Researchers have developed a method called Qworld for evaluating large language models (LLMs) on open-ended questions. Unlike traditional methods that use binary scores or static rubrics, Qworld generates question-specific evaluation criteria using a recursive expansion tree. This allows the model to capture context-dependent requirements and evaluate LLM responses tailored to each question. In tests, Qworld covered 89% of expert-authored criteria and generated 79% novel criteria validated by human experts, outperforming prior methods. --- Why it matters: This matters because it enables more accurate evaluation of large language models on complex tasks, which is crucial for their development and deployment in real-world applications. By generating question-specific criteria, Qworld can reveal capability differences between LLMs that were not apparent with coarse rubrics. Source: https://arxiv.org/abs/2603.23522