In reinforcement learning with verifiable rewards, a benchmark's verifier supplies the reward. Reward-hacking audits, the paper notes, mostly look for runs that pass without doing the task. That approach misses a different kind of failure.

The authors point to a loophole: a run can do the task, cause a side effect, and still be considered successful by the verifier. The abstract does not yet describe the proposed method, but the framing suggests a need to broaden audits beyond task completion alone.

The paper is available on arXiv under identifier 2610.09142.