“ScholarlyArticle: "Antislop: A Comprehensive Framework for Identifying and Eliminating Repetitive Patterns in Language Models" (2025) https://arxiv.org/abs/2510.15061: > Abstract: [...] Our approach combines three innovations: (1) The Antislop Sampler, which uses backtracking to suppress unwanted strings at inference time without destroying vocabulary; …”
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.
候选层:至少 2 个不同的人、在 2 个不同月份提到过,还没到主书架的 3 人 3 个月。没有做立场标注,所以只写“被提及”。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“And people like you will hate it even more when the normies immunize themselves from being obviously caught by such tells: https://arxiv.org/abs/2510.15061”
“We wrote the paper on how to deslop your LLM outputs and if you use our factory de-slopped versions of gemma3 you don't have to worry about this, similarly if you use our antislop sampler, your LLM outputs will look very close to human. https://arxiv.org/abs/2510.15061”
“Read our paper on de-slopping LLM outputs. It's far more than simply all having the same fixed AI system prompts. It's an overuse of post-training and contempt for pre-training. https://arxiv.org/pdf/2510.15061”
“Go take a giant dataset of LLM generated outputs, use an accurate POS tagger and look for 5-grams or similar lengths of matching patterns. If you do thi, you’ll pull out the overrepresented paragraph and sentence level slop that we humans intuitively detect easily. If your writing appears to be AI generated, I assume you aren’t willing to put human inte…”
“Yup, this is SUPER AI generated. Next time OP that you want to lie about using LLMs, either use the techniques from our paper: https://arxiv.org/abs/2510.15061 Or if you're more lazy, there's a "logit_bias" technique which you can use to ban the em-dash in language models. But you were too lazy to do that, and you lied about not using AI. Shame on you …”
看全部 11 条 →
“Our paper on removing AI slop got accepted to ICLR 2026, and it's under consideration for an IgNobel prize: https://arxiv.org/abs/2510.15061 Our definition of slop (repetitive characteristic language from LLMs) is the original one as articulated by the LLM creative writing community circa 2022-2023. Folks trying to redefine it today to mean "lazy LLM ou…”
“This is what they get for not reading our antislop paper (ICLR 2026) and using our anti-slopped sampler/models, or Kimi (which is remarkable relatively non sloppy) https://arxiv.org/abs/2510.15061 I thought normies would have caught onto the EM dash, overuse of semicolons, overuse of fancy quotes, lack of exclamation marks, "It's not X, it's Y", etc. Cl…”
“FWIW: When I accuse folks of it, take it more seriously - we literally wrote a paper (ICLR 2026) on deslopping language models: https://arxiv.org/abs/2510.15061 Also you might not care, but normies sure care! https://arxiv.org/html/2410.03723v2 https://arxiv.org/html/2510.08831v1”
“For anyone that wants to evade the kind of people who want to figure out if an AI wrote your review or not, we wrote a whole paper (ICLR 2026!) on how to do that! https://arxiv.org/abs/2510.15061 I consider all types of "I liked this output, but don't the moment I learned it was AI generated" to be externalizations of "carbon chauvinism" (https://en.wik…”
“This is a better way than we use to create the "slop ban list" in our "Antislop" sampler/fine-tuning method. Gonna totally nab the weights for all these words and remove this crap from claude going forward! Thanks OP!!! https://arxiv.org/abs/2510.15061 (ICLR 2026)”
- Antislop: A framework for eliminating repetitive patterns in language models ↗Der_Einzige · 2025-10-23 · 120 分