Antislop Framework: Detecting and Eliminating Repetitive Patterns in Language Models
By
Der_Einzige
A respectable bake. You'd come back tomorrow for another.
Summary
Researchers present Antislop, a comprehensive framework for identifying and eliminating repetitive phraseology ("slop") in language model outputs. The framework includes three key innovations: the Antislop Sampler for suppressing unwanted strings during inference, an automated pipeline for profiling model-specific slop against human baselines, and Final Token Preference Optimization (FTPO) for fine-tuning models. The system demonstrates that some slop patterns appear over 1,000x more frequently in LLM output than human text, and FTPO achieves 90% slop reduction while maintaining or improving performance across various evaluation tasks.
Key quotes
· 4 pulledWidespread LLM adoption has introduced characteristic repetitive phraseology, termed 'slop,' which degrades output quality and makes AI-generated text immediately recognizable.
We demonstrate that some slop patterns appear over 1,000x more frequently in LLM output than human text.
FTPO achieves 90% slop reduction while maintaining or improving performance in cross-domain evals including GSM8K, MMLU, and creative writing tasks.
The Antislop Sampler successfully suppresses 8,000+ patterns while maintaining quality, whereas token banning becomes unusable at just 2,000.
You might also wanna read
DeepSeek-V4: Hybrid Sparse-Attention Architecture Enables Efficient Million-Token Context Inference
DeepSeek-V4 introduces a hybrid sparse-attention architecture combined with on-policy distillation across domain specialists, enabling 1M-to
artgor.medium.com·11h ago
Orthrus: A Dual-Architecture Framework for Fast, Lossless LLM Inference via Diffusion Decoding
Orthrus is a dual-architecture framework that combines autoregressive LLMs with diffusion models to enable fast, lossless parallel token gen
github.com·16d ago
Evaluating LLMs for TLA+ System Modeling: The Specula Team's Experience with Claude and Raft
The Specula team evaluates LLMs (specifically Claude) on their ability to model real-world systems using TLA+, a formal specification langua
sigops.org·23d ago
Zyphra's ZAYA1-8B Matches Frontier AI Models on Benchmarks Using Under 1 Billion Active Parameters, Trained on AMD Hardware
Zyphra released ZAYA1-8B, a model that matches or competes with frontier AI models like DeepSeek-R1, Claude Sonnet 4.5, and Gemini 2.5 Pro o
firethering.com·24d ago
Testing Karpathy's Autonomous Research Loop on CPU Architecture Optimization
This article explores whether Andrej Karpathy's autonomous research loop (autoresearch) — a coding agent that proposes, implements, measures
github.com·1mo ago
Research on Hierarchical JSON Representations for Preserving Scientific Sentence Meaning
This research paper investigates whether structured hierarchical JSON representations can effectively preserve the meaning of scientific sen
arxiv.org·1mo ago