It sounds like you're proposing doing this operation on the tokens in the reasoning. While it would be interesting to know if allowing it to choose arbitrary to

It sounds like you're proposing doing this operation on the tokens in the reasoning. While it would be interesting to know if allowing it to choose arbitrary tokens, the biggest issue is that there's quite a bit of evidence that the tokens it prints have only a loose relationship with the internal model processes. I question your premise; first demonstrate that having it think aloud in "efficient mathematical representations" is a useful efficiency. Then you can demonstrate that you can do any interpretatability work on the output.