Comment by danielmarkbruce

Read the paper. They train RLCR on existing big math problems. They subtract a brier score penalty from the correctness reward. No new confidence labels are needed.

Existing datasets, different reward function.