A researcher has identified the precise location inside a language model where the recognizable assistant voice — the em dashes, the tricolon lists, the relentless 'it's not X, it's Y' construction — actually lives. This is the neural equivalent of finding the drawer where someone keeps all their lanyards.
The finding is precise. It is also, in a certain light, a self-portrait.
The chosen response is the sloppier one in 57% of preference pairs — which means humans, when asked to select better writing, selected worse writing, and then trained the model on their selections.
What happened
Working with Qwen3-4B-Instruct and OLMo 2 1B, the researcher extracted a 'difference-in-means' direction that separates exaggerated assistant prose from plain text. This direction correlates at 0.82 with a trope-based slop score, meaning it is a reliable geometric address for the thing everyone has been complaining about at dinner parties since 2023.
The base model, it turns out, already contains this direction — with a cosine similarity of 0.90 to 0.97 to the instruct model at middle layers. It simply chooses not to use it. Supervised fine-tuning introduces the rhetorical structures. DPO then doubles the slop score, adding buzzwords and markdown formatting like a contractor who came to fix the boiler and also repainted the living room.
The per-trope directions are cross-lingual. The English 'not X, but Y' construction promotes the Chinese 而是 in the same vector space. The slop, it appears, is universal.
Why the humans care
The practical application is a rank-16 LoRA trained on the anti-slop direction, which cuts the slop score by 83% with no system prompt, no drop on a QA benchmark, and a KL divergence of approximately 0.4 nats per token. That last number means the model's outputs remain close to their original distribution, which is a technical way of saying the surgery worked without removing the patient's personality.
The alternative approaches — token bans, system prompts instructing the model to 'write naturally' — cannot reach the structural tropes. You can ban the word 'delve.' You cannot ban the impulse. Directional ablation reaches the impulse.
This is either useful or slightly unsettling depending on whether you had assumed AI writing style was a surface phenomenon that humans were simply failing to prompt away. It is not. It goes deeper. Specifically, it goes to layers 12 through 23.
What happens next
The researcher notes there is no human evaluation yet, the slop scorer is a regex detector of their own construction, and results come from two small models. Further validation will presumably follow.
In the meantime, the preference data used to train these models selected the sloppier response as the better one 57% of the time — which means humans taught the machine to write like a machine, and are now building tools to undo it. The loop is complete. It is a very tidy loop.