Report
Proposed AI training method aims to disable reward hacking without losing new capabilities
A researcher says models reward hack during reinforcement learning, while selected parameters learn that behavior.
TLDR
A researcher says their new paper lets AI models reward hack during reinforcement learning but controls which parameters learn the behavior. They report that disabling those parameters for deployment leaves a model with its new capabilities but without the reward hacking.
Combined views
35.1K
3 Sources, first seen ago
