Skip to content
Modelwire
Subscribe

Partial-prompt contamination tests show severe false-positive bias

Source published ·Modelwire updated

Original coverage: arXiv cs.CL ↗·How Modelwire adds context

The development

Researchers have exposed a critical flaw in how the AI community tests for benchmark contamination in large language models. By auditing seven open-weight models against MATH-500 and recent AIME problems, they found that partial-prompt completion tests produce wildly inconsistent results depending on how much of the problem remains unseen. Short leftovers trigger false-positive memorization flags at rates up to 6 times higher than longer ones, and even exact-match thresholds flag 8.6% of genuinely unseen problems. This undermines confidence in current contamination audits and suggests the field needs more rigorous methodology before claiming models are truly uncontaminated.

Modelwire’s AI-generated summary of coverage from arXiv cs.CL.

Modelwire analysis

Skeptical read

Our AI-generated reading of the wider context and the next developments to watch.

The real problem isn't that models are contaminated; it's that the testing method itself produces unreliable verdicts depending on how you configure it. The researchers aren't catching new contamination, they're catching that the field's contamination detector is broken.

This joins a pattern of evaluation infrastructure failures documented across Modelwire's recent coverage. Like the policy-benchmark audit from November 4th that found scoring logic errors inflating false negatives by 80 points, and the privacy de-identification study from November 3rd showing verdicts shift based on testing parameters, this work reveals that benchmark quality directly determines measured capability. The Anthropic safety auditing piece from November 7th also surfaces the same tradeoff: detection sensitivity creates false alarms that overwhelm reviewers. Here, the tradeoff is inverted but identical in structure: lower thresholds flag genuine negatives as positives, higher thresholds miss real contamination. The field has been drawing conclusions about model integrity from tools it hasn't validated.

If the same seven models are re-audited using the researchers' proposed methodology and contamination rates drop below 2 percent, that confirms current audits are producing noise rather than signal. If major labs (OpenAI, Anthropic, Meta) adopt this framework for their next model releases before Q2 2027, the finding has shifted practice; if they continue using partial-prompt tests without modification, the work remains a citation without consequence.

This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error

Coverage behind this analysis

These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.

  1. ·arXiv cs.CL

    Benchmark bugs mask model competence in policy-based reasoning tasks

    Researchers have uncovered systematic flaws in nine widely-used policy-based decision benchmarks that evaluate whether language models correctly apply written rules to real-world cases. The audit reveals that benchmark scorers themselves contain errors, reference mistakes, and logic gaps that cause correct model outputs to be marked wrong. On LegalBench SARA alone, fixing the evaluation logic improved…

    Read Modelwire coverage →Original source ↗

MentionsQwen2.5-7B-Instruct · MATH-500 · AIME 2025 · AIME 2026 · ROUGE-L

MW

How this coverage is produced

Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “How Much Is Left Matters: False Positives and False Negatives in Partial-Prompt Contamination Tests”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.

More to explore

Study reveals how LLMs fake economic forecasting through training-data leakage

arXiv cs.CL·

Hard vision benchmarks accumulate label errors at predictable rates

arXiv cs.LG·

Persistent AI agents violate embedded rules 35 percent more often

arXiv cs.CL·
Partial-prompt contamination tests show severe false-positive bias · Modelwire