Tracing where LLM "slop" style comes from in the model and in training, and cutting it out [R]

A researcher traced the recognizable LLM assistant writing style — em dashes, "it's not X, it's Y" constructions, lists of three and buzzwords — to a difference-in-means direction in Qwen3-4B-Instruct, computed from the model's own exaggerated versus plain writing. The work also set out to identify which training stage installs the style and to remove it from the model's output.

I wanted to find where the recognizable assistant style em dashes, "it's not X, it's Y", lists of three, buzzwords lives in a model and which training stage installs it. Where it lives: in Qwen3-4B-Instruct, a difference-in-means direction from the model's own exaggerated vs plain writing separate