Tag
StalePO introduces a token-level preference optimization method for machine translation that uses legacy post-edits to improve model performance, showing significant gains in quality metrics.