Threat Models
Language-shift, translation-mediated, code-switching, red-teaming, and post-deployment adaptation attacks.
Findings of ACL 2026
A Survey of Toxicity Detection and Mitigation Strategies for Multilingual Language Models
A survey of multilingual toxicity detection and detoxification for LLMs, organized around threat models, task setups, detection methods, mitigation strategies, and open evaluation challenges.
Abstract
Multilingual safety is not English safety translated.
Large language models (LLMs) are increasingly deployed across languages, but their safety behavior remains uneven across linguistic and cultural contexts. This survey synthesizes work on toxicity detection and detoxification for multilingual LLMs. We first catalogue threat models that exploit language choice, translation pivots, code-switching, orthographic variation, multi-turn interaction, and post-deployment fine-tuning to weaken safety alignment. We then organize task formulations, multilingual detection approaches, and mitigation strategies spanning data filtering, supervised and preference-based tuning, decoding-time steering, representation editing, and multilingual guardrails. Across these areas, we identify persistent challenges: uneven language coverage, culturally contingent definitions of harm, fragmented evaluation protocols, and the risk that detoxification suppresses legitimate dialectal or identity-related expression.
Paper Map
What the survey organizes
Datasets and Metrics
Toxic-to-neutral rewriting, toxic text detection, toxic-generation evaluation, and preservation metrics.
Detection
Multilingual transformers, translation pipelines, representation-level probes, and LLM-based detectors.
Detoxification
Data filtering, supervised and preference tuning, decoding-time steering, representation editing, and guardrails.
Open Challenges
Cross-lingual gaps, cultural misalignment, fragmented evaluation, over-suppression, and code-switching.
Taxonomy
A compact view of multilingual toxicity work
Takeaways
Research needs highlighted by the paper
Cross-lingual transfer remains unreliable
Safety methods that work in English can underperform in low-resource, morphologically rich, or culturally distant settings.
Evaluation needs cultural grounding
Benchmarks need to move beyond translated prompts and include community-aware definitions of harm.
Detoxification can erase legitimate expression
Mitigation should measure toxicity reduction alongside preservation of meaning, style, dialect, and identity-related speech.
References