It’s actually pretty bad at it. I think there just isn’t enough literature to get a good effect from the LLM approach. Good luck with verifiable rewards when

It’s actually pretty bad at it. I think there just isn’t enough literature to get a good effect from the LLM approach. Good luck with verifiable rewards when the target discourse is effectively pure self-criticism. Maybe ‘disputable rewards’ … I have found that with proper framing I can get good help from Claude and ChatGPT on questions of translation of haute German philosophy and, to my amazement, Ancient Greek. An immediate ‘translate this passage’ request is a cataclysmic disaster. The nexus of sentences differs from other forms of discourse.