There's no reward function in the sense that optimizing the reward function means the solution is ideal. There are objective criteria like 'compiles correctly'

There's no reward function in the sense that optimizing the reward function means the solution is ideal. There are objective criteria like 'compiles correctly' and 'passes self-designed tests' and 'is interpreted as correct by another LLM instance' which go a lot further than criteria that could be defined for most kinds of verbal questions.