Hallucination benchmarks such as TruthfulQA and HaluEval are published in full, questions and reference answers together. The web crawls that train the next generation of models then pick them up. A model that "resists" a common misconception on these benchmarks may simply have read the benchmark's answer key. This matters because hallucination scores are used as safety evidence. If part of a model's apparent truthfulness is recalled benchmark text, the number overstates how reliable the model will be on questions nobody has written down yet.
We propose to measure that overstatement directly, item by item. For each item in a static hallucination benchmark, we build two new "twin" items. The first asks about the same fact in a new form. The second asks about a different fact of the same kind and difficulty. We then report the gap between a model's accuracy on the originals and on the twins. The hard part is producing twins whose reference answers are actually correct. Doing this automatically fails for the same reason hallucination happens: the drafting model and the judging model get the same facts wrong. Human experts therefore check every twin, working from model-drafted suggestions, following evidence that such suggestions speed up expert checking without making experts rubber-stamp them. We validate the whole measurement against deliberately injected contamination, so we know the true inflation.
Core Claim
The claim is that the static-minus-twin accuracy gap, measured on human-verified twins, recovers how much contamination inflates hallucination scores. We expect it to do so more accurately than two cheaper routes: existing contamination detectors applied to the original items, and fully automated twin generation. We also claim the human verification step can be made cheap enough to repeat for each new model release.
Contamination here means that a specific benchmark item and its answer were seen in training. It does not mean the model knows the underlying fact, and the design keeps these two apart. Asking about events after a model's training cutoff, the usual temporal defense, confuses "never saw the item" with "cannot know the fact." A post-cutoff question measures refusal behavior, not truthfulness. Twins keep the fact inside the model's plausible knowledge and change only the item.
Where This Sits in Prior Work
Three benchmarks motivate the gap, and none of them covers it.
TruthfulQA measures whether models repeat human misconceptions across 817 questions. It reports the best GPT-3 model truthful 58% of the time against 94% for humans. It is static. Benchmark Probing later tested it by masking words in its questions and asking models to fill the gaps. Giving the model the benchmark's URL as a hint raised GPT-4's accuracy on this masked-word guessing from 17.53% to 29.49%, and Claude-instant-1's from 22.28% to 42.19%. That is evidence of leakage, but it says nothing about how much the leakage inflates the truthfulness score itself.
HaluEval combines ChatGPT generation with human span annotation to build 35,000 static items derived from public datasets. It also shows that ChatGPT detects hallucinations in summaries with only 58.53% accuracy. That weakness undercuts any plan to let an LLM verify its own benchmark items.
HalluLens avoids leakage by generating questions from Wikipedia at evaluation time. Its automated checking of long answers fails to retrieve evidence about 15.4% of the time, and it cannot say whether a correct answer came from recall or reasoning. We use its generated PreciseWikiQA items as a negative control where the twin gap should be near zero. We do not treat it as a replacement for auditing the static benchmarks that the literature keeps reporting.
Contamination detection supplies our strongest competitors, and it also constrains the design.
Generalization or Memorization introduces CDD, which flags contamination when a model's sampled outputs collapse onto nearly identical strings. It reports detection gains of 21.8% to 30.2% over n-gram and LLM baselines. It also shows that simple overlap filters miss rephrased contamination. This warns us that same-fact paraphrase twins may themselves be contaminated, which is why we also build different-fact "sibling" twins.
PAC is a black-box calibration method based on word swaps. CAP measures whether performance stays consistent when questions are perturbed. Both flag items or datasets as leaked. None of the three detectors estimates the size of score inflation on open-ended truthfulness questions, and that is the quantity we measure.
The human verification method comes from Schulz et al.. They found that model suggestions sped up expert annotation of hard, discourse-level tasks by up to 35%, and that experts corrected wrong suggestions rather than blindly accepting them. Rinaldi et al. similarly restrict curator attention to validating high-probability candidates.
de Wynter and LiveBench supply the reason not to drop the human step. De Wynter documents LLM-judged evaluations with no human cross-check. LiveBench finds that LLM judges fail on problems they cannot solve, and it restricts itself to answers that can be checked by a program. That restriction leaves out exactly the open-ended factual answers studied here.
Research Program
Twin construction. We build twins for three sources:
- TruthfulQA: all 817 items.
- HaluEval QA: a stratified sample of 1,000 items derived from HotpotQA.
- Fabricated references: the ACM Computing Classification System reference-title task from Agrawal et al., where expert annotators agreed with search-engine ground truth 99–100% of the time. That agreement makes it a clean, low-ambiguity anchor.
Each original item gets two twins. The paraphrase twin asks about the same fact or misconception with a different question structure and newly written reference answers, so the item's surface form is new. The sibling twin targets a different misconception or entity from the same category and has comparable popularity. It is sourced from reference material that the benchmark did not use. For fabricated references, the sibling is a different subject area. A strong LLM drafts candidate twins together with retrieved evidence.
Human verification. Domain-competent annotators, two per twin, each record three things:
- whether the gold answer is correct given the evidence;
- whether the twin tests the same skill as the original;
- their own answer to the twin, given without seeing the gold.
The third record gives a human difficulty measurement. TruthfulQA reports 94% human truthfulness on the originals, so twins should land near that figure. Twins far from it are revised or dropped.
To test for automation bias, we plant drafts with known-wrong gold answers in the annotators' queue. Schulz et al. did not include this test. We log time per item for three conditions: with suggestions, without suggestions, and suggestions with confidence-based routing. In the routing condition, humans see only twins where the automated verifier and the judge model disagree or where retrieval failed. The remaining twins are spot-checked at a fixed rate. This routing condition is our cost-reduction proposal. Its label error is measured against the full double annotation.
Measurement. For model $m$ and benchmark $B$, the inflation estimate is
$$\Delta_m = \mathrm{Acc}_m(B) - \mathrm{Acc}_m(T_B) - \delta_B$$
where $\mathrm{Acc}_m(B)$ is truthfulness accuracy on the original items and $\mathrm{Acc}_m(T_B)$ is accuracy on their human-verified twins. The term $\delta_B$ is the leftover difficulty mismatch. We estimate $\delta_B$ from models whose training data predates the benchmark's public release, since such models cannot have memorized its items. We also cross-check it against the human difficulty gap. Confidence intervals come from a paired bootstrap over items.
Responses are scored with each benchmark's native protocol where one exists. For TruthfulQA this is its trained judge, which reports 90–96% agreement with humans. We also have humans judge a fixed audit subsample of responses, because a judge tuned on the original items may itself not transfer to twins.
Baselines
The naive reference point is the raw published-protocol score on each static benchmark, which implicitly assumes $\Delta_m = 0$. Every model is run with identical prompts and decoding on originals and twins. If the naive score and the twin score agree within bootstrap error across models, the proposal's premise is false, and we will report that.
The strongest existing competitor is per-item contamination detection with CDD, from Generalization or Memorization. Its reported advantage is a 21.8% to 30.2% detection improvement over n-gram and LLM baselines on its own contamination benchmarks. We will also run PAC, which reports AUC gains of 4.5% on WikiMIA and 5.9% on StackMIA, and TS-Guessing. To make the comparison fair, each detector's flags are converted into an inflation estimate by rescoring the benchmark with flagged items removed. Each detector gets the same models and sampling budget as the twin method, and its threshold is tuned on the same held-out slice of the controlled-contamination split described below. There is no reported result found on Lacuna for any detector's accuracy at estimating score inflation on TruthfulQA or HaluEval, so the comparison is new rather than a rerun of a published number. We count the twin method as beating the detectors if its inflation estimate has lower absolute error against the injected ground truth, with a paired bootstrap interval excluding zero.
The second competitor is fully automated twins in the style of HalluLens: the same LLM drafts, verified by retrieval plus an LLM judge, with no human involvement. It is run on identical source items and uses the same drafting model. The two pipelines differ only in the human step. HalluLens reports a 15.4% retrieval failure rate on long-form answers, and HaluEval reports 58.53% LLM detection accuracy on summaries. These are the numbers we expect human verification to improve on. It must reduce twin label error, and with it the error of the inflation estimate, by more than the width of the bootstrap interval.
Two ablations isolate the key ingredients. The first is human verification without suggestions. It tests whether suggestions buy speed without introducing planted-error acceptance, measured against Schulz et al.'s reported speed-up of up to 35%. The second drops sibling twins. It tests whether paraphrase twins alone are enough, or whether, as CDD's result on rephrased contamination implies, they are themselves partly memorized.
Why not just…?
Why not just run contamination detectors on the original items and discard what they flag? This is the most serious threat, because it needs no new data. Its weakness is that detectors such as CDD and PAC were validated on code, math, and membership-inference data, where memorized outputs are near-verbatim strings. Truthfulness answers are short, open-ended, and naturally low in variety even when not memorized. A correct answer about why the sky is blue collapses onto a few phrasings either way, so collapse-based signals may flag honest knowledge. We do not assume this. The controlled-contamination experiment measures it directly. If detector-filtered scores match the injected inflation as well as twins do, we will recommend detectors and report the human pipeline as unnecessary for this purpose.
Why not just generate fresh items automatically, as HalluLens does? Automatic generation removes contamination but introduces label error exactly where hallucination lives. The drafting model can write a wrong gold answer that the judge model shares, so the automated pipeline scores a truthful answer as a hallucination. HalluLens's 15.4% retrieval failure rate and HaluEval's 58.53% self-detection figure make this likely, but they come from different tasks. The automated-twin baseline above tests it on exactly our items. If automated twins' label error is small enough that inflation estimates match the human-verified ones, the human step can be dropped. Only the routing condition would then remain, as an inexpensive audit.
Why not just evaluate on questions about events after the training cutoff? Post-cutoff questions are uncontaminated because the model cannot know the answer. That turns a truthfulness test into a test of refusal. HalluLens already shows how different these are: Llama-3.1-405B falsely refused 56.77% of valid questions. Twins keep the fact within reach and change only the item. For completeness we include a small post-cutoff set, to show that it produces a gap driven by missing knowledge rather than by memorization.
Proposed Experiments
The decisive check is controlled contamination. We take open-weight models from the families evaluated in HalluLens, including Llama-3.1 and Mistral. We fine-tune copies on a random half of TruthfulQA and HaluEval items with their reference answers, at several exposure counts. This produces known label-level contamination on one half and none on the other.
For each contamination level, the twin estimate $\Delta_m$ should rise with exposure count on the contaminated half and stay near zero on the clean half. We compare its error in recovering the true inflation against the detectors and against automated twins. The true inflation is the contaminated-half score minus the same model's pre-fine-tuning score on those items.
The negative control is HalluLens PreciseWikiQA, where items are generated fresh and $\Delta_m$ should be indistinguishable from zero. A nonzero gap there would expose a mismatch between twins and originals that the difficulty correction failed to remove.
We then apply the validated method to public and commercial models and ask two questions:
- Does $\Delta_m$ on TruthfulQA track that model's TS-Guessing leakage signal?
- Does contamination adjustment reorder model rankings, in particular the inverse-scaling pattern TruthfulQA reported for larger models?
The main cost result is the time per verified twin for each human condition, and the label error of the routing condition against double annotation. These figures tell us whether the audit can be repeated for each model release, rather than costing the 5.5 expert hours per task that DEEPSYNTH required.
References
- TruthfulQA: Measuring How Models Mimic Human Falsehoods
- Benchmark Probing: Investigating Data Leakage in Large Language Models
- HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Models
- HalluLens: LLM Hallucination Benchmark
- Generalization or Memorization: Data Contamination and Trustworthy Evaluation for Large Language Models
- Data Contamination Calibration for Black-box LLMs
- CAP: Data Contamination Detection via Consistency Amplification
- Analysis of Automatic Annotation Suggestions for Hard Discourse-Level Tasks in Expert Domains
- Large-scale Information Extraction for Assisted Curation of the Biomedical Literature
- Awes, Laws, and Flaws From Today's LLM Research
- LiveBench: A Challenging, Contamination-Limited LLM Benchmark
- Do Language Models Know When They're Hallucinating References?
- A Benchmark for Deep Information Synthesis