Explain the discrepancy in the EN test results

Determine the cause of the discrepancy between the English test metrics reported at upload time and the final leaderboard metrics for the same OCR-grounded ensemble submission.

Background

The paper reports that the English results for the final ensemble differed between the values returned at upload time, 0.460 Cor and 0.347 Cor-lbl, and the final leaderboard values in Table 3, 0.466 Cor and 0.357 Cor-lbl. The corresponding Italian and Chinese results agreed across the two reporting views, leaving the source of the English-only discrepancy unresolved.

References

The EN figures for our final ensemble differ between the values returned to us at upload time ($0.460$/$0.347$) and the final leaderboard values in Table~\ref{tab:submission} ($0.466$/$0.357$); the IT and ZH figures for the same submitted file agree to four decimals across the two views. We cannot account for the EN difference.

— Two-Token Features and Small-Large Ensembles for VLM Hallucination Detection  (2609.10244 - Schwartz, 9 Sep 2026) in Appendix, Section 'Extended Limitations', paragraph 'Reproducibility of the EN test row'