Study locates and removes LLM 'slop' style

A study identified a representation direction for assistant style (e.g., em dashes, 'not X but Y') in Qwen3-4B-Instruct, finding that SFT introduces rhetorical structures and DPO doubles the slop score through buzzwords and markdown. Fixed-vector subtraction or distillation into a LoRA cuts the slop score by 83%.

• Researcher Without Institution Seeks arXiv Endorsement for VELM Paper2026-10-15