Yeah, even RL tuning can be applied to improve reasoning paths for specialized applications, something like RLAIF by anthropic but for specialized agents