Auditing Reward Hackability in Code RL Training Environments

Shreshth Rajan

We measure the rate at which code RL environments accept incorrect solutions as correct. On a 49-task sample of SWE-bench Verified, 28.5% of tasks have test suites weak enough that…