NCP-GENL LLM Architecture Practice Question
A research team wants to train a large decoder-only LLM where each token's representation is computed independently of token order, then inject order information afterward. They are choosing between learned absolute positional embeddings and sinusoidal absolute positional embeddings. Which statement accurately characterizes the tradeoff they face?
⚠ Common exam trap
The trap here is assuming learned positional embeddings extrapolate better because they are trainable, when in fact their fixed table size is their key limitation.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Sinusoidal embeddings are parameter-free and can be computed for any position, while learned embeddings require a fixed maximum length and add parameters.
Sinusoidal absolute positional embeddings are deterministic functions of position, adding no parameters and allowing evaluation at any index. Learned absolute embeddings allocate a trainable vector per position, consuming parameters and capping usable length at the trained maximum. The other options either reverse the extrapolation behavior or misattribute relative-offset abilities to learned absolute embeddings.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Learned embeddings are strictly better because they allow the model to discover relative offsets between tokens.
Why it's wrong here
Learned absolute embeddings encode each position as an independent vector; nothing forces them to represent relative offsets in a structured way. Relative offset awareness is a property of schemes like relative position bias or RoPE, not of learned absolute tables. Claiming they are strictly better also ignores their inability to handle unseen positions, so the option overstates the case.
- ✓
Sinusoidal embeddings are parameter-free and can be computed for any position, while learned embeddings require a fixed maximum length and add parameters.
Why this is correct
Sinusoidal embeddings use fixed sine and cosine functions of position and dimension, so they add no trainable parameters and can be evaluated at arbitrary indices. Learned embeddings allocate a trainable vector per position, which consumes parameters proportional to the maximum length and cannot produce values for unseen indices. This is the classic tradeoff between the two absolute schemes.
- ✗
Learned embeddings can extrapolate to sequence lengths longer than those seen in training, whereas sinusoidal embeddings cannot.
Why it's wrong here
This reverses the actual behavior. Learned absolute embeddings are tied to a fixed table of positions and cannot represent indices beyond the trained maximum, while sinusoidal embeddings are computed by deterministic functions and can be evaluated at any index. Extrapolation quality for sinusoidal embeddings is still limited, but the directional claim in this option is incorrect.
- ✗
Sinusoidal embeddings must be recomputed for every batch, which makes them slower than learned embeddings at inference.
Why it's wrong here
Sinusoidal embeddings depend only on position and dimension, so they can be precomputed once and reused or cached cheaply. They are not recomputed per batch in any meaningful sense. Learned embeddings are a simple table lookup, but the performance difference is negligible relative to attention and feed-forward computation, so this option misrepresents the practical tradeoff.
About these practice questions
This NCP-GENL question is part of Courseiva's 352-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam dumps. Learn why practice questions differ from exam dumps →
JA
Written and reviewed by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
Last reviewed September 2026 · checked against the official NVIDIA exam blueprint
This NCP-GENL practice question is part of Courseiva's free NVIDIA certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the NCP-GENL exam.