Experimentation

A pipeline conditions an image generator on a long, detailed prompt. The text encoder has a fixed maximum token length, and prompts that exceed it are silently truncated before the embedding is produced. Users report that trailing details in their long prompts, such as background and lighting, are consistently ignored in the generated images. Which change most directly addresses the root cause of the ignored details?

• ASplit the prompt into segments within the token limit, encode each segment, and combine their embeddings so all details are represented in the conditioningcheck_circle Correct
• BRaise the classifier-free guidance scale so the conditioning embedding exerts a stronger pull on every denoising step
• CIncrease the number of denoising timesteps so the model has more iterations to render the requested details
• DLower the sampler's noise level at the final steps so faint background features are preserved rather than blurred

Why A is correct: The trailing tokens were dropped before encoding, so the fix is to keep them inside the encoder's limit by segmenting and encoding each part, then merging the embeddings so background and lighting are actually present in the conditioning signal.

Why B is wrong: Stronger guidance does make the model adhere more to its conditioning and tempts those who treat weak adherence as the problem, but the trailing details never reached the embedding, so amplifying it cannot restore information that was already discarded.

Why C is wrong: More steps refine detail that the conditioning already specifies and seems like a fidelity lever, but extra iterations cannot synthesise instructions that were truncated away and never embedded, so the missing details stay missing.

Why D is wrong: Reducing late-step noise can sharpen fine texture and feels relevant to faint background elements, but the issue is absent conditioning information, not blurring, so this addresses a different failure than the silent truncation.

A team trains a text-conditional diffusion denoiser and wants to use classifier-free guidance at inference so they can trade prompt fidelity against diversity with a single scale knob. Their training loop only ever passes the real text embedding to the U-Net. At sampling time, raising the guidance scale produces no usable steering and the outputs barely respond to the prompt. Which change to the training procedure is required to make classifier-free guidance work as intended?

• ATrain a separate noise-aware image classifier on the noisy latents and use its gradients to push samples toward the target class at each reverse step.
• BIncrease the number of forward diffusion timesteps and switch from a linear to a cosine noise schedule so the unconditional signal emerges naturally.
• CRandomly replace the text condition with a learned null or empty embedding for a fixed fraction of training steps, so the same network learns both conditional and unconditional denoising.check_circle Correct
• DNormalise every text embedding to unit L2 norm before injection so the conditional and unconditional magnitudes match during sampling.

Why A is wrong: This is classifier guidance, the earlier approach that classifier-free guidance was designed to replace; it is tempting because it also steers samples, but it requires an extra noise-robust classifier and does not make the denoiser itself usable in the classifier-free formulation the team has chosen.

Why B is wrong: A cosine schedule and more timesteps can improve overall sample quality, which makes this sound plausible, but neither produces an unconditional prediction; the schedule never teaches the network to denoise with the condition removed, so guidance still has no baseline to extrapolate from.

Why C is correct: Classifier-free guidance forms its steering vector from the difference between the conditional and unconditional predictions of one shared network, so the denoiser must be trained to produce both by dropping the condition on a fraction of steps; without that the unconditional branch is undefined and guidance has nothing to extrapolate from.

Why D is wrong: Embedding normalisation can stabilise conditioning magnitudes and so seems related to the symptom, but it still only ever feeds a real condition; the model never learns an unconditional mode, so the guided difference remains close to zero and the prompt response stays weak.

An evaluation harness for a document retrieval system computes a single number that rewards placing the first relevant passage as high as possible in the ranked list, averaged over all queries. The team cares specifically about how quickly users reach one correct answer, not about retrieving every relevant passage. Which evaluation metric best matches this requirement for tuning their embedding model?

• AMean reciprocal rank, because it averages the reciprocal of the rank position of the first relevant result across queries, rewarding models that surface a correct answer early.check_circle Correct
• BRecall at k, because it measures the fraction of all relevant passages that appear in the top k results, which captures how early the first correct answer is surfaced for the user.
• CCosine similarity of the query embedding to the gold passage, because a higher pairwise similarity score directly reflects the ranking quality across the whole result list.
• DNormalised discounted cumulative gain, because it discards graded relevance and reduces to counting only the single first relevant hit per query in the ranked list.

Why A is correct: Correct: mean reciprocal rank uses 1 divided by the rank of the first relevant item, so it directly measures how high the first correct passage appears, matching the stated early-answer goal.

Why B is wrong: Tempting because recall at k is a standard retrieval metric, but it counts how many relevant items are retrieved, not the rank of the first one; it ignores ordering within the top k and rewards completeness instead.

Why C is wrong: Tempting because cosine similarity drives the ranking, but a single-pair similarity is not a list-level ranking metric and says nothing about where the first relevant passage lands among distractors.

Why D is wrong: Tempting because normalised discounted cumulative gain is rank-aware, but it uses graded relevance across the whole list rather than discarding it, so the description is wrong and it does not reduce to a first-hit measure.

Two engineers discuss how temperature affects the next-token distribution of a transformer language model during sampling. Temperature scales the logits by dividing them before the softmax is applied. Which statement most accurately describes the effect of raising the temperature above 1.0 on that distribution?

• AIt flattens the probability distribution, raising the relative likelihood of lower-scoring tokens and increasing output diversity.check_circle Correct
• BIt sharpens the probability distribution, concentrating mass on the single highest-scoring token and reducing diversity.
• CIt removes all tokens whose cumulative probability falls outside a chosen mass threshold before renormalising.
• DIt fixes the candidate pool to a constant number of highest-scoring tokens regardless of their probability spread.

Why A is correct: Dividing logits by a value greater than one shrinks the gaps between them, so the softmax produces a flatter distribution that lifts low-probability tokens and yields more varied, less predictable text.

Why B is wrong: This is the effect of temperature below 1.0, not above it; candidates who confuse the direction of the scaling pick this, but raising temperature flattens rather than sharpens the distribution.

Why C is wrong: This describes nucleus (top-p) sampling, a separate truncation step; it is tempting because both control diversity, but temperature rescales logits rather than truncating the candidate set.

Why D is wrong: This describes top-k sampling, which caps the candidate count; it sounds related because it also shapes sampling, but temperature alters the shape of the whole distribution, not the size of a fixed pool.

An engineer normalises every refined embedding to unit length before storing it in a vector index that ranks results by inner product. Conceptually, what does this unit-length normalisation make the inner-product ranking equivalent to?

• ARanking by cosine similarity, because the inner product of two unit vectors equals the cosine of the angle between them.check_circle Correct
• BRanking by Euclidean distance squared, because normalised vectors all lie on a sphere where squared distance and inner product are unrelated.
• CRanking by raw token overlap, because unit normalisation projects each embedding onto its most active dimension.
• DRanking by vector magnitude, because after normalisation the inner product is dominated by whichever vector had the larger original norm.

Why A is correct: When both vectors have unit norm, their inner product reduces to the cosine of the angle between them, so an inner-product index over normalised embeddings produces the same ordering as cosine similarity.

Why B is wrong: On the unit sphere, squared Euclidean distance is a decreasing function of the inner product, so they are tightly related rather than unrelated; the clause contradicts itself and is wrong, though the sphere framing sounds plausible.

Why C is wrong: Normalisation rescales the whole vector and does not project onto a single dimension or recover token overlap, so this misdescribes both the operation and its effect; it may tempt by invoking lexical matching.

Why D is wrong: Normalisation removes magnitude entirely by setting every norm to one, so the inner product cannot be dominated by original norms; this is the opposite of what normalisation achieves.