Do RLVR benefits extend from math/coding to MCQA?

Determine whether the benefits of reinforcement learning from verifiable rewards (RLVR) observed on mathematics and coding tasks extend to multiple-choice question answering (MCQA), which features a much smaller and constrained answer space compared to open-ended math and coding outputs.

Background

Prior work on RLVR has primarily targeted mathematics and coding, tasks with large, open-ended answer spaces where emergent reasoning has been observed. Multiple-choice question answering (MCQA), including medical MCQA, differs substantially in that the model selects from a small set of predefined options, raising uncertainty about whether RLVR’s demonstrated advantages transfer to this more constrained setting.

This paper investigates RLVR in medical MCQA, noting the fundamental difference in answer space and reasoning demands. The authors explicitly state that it is unclear whether the benefits demonstrated in math and coding tasks will carry over to MCQA, motivating their empirical study.