> I'm just wondering why no one tried to RL a model on stuff like "less LOC" and "less overengineering"

Anthropic looked into this and the answer is because it makes the model more stupid

https://www.anthropic.com/research/evaluating-feature-steeri...