What happened
- A team at Longcat, from Meituan, published Data World on arXiv on October 15, 2026, an engine that builds training environments for agents that analyze data with code.
- It works the opposite way from the usual: it first specifies the underlying world and then generates the files, tasks and grading rubrics from explicit mechanisms. Independent audits verify that the answer can be recovered from the files.
- With that engine, 2,135 training tasks were produced across 1,037 distinct worlds. A human selection left 100 tasks across 50 held-out worlds and 16 domains for DataWorld-Bench.
- Using reinforcement learning from scratch, without distilling trajectories from other models, they trained DataWorld-35B-A3B starting from Qwen3.6-35B-A3B.
Why it matters
- The training data does not come from any real company. The abstract cites the cost of curating environments and the limits of privacy and intellectual property as problems of the traditional method. For anyone who holds sensitive data and cannot share it, training on invented worlds is a way out that did not exist in this form.
- The improvement points over the base model are: +7.06 on DataWorld-Bench, +14.28 on DAComp, +20.90 on DABstep-Hard and +12.13 on JobBench. Three of the four benchmarks are not the team’s own, which helps; the fourth is.
- They evaluated ten frontier models and all show a gap between computing and interpreting: they get the calculation right and fail when drawing the conclusion. A Chilean team that asks an agent to “analyze sales” will get the table before the reading.
The number
+20.90 points: the improvement of DataWorld-35B-A3B over its base model on DABstep-Hard, the largest jump among the four tests reported.
Context
We read the paper’s abstract, not the full text. Other recent benchmarks point to the same weak spot, judgment rather than calculation: Taste-Bench measures the judgment of AI agents and email agents fail despite executing well.
What’s next
- In the abstract, the authors do not report a release date for the model or the data.
- No timelines announced.
Bottom line
Training on invented worlds avoids exposing real data. What remains to be seen is whether what is learned in an invented world holds up when the real files come in dirty.
Sources
- Data World: Scaling Data Environments from Scratch, arXiv 2610.21656, October 15, 2026.
Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.
