New method generates question-specific evaluation criteria for LLMs, covering 89% of expert standards while creating novel assessment approaches

Qworld uses recursive expansion trees to decompose questions into scenarios, perspectives, and binary criteria rather than applying static rubrics across datasets

On HealthBench benchmark, the method covers 89% of expert-authored criteria while generating 79% novel criteria validated by human experts

Addresses key limitation of existing LLM evaluation: context-dependent response quality that binary scores and one-size-fits-all rubrics cannot capture

Employs hierarchical and horizontal expansion to create fine-grained, question-specific standards for what constitutes a high-quality answer