translate-test

**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.