>"illegible reasoning in a few reinforcement-learning environments over long rollout"
Yet, I get the point that you're making: those tokens essentially are an internal scratchpad for the LLM which isn't required to logically lead to the output.
You are expecting everyone to be as aware of these issues as you are. I don't think white supremacy has as large of a share of the mindspace in south korea as you think - not every issue (even if it has a wikipedia page!) is burned into everyone's mind across the globe. But I am happy to hear that every time you pick a username, you make sure it is free of dog whistles in all cultures and that the letters and numbers cannot be picked out or rearranged in a way that hints at one.
>"illegible reasoning in a few reinforcement-learning environments over long rollout"
Yet, I get the point that you're making: those tokens essentially are an internal scratchpad for the LLM which isn't required to logically lead to the output.
This video presentation of the paper you linked was interesting: https://www.youtube.com/watch?v=hUp3zh23aHw