Speculative Decoding Across Languages

In three linesResearchers improve multilingual speculative decoding by comparing three strategies: fine-tuning draft models on task-specific data, fine-tuning on unlabeled monolingual corpora, and training n-gram draft models. Evaluation across 11 languages on translation and story generation tasks. N-gram models provide consistent speedups despite lower acceptance rates.

Summary generated by Claude — human-verified