Where AI slop comes from, and how to turn it down inside the model (100toolkit.com)

Recent advancements by researchers have uncovered a method to manipulate the "sloppy" writing style produced by AI language models, specifically the Qwen3-4B. It was found that the model's degree of sloppiness is linked to specific rhetorical structures, such as the phrase "not X, but Y," and its propensity to use buzzwords. By analyzing the model's pre-training and post-training stages, the researchers observed that while the base model, trained solely on web text, exhibited a more restrained writing style, the assistant version generated output 2.5 times sloppier due to preferences established during supervised fine-tuning. This work is significant for the AI/ML community as it highlights the potential to control and refine output quality by targeting specific internal language model behaviors. The study demonstrates that zeroing out the model's sloppiness direction can paradoxically increase it, while strategically pushing the model's output towards clearer language can yield an 83% reduction in slop without prompting. This insight into the internal mechanics of AI writing can lead to improved instructional approaches and model designs, ultimately allowing for more impactful and accurate AI communication across various applications.