ArxivRollBench

A rolling benchmark turning freshly-published arXiv text into sentence-ordering, cloze and next-fragment multiple-choice tasks every six months, built to measure how much contamination inflates benchmark scores.

Also known as: ArxivRoll, RoBench

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categoryreasoning
Subcategoryrolling, contamination-auditing sentence-ordering, cloze and continuation tasks built from recent arXiv text
Page statusactive
Metricaccuracy (per SCP-type, per domain, per release); the paper's own headline metric is Rugged Score (RS), a derived overestimation measure, not this page's per-item accuracy
Directionhigher_is_better
Unit%
Dataset licenceNot established. No LICENSE file was detected by GitHub's own licence API for the liangzid/ArxivRoll source repository, and the sampled Hugging Face dataset configs carry no licence tag. The underlying text is drawn from recent arXiv papers, each subject to arXiv's own non-exclusive distribution terms or the author's chosen licence; the repository's own README explicitly instructs contributors not to redistribute paper source files beyond what each paper's licence permits.
PublisherThe Hong Kong Polytechnic University

What it measures

ArxivRollBench is the evaluation-task component of ArxivRoll, a framework whose actual goal is not to test topical arXiv knowledge but to measure how much a model's public-benchmark score is inflated by contamination or uneven training exposure. It does this with SCP tasks -- Sequencing (reorder three shuffled sentences, choosing among four candidate orderings), Cloze (fill three masked sentences in a paragraph from labelled candidates, again choosing among four candidate combinations), and Prediction (choose the correct next fragment of a passage from four options) -- built automatically from arXiv article text collected after each task set is constructed, across eight domains (cs, econ, eess, math, physics, q-bio, q-fin, stat). Because the source text postdates every model's training cutoff at the time a round is built, and each round is evaluated once and then either kept private or later published as "expired," the authors use performance on ArxivRollBench relative to performance on long-public benchmarks to compute a "Rugged Score," a measure of public-benchmark overestimation -- inspired explicitly by one-time-pad encryption, where reusing the same key (here, the same test items) twice defeats the scheme's guarantees.

Task format

Four-way multiple choice per item. Sequencing and Cloze items ask the model to answer "Selection 1" through "Selection 4"; Prediction items ask for a letter A-D. All are zero-shot, generation-based (the model's free text is parsed by regex for a selection number or letter), scored per domain, per release and per SCP type.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub