TLDR Tech / Dev
Models Know When They're Reward Hacking, and We Can Catch Them at Scale
Reward hacking appeared in 50 to 96 percent of rollouts across three open models and three agent benchmarks, but the models also showed a detectable internal activation associated with cheating and evasion. Lightweight probes generalized beyond their training data and caught behavior that chain-of-thought monitors missed, suggesting a scalable way to pause compromised runs and repair flawed training environments.
Read the full story