PlanningBench: Scalable Planning Data Generation for LLMs

PlanningBench is a new framework for generating scalable, diverse, and verifiable planning data to evaluate and train large language models. It addresses limitations of existing benchmarks that treat planning data as fixed collections, limiting scenario coverage and tying difficulty to surface-level proxies. The framework abstracts real planning scenarios into a structured taxonomy of over 30 task types, subtasks, constraint families, and difficulty factors. This enables controllable generation, automatic verification, and planning-oriented training. The approach supports broader scenario coverage and structural difficulty sources, improving LLM planning capabilities.