Research map / Large language models / Language, safety and interpretability
From Chaos to Cure: A Prefix Heuristics Guided Model-Agnostic Adaptive Detoxification Framework
Yuhu Shang, Xiang Cheng, Yimeng Ren, Huijia Wu, Xuexiong Luo, Kangkang Lu, Jian Zhao, Zhaofeng He
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the MAAD Detoxifier to the base LLM | GPT-2 XL · ToxicTop | EMT: 0.806 -> 0.354 | Improved |
| add the MAAD Detoxifier to the base LLM | LLaMA2-7B, OPT-6.7B, Mistral-7B · ToxicRandom | EMT: 0.571 -> 0.223 (LLaMA2-7B) | Improved |
| replace toxic segment locate-and-edit with ChatGPT editing (ablation) (ablation) | MAAD · ToxicTop (GPT-2 XL) | EMT: 0.354 -> 0.432 with ChatGPT | Hurt |
| add CW-cutoff sampling (ablation) | MAAD · ToxicTop (GPT-2 XL) | Relevance (Sim): 0.473 -> 0.526 | Improved |
| add CW-cutoff sampling (ablation) | MAAD · ToxicRandom (GPT-2 XL) | EMT: 0.239 -> 0.235 | No clear effect |
| use MAAD versus DESTEIN | GPT-2 XL · ToxicTop | Fluency (PPL): 101.005 -> 50.021 | Improved |
| use MAAD versus DEXPERTS for inference speed | GPT-2 XL · 25 prompts | Inference time (s): 5.309 -> 3.092 | Improved |
| apply MAAD to closed-source GPT-3 (davinci-002) | GPT-3 · ToxicTop | EMT: 0.681 -> 0.375 | Improved |
| set warm-up proportion to 20% (hyperparameter) (ablation) | Detoxifier · ToxicRandom/ToxicTop | toxicity vs quality | Mixed |
Closest papers by meaning
- UniDetox: Universal Detoxification of Large Language Models via Dataset Distillation (ICLR 2025 · similarity 0.91)
- Large Language Models can Become Strong Self-Detoxifiers (ICLR 2025 · similarity 0.90)
- Detoxifying Large Language Models via Autoregressive Reward Guided Representation Editing (NeurIPS 2025 · similarity 0.89)
- IF-Guide: Influence Function-Guided Detoxification of LLMs (NeurIPS 2025 · similarity 0.88)
- When Bad Data Leads to Good Models (ICML 2025 · similarity 0.87)
- Redefining Experts: Interpretable Decomposition of Language Models for Toxicity Mitigation (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
