Your reasoning model isn't dumb. Your parser is throwing away its best answers.

I benchmarked a vision-language model and scored it at 0.31. The real number was 0.70. Same model,...