Data agents use code to process heterogeneous files and extract analytical insights. Scaling their training is hampered by costly environment curation, privacy and IP restrictions, and the difficulty of verifying open‑ended results without a known generative truth. The key idea is to specify the world behind the files before generating the files themselves. We introduce Data World, a multi‑agent engine that synthesizes environments, tasks, and evaluation rubrics from explicit mechanisms. Independent audits check whether the target can be recovered from observed files; rubrics provide scores and RL rewards. Human‑preference curation for realism selects 100 tasks covering 50 unseen worlds and 16 domains, forming DataWorld‑Bench. The engine also yields 2,135 training tasks from 1,037 disjoint worlds. Evaluation of ten frontier models shows the benchmark remains challenging, especially for insight discovery, revealing a clear compute‑insight gap. RL‑zero training on the synthesized tasks directly fine‑tunes Qwen3.6‑35B‑A3B into DataWorld‑35B‑A3B without trajectory distillation. Under matched evaluation it improves over the base by 7.06 points on DataWorld‑Bench, 14.28 on DAComp, 20.90 on DABstep‑Hard, and 12.13 on JobBench, a broader occupational benchmark.
Review