Meituan Longcat opens Data World synthesis engine for data agents

Meituan's Longcat Interaction Team has published Data World, a multi-agent engine that specifies the mechanism behind a workflow before generating any files, so training rubrics align with analytical targets rather than whatever a curator or mining model happened to notice. The paper, posted to Hugging Face on October 16 with collaborators from Nanjing University, Zhejiang, Shanghai Jiao Tong, the Chinese University of Hong Kong, the Chinese Academy of Sciences' Institute of Automation and Zhongnan University of Economics and Law, puts the motivating problem bluntly: "ten experts or strong models could offer ten different explanations for the same decline, as data complexity and inadequately controlled confounders can leave the underlying mechanism ambiguous."

From the engine the authors release DataWorld-Bench, with 100 tasks across 50 held-out worlds and 16 domains, plus a world-disjoint RL training set of 2,135 tasks drawn from 1,037 worlds. "Evaluation of ten frontier models shows that the benchmark remains challenging, particularly for insight discovery, with a consistent computation–insight gap," the abstract reports.

RL-zero training on the synthesized tasks, "without trajectory distillation" from stronger models, produces DataWorld-35B-A3B from a Qwen3.6-35B-A3B base. Under matched evaluation it "improves over its base by 7.06 points on DataWorld-Bench, 14.28 on DAComp, 20.90 on DABstep-Hard, and 12.13 on JobBench," the last described as a broader occupational benchmark. The paper is the thirteenth synthetic-data story in our tracker over the past 90 days.