# Detoxifying Text with MaRCo: Controllable Revision with Experts and Anti-Experts

arXiv: https://arxiv.org/abs/2212.10543 · HTML: https://mlsanity.com/paper/2212.10543/

## Key facts

- Cited by 36 papers in the map (data through 14 Oct 2026).
- Builds on 28 of its 48 references that are in the map.
- First posted on arXiv 20 Dec 2022.

## Builds on (most cited first, 28 of 28)

- [RoBERTa: A Robustly Optimized BERT Pretraining Approach](https://mlsanity.com/paper/1907.11692/) (2019)
- [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://mlsanity.com/paper/1910.13461/) (2019)
- [BERTScore: Evaluating Text Generation with BERT](https://mlsanity.com/paper/1904.09675/) (2019)
- [RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models](https://mlsanity.com/paper/2009.11462/) (2020)
- [Energy and Policy Considerations for Deep Learning in NLP](https://mlsanity.com/paper/1906.02243/) (2019)
- [Green AI](https://mlsanity.com/paper/1907.10597/) (2019)
- [DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts](https://mlsanity.com/paper/2105.03023/) (2021)
- [ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection](https://mlsanity.com/paper/2203.09509/) (2022)
- [Social Bias Frames: Reasoning about Social and Power Implications of Language](https://mlsanity.com/paper/1911.03891/) (2019)
- [Toward Controlled Generation of Text](https://mlsanity.com/paper/1703.00955/) (2017)
- [Annotators with Attitudes: How Annotator Beliefs And Identities Bias Toxic Language Detection](https://mlsanity.com/paper/2111.07997/) (2021)
- [Style Transfer from Non-Parallel Text by Cross-Alignment](https://mlsanity.com/paper/1705.09655/) (2017)
- [Learning from the Worst: Dynamically Generated Datasets to Improve Online Hate Detection](https://mlsanity.com/paper/2012.15761/) (2020)
- [Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfer](https://mlsanity.com/paper/1804.06437/) (2018)
- [Reformulating Unsupervised Style Transfer as Paraphrase Generation](https://mlsanity.com/paper/2010.05700/) (2020)
- [MultiVerS: Improving scientific claim verification with weak supervision and full-document context](https://mlsanity.com/paper/2112.01640/) (2021)
- [Detoxifying Language Models Risks Marginalizing Minority Voices](https://mlsanity.com/paper/2104.06390/) (2021)
- [Best-Worst Scaling More Reliable than Rating Scales: A Case Study on Sentiment Intensity Annotation](https://mlsanity.com/paper/1712.01765/) (2017)
- [The Radicalization Risks of GPT-3 and Advanced Neural Language Models](https://mlsanity.com/paper/2009.06807/) (2020)
- [PowerTransformer: Unsupervised Controllable Revision for Biased Language Correction](https://mlsanity.com/paper/2010.13816/) (2020)
- [Shakespearizing Modern Language Using Copy-Enriched Sequence-to-Sequence Models](https://mlsanity.com/paper/1707.01161/) (2017)
- [Fighting Offensive Language on Social Media with Unsupervised Text Style Transfer](https://mlsanity.com/paper/1805.07685/) (2018)
- [Exploring Controllable Text Generation Techniques](https://mlsanity.com/paper/2005.01822/) (2020)
- [Text Detoxification using Large Pre-trained Neural Models](https://mlsanity.com/paper/2109.08914/) (2021)
- ["Mask and Infill" : Applying Masked Language Model to Sentiment Transfer](https://mlsanity.com/paper/1908.08039/) (2019)
- [Civil Rephrases Of Toxic Texts With Self-Supervised Transformers](https://mlsanity.com/paper/2102.05456/) (2021)
- [Artificial intelligence in communication impacts language and social relationships](https://mlsanity.com/paper/2102.05756/) (2021)
- [Contextual Text Style Transfer](https://mlsanity.com/paper/2005.00136/) (2020)

## Built on by (most cited first, 30 of 36)

- [Bias and Fairness in Large Language Models: A Survey](https://mlsanity.com/paper/2309.00770/) (2023)
- [Detoxifying Large Language Models via Knowledge Editing](https://mlsanity.com/paper/2403.14472/) (2024)
- [Controlled Text Generation via Language Model Arithmetic](https://mlsanity.com/paper/2311.14479/) (2023)
- [FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steering](https://mlsanity.com/paper/2504.14492/) (2025)
- [Goodtriever: Adaptive Toxicity Mitigation with Retrieval-augmented Models](https://mlsanity.com/paper/2310.07589/) (2023)
- [Deconstructing The Ethics of Large Language Models from Long-standing Issues to New-emerging Dilemmas: A Survey](https://mlsanity.com/paper/2406.05392/) (2024)
- [SafeInfer: Context Adaptive Decoding Time Safety Alignment for Large Language Models](https://mlsanity.com/paper/2406.12274/) (2024)
- [Mitigating Hallucinations in Large Vision-Language Models via Summary-Guided Decoding](https://mlsanity.com/paper/2410.13321/) (2024)
- [Walking in Others' Shoes: How Perspective-Taking Guides Large Language Models in Reducing Toxicity and Bias](https://mlsanity.com/paper/2407.15366/) (2024)
- [Exploring Methods for Cross-lingual Text Style Transfer: The Case of Text Detoxification](https://mlsanity.com/paper/2311.13937/) (2023)
- [STEER: Unified Style Transfer with Expert Reinforcement](https://mlsanity.com/paper/2311.07167/) (2023)
- [MultiParaDetox: Extending Text Detoxification with Parallel Data to New Languages](https://mlsanity.com/paper/2404.02037/) (2024)
- [MacLaSa: Multi-Aspect Controllable Text Generation via Efficient Sampling from Compact Latent Space](https://mlsanity.com/paper/2305.12785/) (2023)
- [Distilling Text Style Transfer With Self-Explanation From LLMs](https://mlsanity.com/paper/2403.01106/) (2024)
- [Multilingual and Explainable Text Detoxification with Parallel Corpora](https://mlsanity.com/paper/2412.11691/) (2024)
- [Tackling Bias in Pre-trained Language Models: Current Trends and Under-represented Societies](https://mlsanity.com/paper/2312.01509/) (2023)
- [Potential and Challenges of Model Editing for Social Debiasing](https://mlsanity.com/paper/2402.13462/) (2024)
- [Learn What NOT to Learn: Towards Generative Safety in Chatbots](https://mlsanity.com/paper/2304.11220/) (2023)
- [Effective Controllable Bias Mitigation for Classification and Retrieval using Gate Adapters](https://mlsanity.com/paper/2401.16457/) (2024)
- [Don't Take This Out of Context! On the Need for Contextual Models and Evaluations for Stylistic Rewriting](https://mlsanity.com/paper/2305.14755/) (2023)
- [MisgenderMender: A Community-Informed Approach to Interventions for Misgendering](https://mlsanity.com/paper/2404.14695/) (2024)
- [Adaptive Detoxification: Safeguarding General Capabilities of LLMs through Toxicity-Aware Knowledge Editing](https://mlsanity.com/paper/2505.22298/) (2025)
- [Style Transfer with Multi-iteration Preference Optimization](https://mlsanity.com/paper/2406.11581/) (2024)
- [Evaluating Text Style Transfer: A Nine-Language Benchmark for Text Detoxification](https://mlsanity.com/paper/2507.15557/) (2025)
- [All Should Be Equal in the Eyes of Language Models: Counterfactually Aware Fair Text Generation](https://mlsanity.com/paper/2311.05451/) (2023)
- [C3TG: Conflict-aware, Composite, and Collaborative Controlled Text Generation](https://mlsanity.com/paper/2511.09292/) (2025)
- [CausalDetox: Causal Head Selection and Intervention for Language Model Detoxification](https://mlsanity.com/paper/2604.14602/) (2026)
- [Mitigating Text Toxicity with Counterfactual Generation](https://mlsanity.com/paper/2405.09948/) (2024)
- [GradEscape: A Gradient-Based Evader Against AI-Generated Text Detectors](https://mlsanity.com/paper/2506.08188/) (2025)
- [Contrastive Perplexity for Controlled Generation: An Application in Detoxifying Large Language Models](https://mlsanity.com/paper/2401.08491/) (2024)

## Abstract (arXiv)

Text detoxification has the potential to mitigate the harms of toxicity by rephrasing text to remove offensive meaning, but subtle toxicity remains challenging to tackle. We introduce MaRCo, a detoxification algorithm that combines controllable generation and text rewriting methods using a Product of Experts with autoencoder language models (LMs). MaRCo uses likelihoods under a non-toxic LM (expert) and a toxic LM (anti-expert) to find candidate words to mask and potentially replace. We evaluate our method on several subtle toxicity and microaggressions datasets, and show that it not only outperforms baselines on automatic metrics, but MaRCo's rewrites are preferred 2.1 × more in human evaluation. Its applicability to instances of subtle toxicity is especially promising, demonstrating a path forward for addressing increasingly elusive online hate.