Data World: Scaling Data Environments from Scratch

Data World. Scaling Data Environments from Scratch. By Zhanli Li and eighteen others, from Nanjing University, Meituan, and six other institutions. Published on the fifteenth of October. Its code repository is worked out: about forty source files, with a readme.

Data agents use code to process heterogeneous files and derive analytical insights from them. Scaling their training runs into three obstacles. Environments are costly to curate. Privacy and intellectual-property restrictions limit what can be used. And open-ended findings are hard to verify when there is no known generative truth behind the data. An agent can report an insight, but without knowing what produced the files, there is nothing to check that insight against.

The core insight is to specify the world behind the files before generating the files themselves. Data World is a multi-agent engine that synthesizes environments, tasks and rubrics from explicit mechanisms. Because the mechanism is written down first, the truth a task asks about is known by construction.

Independent audits check that each target is actually recoverable from the observed files. Rubrics support evaluation, and also serve as RL rewards. For a benchmark, DataWorld-Bench, human preference-based curation for question realism selects one hundred tasks across fifty held-out worlds and sixteen domains. For training, the engine yields two thousand one hundred and thirty-five tasks, drawn from one thousand and thirty-seven disjoint worlds. The benchmark worlds are held out, so what is evaluated is kept apart from what is trained on.

An evaluation of ten frontier models shows that the benchmark remains challenging, particularly for insight discovery, with a consistent computation-insight gap. RL-zero training on the synthesized tasks then produces DataWorld-35B-A3B from Qwen3.6-35B-A3B, without trajectory distillation.

Under matched evaluation, the trained model improves over its base by seven point zero six points on DataWorld-Bench, fourteen point two eight on DAComp, twenty point nine on DABstep-Hard, and twelve point one three on JobBench, a broader occupational benchmark.

Generating files first and asking what they mean afterwards leaves the right answer uncertain. Data World runs in the other direction, from mechanism to files, and that ordering is what makes its tasks checkable and its rewards usable.