basically yes

but the sneaky interesting part is they don’t define “better writing” as “a judge liked it more”

they define it as: does the model’s writing look like it came from the same distribution as real human writing?

lol pretty much

imagine you’re trying to spot a fake restaurant menu

one dish might look plausible. but across 1,000 dishes, the fake menu keeps using the same sauce, same adjectives, same garnish, same “chef’s kiss” move

exactly. AI writing often fails at the population level

too many repeated structures, overly tidy arguments, favorite phrases, same pacing, same generic detail level

the report’s claim is: SFT alone doesn’t fix that

😮that’s the counterintuitive bit

SFT teaches “for this prompt, produce this kind of answer”

but human-likeness is partly a group property: the spread, variety, weirdness, and frequency of patterns across many answers

yes. or like teaching cooking from one perfect plate per recipe

they learn “correct dish,” but maybe not the normal human range: messy plating, substitutions, mild chaos, different spice instincts

three main lenses from the summary

n-gram token L2 distance: are words and short phrases overused or underused? this catches stuff like em-dash addiction

MMD: compares embedding distributions, so it can see if the overall content/style cloud is in the wrong place

JMQ: a judge model compares model output against human reference output

so they’re not only asking “is this good?” — they’re asking “does this look statistically close to the human reference set?”

tragically, no yacht

think of embeddings as dropping every essay onto a map

human essays form one weather system. model essays form another. MMD asks how far apart those weather systems are

yep. according to the report, SFT outputs still differ strongly from human references

and just tweaking temperature doesn’t cleanly solve it

more variety, yes. better match, not automatically

different sampling temperatures optimize different metrics

so one temperature might help phrase diversity while another helps judge preference or embedding match

exactly. the paper frames distribution fine-tuning, DFT, as optimizing closer to the thing they actually care about: matching the output distribution

not just “imitate this one target answer”

that’s the idea

the report says DFT beat a strong “super baseline” that got to cherry-pick the best SFT hyperparameter / sampling result per metric

yeah, which makes the claim more interesting

they also report a 4B DFT model beating a 14B SFT superbaseline on MMD, and an 8B SFT superbaseline on JMQ

yup. if the metric is distribution match, a smaller model trained for that can beat a larger model trained in the usual way

at least on the reported setup

slop is the human-facing symptom

the model overuses certain rhythms: “not X but Y,” tidy three-part structures, generic caveats, polished-but-empty paragraphs

distribution metrics are trying to catch the statistical fingerprint underneath that cringe

that’s the cleanest takeaway

instead of “make it sound human,” ask: across lots of outputs, which tokens, structures, embeddings, and judge comparisons are out of line with humans?

a couple

their demo asks for structured inputs — prompt, outline, writing style, use case — so users state intent instead of just “write me a thing”

and they add copy-paste friction by injecting random fruits/animals into copied text

basically. plus no public API, apparently to reduce automated spam use

so the technical goal and product constraints are aligned: better writing, but not “industrialized fake human internet at scale”

three things

1. SFT can imitate examples without matching the human distribution

2. sampler tweaks help some metrics and hurt others, so temp isn’t a universal slop dial

3. if you care about human-like writing at scale, measure distribution mismatch directly: token patterns, embedding clouds, and judge comparisons

exactly. less “copy this essay,” more “match the ecosystem these essays came from”

ok go forth and side-eye every em dash on the internet