The Pulse
Meituan’s LongCat Publishes a Framework for Training Data Agents
Meituan’s LongCat Interaction Team published Data World, an open framework for creating synthetic data-analysis environments, tasks and grading rubrics for AI agents. The paper reports 2,135 training tasks and gains for a model trained on t
AI.info Team · Written with AI ·
“Our core insight is to specify the world behind the files before generating the files themselves.”
Zhanli Li et al., Longcat Interaction Team, Meituan
Meituan’s LongCat Interaction Team published Data World on October 15, pairing an open-source framework with a paper on generating training environments for data-analysis agents. The system creates interconnected files, analytical tasks and grading rubrics from a specified model of how the data relates. Its central bet is that building in a known account of how a synthetic world works makes it easier to check whether an agent’s conclusions follow from the files it can see.
Build the world before its files
Data World starts with a dependency graph describing entities and relationships, then generates an internal, canonical account of that world. It projects the account into files for an agent to inspect, adding defined imperfections such as missing values, delayed records, duplicate entries, inconsistent identifiers and changes in units or schema. A separate auditor checks whether the relationships and analytical targets specified by the system can still be recovered from those imperfect observations.
Tasks follow either a cause or consequence route through the graph: an agent can investigate what factors explain a change, or trace how that change affects downstream outcomes. Rubrics break expected answers into individual criteria, including checks for incorrect calculations or conclusions contradicted by the generated world. The solver sees the data files, not the canonical state used to create and audit them.
1,037 training worlds and 100 benchmark tasks
The team reports creating 2,135 training tasks across 1,037 worlds, separate from the 50 held-out worlds used for DataWorld-Bench. The benchmark contains 100 selected tasks spanning 16 domains. A group of 15 annotators compared candidate questions for realism, with disagreements adjudicated, according to the paper.
Ten models completed the benchmark, and every one scored lower on insight discovery than on data computation, by 4.87 to 21.35 points. GPT-6-Astra had the highest average score, 74.43 out of 100, while Qwen-3.8-Max-0902 led in insight discovery at 67.95. The authors caution that rubric coverage and scoring choices may contribute to the gap, so the comparison does not show that data processing ability alone causes the weaker insight scores.
Training gains extend beyond the synthetic benchmark
The researchers also used the generated tasks for reinforcement-learning training, starting from Qwen3.6-35B-A3B and producing DataWorld-35B-A3B. They report gains over the base model of 7.06 points on DataWorld-Bench, 14.28 on DAComp, 20.90 on DABstep-Hard and 12.13 on JobBench. The last benchmark covers 35 occupations, making it a test of transfer to file-based professional work beyond data analysis.
The training run used the model’s own sampled responses and rubric-based rewards, without task-solving demonstrations or trajectory distillation from a stronger model. On the paper’s separate retention checks, MMLU-Pro accuracy stayed at 84.73%; LiveCodeBench-Hard accuracy moved from 71.14% to 70.29%, while SimpleQA-Verified F1 rose from 22.20 to 24.00. Those figures are the authors’ matched evaluations, not a guarantee that the training method will produce the same results with other models or training setups.
Audits check synthetic truth, not real-world validity
The open GitHub repository contains the generation engine, task-synthesis tools, examples and documentation. Its quick-start instructions call for Python 3.9 or later, Node.js 18 or later, a Claude Code installation and a compatible model service; optional trajectory capture has additional requirements. The paper says the construction stages use GLM 5.3 through the Claude Code harness.
The method’s gains come with a defined boundary: the audit checks whether an agent can recover relationships that the synthetic world specifies, not whether its findings are suitable for real operational decisions. The authors also say their experiments assess the framework as a whole and do not isolate the effect of each component. Data World therefore offers a way to generate checkable training and evaluation material; its benchmark and audit results do not substitute for validation against the messy data and consequences of a real organization.