In plain words: A toolkit finds overused phrases AI text leans on and removes them by rethinking a word when a banned phrase starts, or by nudging word choices during training. That training cut these phrases by 90% without hurting quality, while the usual preference training hurt writing.
Abstract · Antislop: A Comprehensive Framework for Identifying and Eliminating Repetitive Patterns in Language Models
Widespread LLM adoption has introduced characteristic repetitive phraseology, termed "slop," which degrades output quality and makes AI-generated text immediately recognizable. We present Antislop, a comprehensive framework providing tools to both detect and eliminate these overused patterns. Our approach combines three innovations: (1) The Antislop Sampler, which uses backtracking to suppress unwanted strings at inference time without destroying vocabulary; (2) An automated pipeline that profiles model-specific slop against human baselines and generates training data; (3) Final Token Preference Optimization (FTPO), a novel fine-tuning method that operates on individual tokens, surgically adjusting logits wherever a banned pattern has appeared in an inference trace. We demonstrate that some slop patterns appear over 1,000x more frequently in LLM output than human text. The Antislop Sampler successfully suppresses 8,000+ patterns while maintaining quality, whereas token banning becomes unusable at just 2,000. Most importantly, FTPO achieves 90% slop reduction while maintaining or improving performance in cross-domain evals including GSM8K, MMLU, and creative writing tasks. In contrast, DPO suffers significant degradation in writing quality and lexical diversity despite achieving weaker suppression. We release all code and results under MIT license: https://github.com/sam-paech/auto-antislop.
Samuel Paech, Allen Roush, Judah Goldfeder, Ravid Shwartz-Ziv
arXiv:2510.15061 · cs.LG, cs.CL · submitted Oct 16, 2025 · updated Oct 21, 2025
abstract · pdf · html · 11 pages + appendices, 16 figures
Lots of things have changed in that year, but the things that haven't are:
* So, so many em-dashes. All over the place. (I've tried various ways to get it to stop. None of them have worked long term).
* Random emojis.
* Affirmations at the start of messages. ("That's a great idea!") With a brief pause when 5 launched. But it's back and worse than ever now.
* Weird adjectives it gets stuck on like "deep experience".
* Randomly bolded words.
Honestly, it's kind of helpful because it makes it really easy to recognize content that people have copied and pasted out of ChatGPT. But apart from that, it's wild to me that a $500bn company hasn't managed to fix those persistent challenges over the course of a year.