Loading today’s stories
Build1 distinct publisher2 min readPublished
A paper rebuilt a hidden holdout for TruthfulQA and found some of twenty models score as much as 16 points higher on the public version. A flat discount will not repair your shortlist.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
A holdout earns its name from two properties held at the same time: it comes from the same distribution as the target set, and it was hidden while the model was being built [13]. The first is testable after the fact. The second is not, strictly, and the authors say as much: a clean proof of a performance gap would need an IID split that you know could not have influenced any aspect of model development [14]. A question set written later is a reconstruction of that condition rather than the condition itself. That is the seam in the result, and it is where a vendor whose score moves will push.
Take the gap at face value anyway and the awkward part is its shape. It is a maximum drawn from twenty models on one target benchmark [15], reported as what some models do and not as what the median model does [4]. That rules out the correction an operator would actually want, which is a fixed haircut on published scores. If every model were padded equally, leaderboards would still rank correctly and would still be usable for narrowing a field. Padding that varies by model reorders the field, and a public score carries no information about how much of it belongs to your candidate.
The paper's account of why this persists is worth more than its arithmetic. Its starting premise is that training data for many models is contaminated with test data, so the public benchmarks are compromised [1], and it reads the accumulated evidence of evaluation data sitting in training corpora as a sign that the gaming is already underway [7]. Alongside that, it notes incentives for higher scores strong enough that optimising benchmark performance can take precedence over real-world effectiveness and safety [12]. Contamination on that reading is not a defect to be patched but the expected output of the structure the leaderboards set up.
The release creates a shelf-life problem of its own. Retro-Misconceptions is now published [3], which means the hiddenness property stops holding for anything trained after that point [16], and the measurement it supports decays with every subsequent training run. A retro-holdout is a consumable. Anyone reusing this one for an internal decision is buying a number whose validity is bounded by the candidate model's training cutoff, and that boundary is not visible in the score itself.
Ranked by verification strength, evidence, and original report placement.
The authors introduce a systematic methodology for retrospectively constructing a holdout dataset for a target dataset, demonstrating the statistical indistinguishability of that retro-holdout, and comparing LLMs on both datasets to quantify the performance gap due to the target dataset's public availability.
The authors find that some of the evaluated models have inflated scores by as much as 16 percentage points.
The paper states that post-hoc construction of sufficiently similar datasets is non-trivial.
The paper states that the training data for many LLMs is contaminated with test data, meaning public benchmarks used to assess LLMs are compromised and there is a suggested gap between benchmark scores and actual capabilities.
Applying the methods to TruthfulQA, the authors construct and release a retro-holdout named Retro-Misconceptions and evaluate twenty LLMs on it.
Holdout datasets for benchmarks are typically not available because benchmark developers usually release all evaluation data, with notable exceptions.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One self-labeled preliminary preprint with a released artifact
The methodology, the released dataset, and the twenty-model comparison are all stated directly in the source, and the statistical-indistinguishability tests are described rather than merely asserted. But the entire cluster rests on a single arXiv preprint that carries the keyword 'Preliminary', covers one target benchmark, reports an upper bound without per-model numbers in the supplied excerpt, and has no independent replication.
Artifact released, uptake only author-side
There is real, dated artifact activity - a public dataset release and a twenty-model evaluation run - but all of it is by the authors. The supplied sources show no third-party use of Retro-Misconceptions, no lab adopting retro-holdouts in reporting, and no leaderboard integration, and the paper notes the artifact's validity is bounded to pre-2024 training cutoffs.
Framing outruns a preliminary single-benchmark result
The measured gap itself is modestly stated ('some... as much as 16 percentage points'), but the interpretive framing goes further than the evidence: the introduction says results 'conclusively indicate' evaluation gaming while the abstract only claims scores 'do not always accurately assess model properties' and the paper is keyworded Preliminary. One benchmark, one retroactively built holdout, an upper-bound statistic, and the paper's own admission that a definitive test would require a never-exposed IID split all argue for a positive but not extreme overstatement score.
Named developer score incentives; author stake in the method
The source itself documents one side of the incentive picture: strong rewards for higher benchmark scores that could put leaderboard optimization ahead of real-world effectiveness and safety. On the other side, the authors are advocating a method and dataset they created, so a positive finding of evaluation gaming supports their contribution - visible in the escalation to 'conclusively indicate'. The supplied excerpt discloses no funding, affiliation, or competing-interest details, so this reading stops at what the text shows.
Clear primary text, narrow and unreplicated base
Confidence in what the paper says is high because the source is the primary document and its key statements are verbatim. Confidence in the finding's generality is moderate at best: one publisher, one benchmark, preliminary status, no per-model detail, no external replication, and a dataset whose validity window is already bounded.
build
Multi-agent LLM gains largely vanish once the thinking-token budget is held constant1 distinct publisher
science
GJ 523b gives 'Mega-Earth' a number: 23 Earth masses inside 2.5 Earth radii1 distinct publisher
build
A prompt A/B on six inputs is a coin flip until you have measured the noise floor1 distinct publisher
security
Mandiant found 100 high-severity bugs in two days. Plan for the other side doing the same.1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 26, 2026