Retrieving external knowledge can improve large language model (LLM) translation by providing supplementary information such as idiom interpretations and terminology. However, existing retrieval-augmented machine translation approaches appear to lack a mechanism for deciding when retrieval is genuinely beneficial, potentially retrieving indiscriminately or at the wrong times.

The paper introduces an agentic reinforcement learning framework for selective retrieval. Instead of always retrieving, the model learns a policy that determines when to fetch external knowledge—only when it helps—and when to rely on its own parameters. This is framed as an agentic decision process, where the model acts as an agent deciding on retrieval actions.

The approach aims to improve translation quality while reducing unnecessary retrieval, which could also lower computational overhead and avoid introducing irrelevant or noisy context. The abstract suggests that this selective behavior is learned through reinforcement learning, though the specific algorithm details are not fully described in the available excerpt.