So presumably, sometimes when asked a question in another language the model should “think” in English? TFA doesn’t say either way.