I happened to come across this while searching for my own LessWrong post: 5 ways to improve CoT faithfulness. I was wondering if you saw the “canonical paraphrasing” idea in that doc.
If either of you is working on something along these lines, I’d be curious to hear about it!
I happened to come across this while searching for my own LessWrong post: 5 ways to improve CoT faithfulness. I was wondering if you saw the “canonical paraphrasing” idea in that doc.
If either of you is working on something along these lines, I’d be curious to hear about it!
Rob / Poseidon (https://poseidonresearch.org/) have done very good work on it. Probably worth getting in touch with him.