Enable javascript in your browser for better experience. Need to know to enable it? Go here.
Research

Antislop: A Comprehensive Framework for Identifying and Eliminating Repetitive Patterns in Language Models

Antislop tackles the repetitive phrasing that makes LLM output instantly recognizable as AI generated. It pairs an inference-time backtracking sampler with Final Token Preference Optimization (FTPO), a fine-tuning method that suppresses overused patterns at the token level with minimal collateral damage. FTPO cuts slop by 90% while holding steady on GSM8K, MMLU and creative writing quality, outperforming DPO and token banning.