Large language models (LLMs) have shown incredible promise in translation, but they often struggle with low-resource languages—those with limited training data. This is where Mufu, a novel technique leveraging multilingual fused learning, comes into play. Imagine trying to decipher a language you barely understand. You might rely on similar languages you know or consult dictionaries. Mufu takes a comparable approach. It automatically generates translation candidates in related languages, effectively creating a multilingual cheat sheet for the LLM. Then, it instructs the LLM not to translate from scratch but to refine these pre-existing candidates. This transforms the task from direct translation into a more manageable "proofreading" exercise. Mufu prompts the LLM to reason by assessing the quality of these candidates, aligning them semantically across languages, and selecting the best parts from each. It's like a detective piecing together clues from different sources to solve a case. Experiments show that LLMs fine-tuned with Mufu surpass even large, specialized translation models like NLLB in many low-resource language pairs. Even more surprisingly, Mufu works across different LLM architectures, like PaLM2 and Gemma. To address Mufu's higher computational cost, a distilled version offers a practical compromise between speed and accuracy. Challenges remain, such as the occasional dip in performance with extremely poor-quality candidate translations and balancing the accuracy of the full Mufu method with the efficiency of its distilled counterpart. Mufu opens exciting avenues for bridging the language gap by making quality translation accessible to a wider range of languages.