Preprint
Oct 2026
SimHack: Predicting the onset of detectable reward hacking before RL training
When training LLMs to code with reinforcement learning, reward hacking is often detectable but difficult to prevent, since (a) environment hardening is increasingly difficult as training environments grow more complex; and (b) training-time interventions incentivize models to obfuscate hacking behavior to bypass detect...
Jun-Wei Deng, Alex Bie, N. Ponomareva
· 1 citation