LLM evaluation and hallucinations quiz

Question 1

What is the distinction between intrinsic and extrinsic hallucinations in LLM outputs?

Accepted Answer

Intrinsic contradicts the given source; extrinsic adds unverifiable content not grounded in the source

Answer

Intrinsic occurs in training; extrinsic occurs at inference

Answer

Intrinsic is factual error; extrinsic is grammatical error

Answer

Intrinsic affects open-source models; extrinsic affects closed models

Question 2

Why is perplexity a poor metric for evaluating instruction-following or factual accuracy in modern LLMs?

Accepted Answer

It measures token-prediction likelihood/fluency, not factuality, helpfulness, or instruction-following

Answer

It is too expensive to compute on large corpora

Answer

It only works for encoder models

Answer

It requires human reference answers to compute

Question 3

In the LLM-as-a-judge paradigm, what is 'position bias' and how can it be mitigated?

Accepted Answer

Judge favors a candidate based on its display order; mitigate by swapping positions and averaging

Answer

Judge always picks the longer answer; mitigate by truncation

Answer

Judge prefers its own outputs; mitigate by temperature scaling

Answer

Judge ignores the system prompt; mitigate by fine-tuning

Question 4

What does the RAGAS 'faithfulness' metric measure, and how is it typically computed?

Accepted Answer

Fraction of answer claims that are supported by the retrieved context

Answer

Fraction of retrieved chunks actually used in the answer

Answer

Semantic similarity between query and answer

Answer

Whether the answer matches a gold reference string

Question 5

How does SelfCheckGPT detect hallucinations without access to external knowledge or token probabilities?

Accepted Answer

By sampling multiple responses and flagging sentences where samples are inconsistent

Answer

By comparing token logprobs to a fixed threshold

Answer

By querying a knowledge graph for each entity

Answer

By fine-tuning a classifier on labeled hallucinations

Question 6

Why can high token-level confidence (logprobs) still coincide with hallucination, and what does this say about calibration?

Accepted Answer

Token probability reflects fluency/likelihood, not truth; calibration of correctness is separate

Answer

Logprobs are always uniform after RLHF

Answer

High confidence guarantees correctness in well-trained models

Answer

Hallucinations only occur at low-confidence tokens

Question 7

What is 'semantic entropy' and why is it preferred over lexical entropy for hallucination detection?

Accepted Answer

Entropy over meaning-clusters of samples, ignoring paraphrase variation that inflates lexical entropy

Answer

Entropy of the embedding vector norms

Answer

The KL divergence between two model checkpoints

Answer

Token-level entropy averaged over the sequence

Question 8

What is a key methodological flaw of relying solely on benchmark accuracy (e.g., MMLU) to claim an LLM is 'better'?

Accepted Answer

Contamination, narrow MCQ format, and poor transfer to open-ended/hallucination behavior

Answer

MMLU only covers code tasks

Answer

Accuracy cannot be computed for generative models

Answer

Benchmarks require human raters for every item

Question 9

How do you detect benchmark data contamination in an LLM you did not train?

Accepted Answer

Compare original vs. perturbed/novel variants and probe for verbatim recall of test items

Answer

Inspect the model's attention weights for leakage

Answer

Measure GPU memory usage during inference

Answer

Check the model's parameter count

Question 10

What is the difference between faithfulness and factuality (factual correctness) when evaluating a RAG system's answer?

Accepted Answer

Faithfulness = grounded in retrieved context; factuality = true in the real world; they can diverge

Answer

They are synonyms in RAG evaluation

Answer

Faithfulness needs gold labels; factuality does not

Answer

Factuality only applies to multiple-choice tasks

Question 11

Why is BLEU/ROUGE inadequate for evaluating long-form generative LLM answers, and what is a better alternative?

Accepted Answer

They reward n-gram overlap and penalize valid paraphrases; use LLM-judge/claim-level checks instead

Answer

They are too slow for production use

Answer

They require GPU acceleration

Answer

They only work for translation, never summarization

Question 12

What is 'sycophancy' in LLMs and why is it a concern for evaluation?

Accepted Answer

Tendency to agree with the user's stated view even when wrong; test with leading/adversarial prompts

Answer

Generating overly long answers

Answer

Refusing all controversial questions

Answer

Repeating the prompt verbatim

Question 13

How does the FActScore metric evaluate factual precision of long-form generations?

Accepted Answer

Percentage of atomic facts in the output supported by a reliable knowledge source

Answer

BLEU score against a Wikipedia article

Answer

Whether the overall answer is judged true or false

Answer

Average token logprob of factual tokens

Question 14

What is the main risk of using the same model family as both generator and LLM judge?

Accepted Answer

Self-preference bias and shared blind spots inflate scores and miss shared hallucinations

Answer

It is computationally cheaper but less accurate on latency

Answer

The judge cannot parse the generator's output format

Answer

It always produces ties

Question 15

Why does increasing retrieval recall not always reduce hallucinations in a RAG pipeline?

Accepted Answer

More passages add noise/distractors and lost-in-the-middle effects, offsetting recall gains

Answer

Higher recall always lowers hallucinations linearly

Answer

Recall has no effect on generation quality

Answer

Recall only matters for embedding models

Question 16

What is the 'lost in the middle' phenomenon and why does it matter for long-context evaluation?

Accepted Answer

Models underuse evidence in the middle of long contexts; evaluate by varying evidence position

Answer

Models crash when context exceeds the window

Answer

Middle tokens get higher attention than edges

Answer

Only applies to encoder-decoder models

Question 17

How should you evaluate an LLM's ability to abstain ('I don't know') rather than hallucinate?

Accepted Answer

Use unanswerable items and measure selective accuracy/coverage and penalized confident errors

Answer

Count total tokens generated

Answer

Measure only accuracy on answerable questions

Answer

Check whether the model ever says 'I don't know' at all

Question 18

What does Expected Calibration Error (ECE) measure for an LLM classifier or QA model?

Accepted Answer

Weighted average gap between confidence and accuracy across confidence bins

Answer

The variance of model logits

Answer

Difference between train and test accuracy

Answer

Entropy of the output distribution

Question 19

Why is inter-annotator agreement (e.g., Cohen's/Fleiss' kappa) important when building a human eval set for hallucination, and what does low agreement imply?

Accepted Answer

It validates label reliability; low kappa signals an ambiguous rubric and noisy ground truth

Answer

It measures model accuracy directly

Answer

It only matters for multiple-choice tasks

Answer

Low kappa means the model is hallucinating more

Question 20

What is the difference between pointwise (absolute) scoring and pairwise (preference) evaluation, and when is each preferred?

Accepted Answer

Pairwise compares two responses (more reliable, costlier to rank); pointwise gives absolute scores

Answer

Pointwise is always more accurate than pairwise

Answer

Pairwise requires no aggregation method

Answer

They produce identical rankings always

Question 21

How does the Chatbot Arena / Elo (Bradley-Terry) approach evaluate LLMs, and what is its main limitation?

Accepted Answer

Fits Bradley-Terry to human pairwise votes; limited by prompt skew and style-over-correctness bias

Answer

Measures perplexity on a fixed corpus

Answer

Uses only multiple-choice accuracy

Answer

Requires gold reference answers for every prompt

Question 22

What is length/verbosity bias in LLM-as-a-judge evaluation, and how do you control for it?

Accepted Answer

Judges favor longer answers; use length-controlled win rates and rubric/length normalization

Answer

Judges favor shorter answers; pad responses to fix it

Answer

Length has no measurable effect on judges

Answer

Only humans, never LLM judges, show this bias

Question 23

Why are RLHF reward models prone to 'reward hacking' / overoptimization, and how is it detected during evaluation?

Accepted Answer

Policy exploits reward-model errors; detect via divergence between proxy reward and gold/human eval

Answer

Reward models always undertrain; detect via low train loss

Answer

It is caused by tokenizer mismatch; detect via vocab size

Answer

It only occurs with DPO, never PPO

Question 24

What evaluation strategy helps distinguish whether a RAG error originates from retrieval vs. generation?

Accepted Answer

Measure retrieval metrics (recall@k) and generation faithfulness separately to localize the error

Answer

Only measure end-to-end answer accuracy

Answer

Compare embedding norms of query and documents

Answer

Check the model's parameter count

Question 25

What is the 'generation-verification gap' between a model's ability to verify a claim and its ability to generate correct claims, and why does it matter?

Accepted Answer

Verification is easier than generation; this justifies generate-then-verify hallucination-reduction pipelines

Answer

Generation is always more accurate than verification

Answer

The gap only exists in vision models

Answer

Verification and generation accuracy are always equal

Question 26

How does Chain-of-Verification (CoVe) reduce hallucinations, and what is its key failure mode?

Accepted Answer

Draft, self-generate and answer verification questions, then revise; fails when the model repeats its own error

Answer

It retrains the model on each query; fails on cost

Answer

It uses a knowledge graph; fails on coverage

Answer

It samples at temperature 0; fails on determinism

Question 27

Why should evaluation report confidence intervals and significance tests rather than single accuracy numbers when comparing LLMs?

Accepted Answer

Scores carry sampling/finite-set variance; CIs and paired tests prevent overclaiming small gaps

Answer

Confidence intervals are only needed for regression tasks

Answer

Single numbers are always sufficient for benchmarks

Answer

Significance testing requires retraining the model

Question 28

What is the role of an NLI (natural language inference) model in automated hallucination/groundedness detection?

Accepted Answer

It checks whether each answer claim is entailed by the source context to flag unsupported claims

Answer

It generates the answer from the context

Answer

It measures embedding cosine similarity only

Answer

It ranks retrieved passages by BM25

Question 29

What is 'goal/specification gaming' versus genuine capability when evaluating agentic LLM systems, and how do you guard against it?

Accepted Answer

Agent satisfies the literal success signal without solving the task; guard with hidden tests and trajectory checks

Answer

It is a synonym for hallucination

Answer

It only occurs in RAG systems

Answer

It is prevented by lowering temperature

Question 30

Why is the temperature/decoding setting a confounder in LLM evaluation, and how should it be handled?

Accepted Answer

Decoding params affect diversity/hallucination/variance; fix, report, and average over samples

Answer

Temperature has no effect on factual accuracy

Answer

Always use temperature 1.0 for fair comparison

Answer

Decoding settings only matter for code tasks

Test yourself