The Price of Reasoning: Cost-Quality Tradeoffs in Reinforcement Learning for Neural Machine Translation
This work investigates the importance of including the model's reasoning trace in the generated responses during both training and inference by systematically omitting it from one of the phases and shows that including the reasoning, specifically during inference, has a positive effect on the overall translation quality.