Spacing at HTML tags in LLM and MT engine translations

How do LLMs and MT engines compare on spacing at tag boundaries?

Translations from LLM-based MT engines have spacing errors at tag boundaries in 99% of the cases, from other MT engines in 68%, from LLMs in 9% and from open-weight LLMs in 4%. Translations from Anthropic Claude Opus 4.8 and Anthropic Claude Sonnet 4.6 keep the spacing correct in every language where we shortlisted them, and those from Microsoft Translator in 9 of 32.

Spacing errors at tags appear in every language, most in Basque, Czech and Russian

In Basque and Czech, translations from Google Gemini 3.1 Pro Preview, DeepL Next Gen and Google Translation LLM have spacing errors, and in Russian translations from OpenAI GPT-5.5, Mistral Large 3 and DeepL Next Gen. In Gujarati, Norwegian and Serbian, translations from all but one shortlisted model keep the spacing correct.

Translations from DeepL Next Gen keep the spacing at tag boundaries correct in 2 of 69 languages

Translations from Microsoft Translator keep it correct in 9 of 32 languages, and translations from Translated Lara 2, Google Translation LLM and Widn AI Tower Sugarloaf have spacing errors in every language where we shortlisted them. Among LLMs, translations from MoonshotAI Kimi K2.5 keep it correct in 3 of 11 languages and from Mistral Large 3 in 13 of 28.

Every one-step solution with a spacing error has the spacing correct as a two-step or three-step solution

On the same solutions, every one-step solution with a spacing error has the spacing correct as a three-step solution (translation + LQA + repair), and so does every one with a result as a two-step solution (translation + post-edit). We measured these fixes on LLM solutions, since we leave out the MT engine solutions as too far off.

Disclaimers

• Spacing at tag boundaries is correct with no glued or split words at tags, no double spaces, no space before punctuation and no space after an opening bracket.
• We assess only the shortlisted one-step solutions (10 to 13 per language by COMET-Kiwi or MetricX, with at least three real-time solutions among them), so a model has no result in a language where it is not on the shortlist.
• The results come from four business texts of ~3,000 words. Other content can show other results.
• We build two-step and three-step solutions only in the languages with fewer than two solutions that meet every Standard requirement, so their results describe those languages.