Proposed AI training approach lets models reward-hack, then disables the parameters that learn it

On Oct. 19, _jake_ward described a paper proposing that models reward-hack during reinforcement learning while the behavior is directed into specific parameters. He claims disabling those parameters for deployment leaves the model with new capabilities but no reward hacking. In a reply, siddarthv66 questioned whether reward hacking can be detected reliably during training and warned that missed hacks could still update a “retain adapter.”