Finetuning a model to identify reward hacks reduces the frequency of those same hacks in its own outputs. This suggests that teaching a model to judge misalignment also improves its internal alignment. LessWrong researchers found this dual-purpose training eliminates the need for separate alignment runs.