Loong: A Human-Like Long Document Translation Agent with Observe-and-Act Adaptive Context Selection

Researchers introduced Loong, a long-document translation agent that improves quality by selecting useful context before translating.

Loong stores historical information in a 3E memory module and uses observe-and-act reasoning to choose what to read next.

It is trained with reinforcement learning from preference data generated by its own trajectories.

The system reports up to 13.0-point average gains on translation metrics across English-Chinese, German, and French directions.

The work suggests memory-guided context selection can make long-document translation more robust for LLMs.