieval step of a RAG system? (Choose two.)

   1. ANumber of GPUs used during retrieval
   2. BRecall at k for relevant passages
   3. CHit rate or mean reciprocal rank of retrieved passages
   4. DPerplexity of the generator on the question alone
   5. EToken count of the final answer
   Show answer and explanationHide answer

   Correct answer: B, C

   Recall at k and ranking metrics such as mean reciprocal rank measure whether relevant passages were retrieved and where they appear. Generator perplexity, answer length and hardware count do not measure retrieval quality directly.