Some labs do it internally because RLVR is very token-expensive. But it degrades CoT readability even more than normal RL pressure does.

It isn't free either - by default, models learn to offload some of their internal computation into the "filler" tokens. So reducing raw token count always cuts into reasoning capacity somewhat. Getting closer to "compute optimal" while reducing token use isn't an easy task.

Yeah the readability suffers, but as long as the actual output (ie the non-CoT part) stays unaffected it’s reasonably fine.

I work on a few agentic open source tools and the interesting thing is that once I implemented these things, the overall feedback was a performance improvement rather than performance reduction, as the LLM would spend much less time on generating tokens.

I didn’t implement it fully, just a few basic things like “reduce prose while thinking, don’t repeat your thoughts” etc would already yield massive improvements.