A new translation method gives you the high quality of slow-thinking AI models at the lightning speed of fast-thinking ones.
ReflectMT internalizes the reflection process of reasoning models so they do not have to output a visible chain-of-thought. This allows the AI to produce near-perfect translations in a single pass, saving over 94% on token costs. We previously thought that high-quality reasoning required a long, visible internal monologue. This research proves that a model can think internally without slowing down or wasting tokens. It makes elite-level AI reasoning affordable for real-time applications like live translation and voice assistants.
ReflectMT: Internalizing Reflection for Efficient and High-Quality Machine Translation
arXiv · 2604.19144
Recent years have witnessed growing interest in applying Large Reasoning Models (LRMs) to Machine Translation (MT). Existing approaches predominantly adopt a "think-first-then-translate" paradigm. Although explicit reasoning trajectories significantly enhance translation quality, they incur prohibitive inference costs and latency. To address these limitations, we propose ReflectMT, a two-stage reflection internalization algorithm for machine translation that employs a "translate-first-think-late