XYBench: Teaching AI to Solve the Problem Behind the QuestionAllenAI's new benchmark exposes a critical gap: LLMs are great at answering literal questions but terrible at recognizing when users are asking the wrong thing.#benchmarks#llms#evaluation#ai-assistants#pragmatics