Whom's messenger? You didn't point us to anyone's research.
I just don't see how sampling tokens constrained to a grammar can be worse than rejection-sampling whole answers against the same grammar. The latter needs to follow the same constraints naturally to not get rejected, and both can iterate in natural language before starting their structured answer.
Under a fair comparison, I'd expect the former to provide answers at least just as good while being more efficient. Possibly better if top-whatever selection happened after the grammar constraint.
I will die on this hill and I have a bunch of other Arxiv links from better peer reviewed sources than yours to back my claim up (i.e. NeurIPS caliber papers with more citations than yours claiming it does harm the outputs)
Any actual impact of structured/constrained generation on the outputs is a SAMPLER problem, and you can fix what little impact may exist with things like https://arxiv.org/abs/2410.01103
As opposed to emitting non-JSON tokens and having to throw away the answer?
Or just run json.dumps on the correct answer in the wrong format.
Don't shoot the messenger
Whom's messenger? You didn't point us to anyone's research.
I just don't see how sampling tokens constrained to a grammar can be worse than rejection-sampling whole answers against the same grammar. The latter needs to follow the same constraints naturally to not get rejected, and both can iterate in natural language before starting their structured answer.
Under a fair comparison, I'd expect the former to provide answers at least just as good while being more efficient. Possibly better if top-whatever selection happened after the grammar constraint.
THIS IS LIES: https://blog.dottxt.ai/say-what-you-mean.html
I will die on this hill and I have a bunch of other Arxiv links from better peer reviewed sources than yours to back my claim up (i.e. NeurIPS caliber papers with more citations than yours claiming it does harm the outputs)
Any actual impact of structured/constrained generation on the outputs is a SAMPLER problem, and you can fix what little impact may exist with things like https://arxiv.org/abs/2410.01103
Decoding is intentionally nerfed/kept to top_k/top_p by model providers because of a conspiracy against high temperature sampling: https://gist.github.com/Hellisotherpeople/71ba712f9f899adcb0...
I use LLMs for Actual Work (boring shit).
I always set temperature to literally zero and don't sample.
I honestly would like to hope people were more up in arms over this, but.. based on historical human tendencies, convenience will win here.