This article is wrong. SynthID-style watermarking does not affect one-shot generation quality. When news about the watermarking went around a couple of months ago I too thought there must be a quality/detectability tradeoff (as there is for images), but I dug into the math more and realized that my intuition was incorrect.
They key idea is this: changing the random seed for LLM generation definitionally does not change the distribution of tokens in expectation over all seeds.
This can already be used to embed information in LLM streams, leading to LLM steganography work (https://arxiv.org/abs/2606.09135, but the idea dates back much further). Take your target message, whiten it via a one-time pad (stream cypher), then use it as the 'compressed stream' under an arithmetic decoding process (https://en.wikipedia.org/wiki/Arithmetic_coding) with the LLM-supplied logits as the assumed probabilitiy distribution.
SynthID pulls off the same kind of trick, save that its 'tournament structure' allows application after the sampling step rather than before. The random seed under SynthID becomes context sensitive (keying on the prior N tokens as well as the watermark seed), but in expectation the next-token distribution is still the same as the unwatermarked case.
If there's no change in token distribution, then there can't be a change in quality.
However, the article and its sources are right about one factor: random generation diversity will suffer. Applying SynthID to an LLM stream acts like fixing the PRNG seed. Repeated generations with the same prompt and same seed will deliver the same result, up to small amounts of nondeterminism in the architecture (mostly GPU scheduling and roundoff). This matters under the same circumstances that it would matter if the provider said "we sample using the XYZ PRNG with fixed seed '42'".
In my opinion, 'sample diversity' and 'sample quality' are not the same thing, and conflating the two concepts leads to poor intuition. SynthID and the like probably matter most if one is trying to repeatedly sample an LLM in order to infer the probability distribution (e.g. using a commercial LLM provider as a Jev-style classifier, looking for calibrated answers).
It's an interesting review, but it seems to mix some things together. Gruber's claim is that we shouldn't sacrifice quality, but none of those sources show that's what's happening. Having less diversity in answers could mean you're more consistently getting a better/worse/same quality answer. There are lots of people complaining about not being able to get a real temperature=0 answer - they'd actually welcome the consistency too.
The practical example of not liking the sentence is not great in my opinion. Instead of just re-rolling the answer, why not continue in that thread and say what you don't like? Or request 20 alternatives, say which go in the right direction and request 20 more? Retrying from scratch with one sample is a waste of time and doesn't indicate anything about quality.
Basically, it's all theoretical - show us the actual benchmarks/samples where the quality is consistently degraded.
> Natzir Turrado, a technical consultant, actually tested this (article in Spanish) by reproducing SynthID-Text on an open model with the paper's settings and a watermark key. He asked the watermarked model five times to describe a memorable trip to a coastal town. Four of the five answers he got back opened in a near-identical way: “My heart still sings/dances when I think back to my trip to the quaint coastal town of…”. For the unwatermarked model, three opened in different ways.
A test of 2/5 vs 4/5 seems to be an obviously invalid experiment.
"X is the useful part" I hate this claudism. But a good thing about recognizing these is that you can pretty much skip any youtuber that uses any claudish phrasing. But it's damn near everyone now a days using claude to write their scripts.
They key idea is this: changing the random seed for LLM generation definitionally does not change the distribution of tokens in expectation over all seeds.
This can already be used to embed information in LLM streams, leading to LLM steganography work (https://arxiv.org/abs/2606.09135, but the idea dates back much further). Take your target message, whiten it via a one-time pad (stream cypher), then use it as the 'compressed stream' under an arithmetic decoding process (https://en.wikipedia.org/wiki/Arithmetic_coding) with the LLM-supplied logits as the assumed probabilitiy distribution.
SynthID pulls off the same kind of trick, save that its 'tournament structure' allows application after the sampling step rather than before. The random seed under SynthID becomes context sensitive (keying on the prior N tokens as well as the watermark seed), but in expectation the next-token distribution is still the same as the unwatermarked case.
If there's no change in token distribution, then there can't be a change in quality.
However, the article and its sources are right about one factor: random generation diversity will suffer. Applying SynthID to an LLM stream acts like fixing the PRNG seed. Repeated generations with the same prompt and same seed will deliver the same result, up to small amounts of nondeterminism in the architecture (mostly GPU scheduling and roundoff). This matters under the same circumstances that it would matter if the provider said "we sample using the XYZ PRNG with fixed seed '42'".
In my opinion, 'sample diversity' and 'sample quality' are not the same thing, and conflating the two concepts leads to poor intuition. SynthID and the like probably matter most if one is trying to repeatedly sample an LLM in order to infer the probability distribution (e.g. using a commercial LLM provider as a Jev-style classifier, looking for calibrated answers).
The practical example of not liking the sentence is not great in my opinion. Instead of just re-rolling the answer, why not continue in that thread and say what you don't like? Or request 20 alternatives, say which go in the right direction and request 20 more? Retrying from scratch with one sample is a waste of time and doesn't indicate anything about quality.
Basically, it's all theoretical - show us the actual benchmarks/samples where the quality is consistently degraded.
A test of 2/5 vs 4/5 seems to be an obviously invalid experiment.
For this example; he tested 30 prompts at 5 pass-pairs each; so 300 times.
https://natzir.com/posicionamiento-buscadores/que-es-como-fu...