Update Qwen2.5-VL-3B mmmu_pro reference for sample 16
With current dependencies the model answers D instead of B on sample 16
(gold is C), so extractive_match stays 0 and aggregate scores are unchanged.
The drift was hidden until now because the slow tests crashed on xxhash.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
