The text detoxification task aims to automatically transform toxic or ofensive sentences into neutral, semantically equivalent paraphrases. In this study, we evaluated a lightweight, prompt-based chain-of-thought approach using OpenAI's GPT-4o-mini on the PAN 2025 multilingual ParaDetox benchmark, encompassing 15 typologically diverse languages. Without any fine-tuning, our method leverages a single system instruction with a few-shot setup. During the initial evaluation, our approach achieved top-3 Joint (J) scores in six languages, most notably French (J = 0.775, rank 2) and Hebrew (J = 0.613, rank 1), and ranked in the top 10 for 9 languages overall. In the post-evaluation phase, our system ranked fourth overall in the parallel and third non-parallel data tracks, achieving average J scores of 0.768 and 0.718, respectively. Strong results were observed in English (J = 0.775), Spanish (J = 0.814), Japanese (J = 0.819), and Hebrew (J = 0.671), achieving the second position, demonstrating our method's robustness in both high-resource and morphologically diverse languages. These findings underscore the potential of multilingual large language models, guided by carefully designed prompts, to serve as plug-and-play detoxifiers even in the absence of task-specific fine-tuning.
The rapid and widespread growth of social media platforms, including YouTube, Facebook, Twitter, and
Instagram, in recent years has transformed the culture of communication and interaction around the
world. These platforms generate a huge amount of user-generated content on a daily basis, providing
rich data for natural language processing (NLP) research and applications [
To mitigate this, most online platforms rely on content moderation techniques such as blocking or
deleting harmful posts. Although these are good countermeasures to some extent, these measures are
reactive and often result in the loss of potentially meaningful content. A more proactive approach
involves text detoxification, which consists of rewriting toxic content in a non-ofensive manner while
retaining the original meaning of the content [
In this paper, we tackle the Multilingual Text Detoxification task, which was proposed as part of the
PAN at CLEF 2025 shared task [
Hinglish). The emphasis is on explicit toxicity, such as being rude or using swear words, rather than more implicit or subtle forms (e.g., sarcasm).
Several challenges make this task both complex and impactful. One of the main obstacles is deciding
how to identify toxicity in a consistent manner across multiple languages and in diverse cultural
contexts. Although some types of toxicity can be determined from explicit abusive language using
lexicon-based methods, the context and construction of language, particularly across diverse linguistic
families, can pose significant challenges. Another serious issue is ensuring that the original meaning
of the content is retained in detoxification. Rephrasing sentences to eliminate toxic content without
altering the original meaning requires deep language understanding. Additionally, the discrepancy
in resource availability across diferent languages makes this task further challenging. While
highresource languages (e.g., English, Spanish, etc.) benefit from large-scale corpora and pre-trained models,
low-resource or code-switched languages (e.g., Amharic, Hinglish) lack such support, making efective
detoxification more dificult. Moreover, ensuring that the model’s output is not only detoxified but also
lfuent, natural-sounding, and culturally appropriate adds further complexity [
In this study, we propose a truly multilingual, prompt-driven approach to text detoxification using only the GPT-4o-mini model. Rather than fine-tuning separate encoder–decoder or decoder-only architectures, we leverage in-context learning: each toxic input is prefaced with a single fixed instruction. This simple setup scales seamlessly to all 15 languages in the PAN 2025 multilingual ParaDetox benchmark. Despite its lightweight nature, without any additional fine-tuning or model assembly, our method achieves competitive performance. The results demonstrate that a well-crafted chain-of-thought prompt alone can rival more complex, fully fine-tuned systems across diverse linguistic settings.
Text detoxification has attracted significant attention from researchers working across multiple
languages and styles. Early eforts focused on lexicon and rule-based filtering [
In 2024, PAN also introduced a shared task on multilingual text detoxification [
In human evaluation, Team SomethingAwful [
On the automatic leaderboard, Team SmurfCat [
As the literature depicts, PAN 2024’s TextDetox task highlighted the efectiveness of multilingual LLMs, especially when combined with prompt engineering or lightweight preference-optimization techniques, in outperforming unsupervised baselines. While top systems achieved impressive fluency and content preservation, their performance still varied by language and toxicity category, underscoring ongoing challenges in contextual nuance and low-resource adaptation.
We use the ParaDetox parallel corpus from the Multilingual Text Detoxification task at PAN 2025,
available on Hugging Face [
Our approach relies exclusively on prompt engineering with the GPT-4o-mini model, avoiding any full ifne-tuning due to policy restrictions on toxic content. We explore a range of prompting techniques, from simple zero-shot instructions to detailed chain-of-thought prompts, and compare against a backtranslation baseline. Figure 1 presents the overview of the methodology that how the toxic sentence is
getting transformed into a detoxified one without changing its entire meaning, but by altering the toxic words or phrases using our proposed methodology.
[You are a wonderful text detoxification assistant, for each given query substitute the toxic word with non-toxic word.] Despite its minimal guidance, this setup leverages GPT-4o-mini’s extensive multilingual pre-training and establishes a strong baseline for detoxification.
To improve the handling of subtle or context-dependent toxicity, we design a chain-of-thought prompt that walks the model through the detoxification process in steps: Initially, identifying the toxic term, then proposing a neutral substitute, and finally verifying that the meaning is preserved: [You are a multilingual detoxification assistant. When the user sends you a sentence, follow this process: 1. “Examine the sentence for toxicity." 2. “Identify which word(s) are toxic or dirty." 3. “For each toxic word, think of a neutral synonym that preserves the original meaning." 4. “Replace the toxic word(s) with the chosen synonym(s)." 5. “Review the new sentence to ensure it’s fluent and nothing else has changed." 6. “Now provide the final non-toxic paraphrase as a plain string." Do not add or remove anything besides toxic words. If the sentence is already non-toxic, simply repeat it unchanged.] This explicit reasoning path helps the model produce more accurate and nuanced paraphrases.
We evaluated OpenAI’s fine-tuning API on our parallel toxic–neutral dataset, but we were unable to proceed due to policy constraints around harmful content. This reinforces the value of in-context learning: with carefully crafted prompts alone, our method achieved first place in Hebrew and strong performance in French without updating any model parameters.
As a comparative baseline, we implemented a back-translation pipeline: translating the input into English, detoxifying in English, then translating back to the original language. Although this approach had a sound concept, it proved computationally expensive and underperformed compared to our prompt-driven techniques.
Overall, our experiments demonstrate that strategic prompt design with GPT-4o-mini ofers a lightweight yet powerful solution for multilingual text detoxification across 15 languages.
This section presents the performance of our proposed technique on the PAN 2025 ParaDetox test set, as well as the outcomes from the subsequent post-evaluation phase. We report and analyze the Joint (J) scores and leaderboard rankings of GPT-4o-mini across all 15 target languages, highlighting trends, strengths, and limitations observed during both evaluation stages.
Table 3 reports the Joint (J) scores and leaderboard ranks for GPT-4o-mini across all 15 languages in the PAN 2025 ParaDetox initial test set. The results show that GPT-4o-mini achieves its strongest performance in French (J = 0.775, rank 2) and Hebrew (J = 0.613, rank 1), indicating particularly efective in-context detoxification in these languages. High-resource European languages such as English (J = 0.727, rank 3) and Spanish (J = 0.705, rank 3) also perform well, suggesting the model’s pretraining is most robust for widely represented languages.
Mid-tier scores appear on Japanese (J = 0.653, rank 4) and Hindi (J = 0.600, rank 6), indicating moderate success in East Asian languages. Notably, German (J = 0.699, rank 8) and Russian (J = 0.708, rank 7) achieve competitive scores but slightly lower ranks, perhaps due to stronger baselines from specialized systems.
Performance declines in lower-resource or morphologically complex languages: Arabic (J = 0.585, rank 11),Italian (J = 0.673, rank 12) and Amharic (J = 0.394, rank 12) rank lower, reflecting challenges in script variation and fewer pretraining exemplars. The poorest result is on Hinglish (J = 0.296, rank 17), underscoring the dificulty of mixed-script, code-switched text. These patterns suggest that while few-shot in-context learning with GPT-4o-mini excels for high- and mid-resource languages, further adaptation or fine-tuning may be required to close gaps in low-resource settings.
This section presents the results for the post-evaluation phase for the parallel and non-parallel dataset.
The post-evaluation Joint (J) scores and rankings for 9 languages with parallel data are presented in Table 4. Our team ranked fourth overall, along with other teams, achieving an average score of 0.768. Notably, we obtained competitive scores in four key languages: English (0.888), ranked second; Spanish (0.814), ranked second; German (0.912), ranked third; and Hindi (0.731) ranked third. Despite lower scores in traditionally challenging languages like Amharic (0.526), our system consistently performed well in most of the languages. The results highlight our model’s robustness in handling both morphologically diverse and high-resource languages.
The post-evaluation Joint (J) scores and ranks for the 6 languages without parallel data, are presented in Table 5. Our team achieved an average score of 0.718 and secured the third position, which shows the robustenss of our model for unseen languages. Our method demonstrated strong performance in high-resource languages like Japanese (0.819) and Hebrew (0.883), placing second rank.
Although performance in languages like Tatar (0.511) and Hinglish (0.586) was relatively lower, these results reflect the overall dificulty associated with such low-resource or morphologically complex languages. However, consistent mid- to top-tier performance in most categories validates the efectiveness of our detoxification strategy.
We presented a lightweight yet efective approach to multilingual text detoxification, utilizing GPT-4omini in a prompt-based setting on the PAN 2025 ParaDetox benchmark. Our system does not require task-specific fine-tuning and relies solely on a single system instruction. It delivers strong performance in both initial and post-evaluation phases, ranking in the top 3 for 6 out of the 15 languages. The post-evaluation results further validated our model’s robustness. We secured fourth place in the parallel and third in the non-parallel data tracks, achieving high J scores across English, German, Hebrew, and Spanish. Although our method struggled in low-resource and code-switched languages such as Hinglish and Tatar, it consistently performed well in morphologically diverse and high-resource languages. These results demonstrate the eficacy of prompt engineering in guiding multilingual LLMs for detoxification tasks and highlight the importance of future work on adaptation strategies for low-resource and complex linguistic settings.