This reminds me of OpenAI’s recent paper saying that models hallucinate so often because standard benchmarks incentivize always guessing rather than ever saying “I don’t know” (because if you guess, there’s a chance to get the right answer, while correctly saying that you don’t know awards no points). This would then be an instance of the same phenomenon. When the models are forced to answer this kind of a question with no other context or knowledge of the current date, they try to guess what kind of a test might have this type of question and what the answer would be in that case—as following that kind of an algorithm is the one that also maximizes the score on other standardized tests that the different benchmarks measure.
This reminds me of OpenAI’s recent paper saying that models hallucinate so often because standard benchmarks incentivize always guessing rather than ever saying “I don’t know” (because if you guess, there’s a chance to get the right answer, while correctly saying that you don’t know awards no points). This would then be an instance of the same phenomenon. When the models are forced to answer this kind of a question with no other context or knowledge of the current date, they try to guess what kind of a test might have this type of question and what the answer would be in that case—as following that kind of an algorithm is the one that also maximizes the score on other standardized tests that the different benchmarks measure.