We present Command A Translate, an LLMbased machine translation model built off Cohere’s Command A. It reaches state-of-the-art machine translation quality via direct preference optimization. Our meticulously designed data preparation pipeline emphasizes robust quality control and a novel difficulty filtering – a key innovation that distinguishes Command A Translate. Furthermore, we extend our model and participate at WMT with a system (CommandA-WMT) that uses two models and post-editing steps of step-by-step reasoning and limited Minimum Bayes Risk decoding.

We investigate a surprising limitation of LLMs: their inability to consistently generate text in a user’s desired language. We create the Language Confusion Benchmark (LCB) to evaluate such failures, covering 15 typologically diverse languages with existing and newly-created English and multilingual prompts. We evaluate a range of LLMs on monolingual and cross-lingual generation reflecting practical use cases, finding that Llama Instruct and Mistral models exhibit high degrees of language confusion and even the strongest models fail to consistently respond in the correct language. We observe that base and English-centric instruct models are more prone to language confusion, which is aggravated by complex prompts and high sampling temperatures. We find that language confusion can be partially mitigated via few-shot prompting, multilingual SFT and preference tuning. We release our language confusion benchmark, which serves as a first layer of efficient, scalable multilingual evaluation.