SFT installs the slop, DPO doubles it: representation-level ablation cuts AI-slop score 83%
Imaginary_Money2285 · reddit · 2026-10-15
A reproducible Reddit study traces where LLM "slop" style (em dashes, "it's not X, it's Y", lists of three, buzzwords) lives, where it comes from, and how to cut it.
Where it lives: in Qwen3-4B-Instruct, a difference-in-means direction between the model's exaggerated vs plain writing separates 64 handwritten minimal pairs and correlates up to 0.82 with a trope-based slop score. Logit-lens reveals a register axis (grand abstract vocabulary vs plain hedges), and trope directions are cross-lingual — "not X, but Y" promotes Chinese 而是.
Where it comes from: the base model has the same direction (cosine 0.90–0.97 at middle layers) but writes near the plain end. Across OLMo 2 1B checkpoints, SFT introduces the rhetorical structures and DPO doubles the slop score; 57% of 3,000 OLMo preference pairs chose the sloppier response.
Cutting it: directional ablation backfires (+89% slop, clean text projects negatively), but fixed-vector subtraction at layers 12–23 works within 0.4 nats/token KL. Distilled into a rank-16 LoRA, it cuts the slop score 83% with no system prompt, lower KL than steering, and no QA drop.
Caveats: self-written regex scorer, one 4B model plus OLMo 2 1B, no human eval yet. Complementary to token-level Antislop (Paech et al., 2025).
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23