Translate-Test (or Translate-Then-Test) is a cross-lingual transfer strategy where input data in a target language is translated into the source language (usually English) at inference time, allowing a source-trained model to process it — essentially adapting the input to the model rather than the model to the input.

Mechanism

• Model: Train a powerful model on English data (e.g., English BERT on SQuAD).
• Inference: User asks a question in Japanese.
• Translation: Translate Japanese Query → English.
• Prediction: Model predicts answer in English.
• Back-Translation: Translate Answer English → Japanese (optional).

Why It Matters

• SOTA Access: Allows using the absolute best English models (like GPT-4) on any language immediately.
• Latency: High latency due to explicit translation steps.
• Error Propagation: Translation errors in the query can lead to nonsense answers.

Translate-Test is using an interpreter — translating the world into the model's native language so it can perform the task.