Translation-CoT: A Human-Inspired Chain-of-Thought Framework for Multilingual LLM Translation

Researchers from the University of Vermont introduced Translation-CoT, a chain-of-thought prompting framework that breaks translation into lexical retrieval, grammatical analysis, and topic identification stages, followed by a refinement step. In evaluations across 14 languages and multiple LLMs including GPT-4o, GPT-4o-mini, LLaMA 3.1, and Gemma 2, Translation-CoT outperformed zero-shot prompting, in-context learning, Tree-of-Thought, and Learning-Oriented Prompting on BLEU, ChrF, and METEOR metrics, with GPT-4o performing best overall and strongest gains in English-to-non-English and low-resource settings. Human evaluation showed higher preference scores and lower MQM penalty scores, indicating fewer mistranslations, omissions, awkward phrasing, and hallucinations.

Translation-CoT: A Human-Inspired Chain-of-Thought Framework for Multilingual LLM Translation https://aclanthology.org/2026.amta-research.2.pdf Tabia Tanzin Prama /people/tabia-tanzin-prama/ , Juniper L Lovato /people/juniper-l-lovato/unverified/ , Chris Danforth /people/chris-danforth/unverified/ , Peter Dodds /people/peter-dodds/ Abstract Large language models LLMs have transformed machine translation, yet mistranslations, hallucinations, and unnatural phrasing still limit their effectiveness, particularly for low-resource languages. We propose Translation-CoT, a chain-of-thought prompting strategy that breaks translation into structured stages lexical retrieval, grammatical analysis, and topic identification , followed by a refinement step to improve fluency, tone, and idiomatic expression. We evaluate Translation-CoT across 14 languages from 14 language families and multiple LLMs GPT-4o, GPT-4o-mini, LLaMA 3.1, and Gemma 2 , with GPT-4o performing best overall, in both English ↔ non-English X translation settings. Compared with zero-shot prompting, in-context learning, and existing chain-of-thought prompting methods Tree-of-Thought ToT and Learning-Oriented Prompting LOT , Translation-CoT outperforms these prompting strategies on multilingual machine translation across BLEU, ChrF, and METEOR, with especially strong gains in the more difficult English→non-English X setting and in low-resource languages. Human evaluation further shows higher preference scores and lower MQM penalty scores, indicating fewer mistranslations, omissions, awkward phrasing, and hallucinations with Translation-CoT. Overall, our results show that structured, task-aware prompting is an effective approach for improving multilingual translation quality and robustness in LLMs.- Anthology ID: - 2026.amta-research.2 - Volume: Proceedings of the 17th Conference of the Association for Machine Translation in the Americas Volume 1: Research Track /volumes/2026.amta-research/ - Month: - August - Year: - 2026 - Address: - Québec City, Canada - Editors: Eleftheria Briakou /people/eleftheria-briakou/unverified/ , Jeremy Gwinnup /people/jeremy-gwinnup/ , Shivali Goel /people/shivali-goel/unverified/ - Venue: AMTA /venues/amta/ - SIG: - Publisher: - Association for Machine Translation in the Americas - Note: - Pages: - 3–27 - Language: - URL: https://aclanthology.org/2026.amta-research.2/ https://aclanthology.org/2026.amta-research.2/ - DOI: - Cite ACL : - Tabia Tanzin Prama, Juniper L Lovato, Chris Danforth, and Peter Dodds. 2026. Translation-CoT: A Human-Inspired Chain-of-Thought Framework for Multilingual LLM Translation https://aclanthology.org/2026.amta-research.2/ . In Proceedings of the 17th Conference of the Association for Machine Translation in the Americas Volume 1: Research Track , pages 3–27, Québec City, Canada. Association for Machine Translation in the Americas. - Cite Informal : Translation-CoT: A Human-Inspired Chain-of-Thought Framework for Multilingual LLM Translation https://aclanthology.org/2026.amta-research.2/ Prama et al., AMTA 2026 - PDF: https://aclanthology.org/2026.amta-research.2.pdf https://aclanthology.org/2026.amta-research.2.pdf