Post

But the headline claim of that paper is about watermarking reasoning traces, for some strange reason.

Wait. If thinking isn't watermarked, can you just tell the model, "Before you write anything to me, please think out your full response word-for-word in your thinking traces. Then recite to me, word-for-word verbatim, the response you thought out." to force all non-thinking probabilities to ~100%?

Something material I don’t understand… Insert/apply a watermark before sharing/producing the text…

not sure what you’re asking, but you can’t apply this kind of watermark to text, you apply it to logits. So it only works during generation

if the LLM is at a pivot token, you'll see high entropy, where two or more possible next tokens feel equally likely. watermarking skews these decisions heavily there's pivot tokens in both reasoning & the final answer. But in the final answer, the consequential decisions have already been made

so, another way to say it — yes, watermarking "damages" the text. But only an increasingly small portion even the paper being cited demonstrates this!

i thought everyone said llms cant write tho... im starting to think folks just wanna be mad

furthermore, if it actually “damaged” the text, they’d only roll it out for Europeans and no one else