Meituan Trained Qwen on 1,037 Invented Worlds, Not Real Data

A team led by Meituan’s Longcat Interaction Team, with Nanjing University and other Chinese universities, published a paper on 15 October 2026 about Data World, an engine that invents the hidden rules of a world first and then generates the messy files an analyst would receive. They used 2,135 training tasks from 1,037 of these invented worlds to train an open Qwen model, which the paper reports improved by 20.90 points on the DABstep-Hard benchmark.

The trained model, which the authors call DataWorld-35B-A3B, was built from Qwen3.6-35B-A3B with reinforcement learning on its own attempts, without copying a stronger model’s answers, according to the paper.

Business Pill · SYNTHETIC DATA

A one-minute explainer of synthetic data. It teaches the general idea only and says nothing about Meituan or Data World.

How Data World Builds a Training Ground

The paper says the motivation is a supply problem: training data agents “faces costly environment curation, privacy and intellectual-property restrictions, and difficulty verifying open-ended findings without known generative truth.”

Its answer, in the authors’ words: “Our core insight is to specify the world behind the files before generating the files themselves.” Data World is described as “a multi-agent engine that synthesizes environments, tasks, and rubrics from explicit mechanisms.”

Because the engine specifies the mechanisms before it generates the files, the generative record shows which findings the files should support, according to the paper. The paper says independent audits check that each target can be recovered from the generated files, and the rubrics serve both to grade models and as rewards for reinforcement learning.

The engine produced two sets, according to the paper: a benchmark, DataWorld-Bench, with 100 tasks from 50 held-out worlds across 16 domains, and a training set of 2,135 tasks from 1,037 separate worlds.

What Was Posted on X

Alex Veremeyenko summarised the Data World paper on X on 20 October 2026, writing that the trained model was 81% better on hard financial tasks. That is the 20.90-point DABstep-Hard gain over a base of 25.66 in the paper’s Table 3; the paper reports the gain in points. The post asks how much private data AI companies need if invented worlds can train a better analyst.

What the Training Changed

The authors trained Qwen3.6-35B-A3B in what they call an RL-zero setting. The paper says optimization “uses only the training model’s own rollouts and rubric rewards”, avoiding supervised fine-tuning on the tasks or distillation from a stronger model’s trajectories.

Against its own base model, the paper reports DataWorld-35B-A3B scoring 46.56 on DABstep-Hard against 25.66, a gain of 20.90 points. It reports 68.00 against 53.72 on DAComp, 30.43 against 18.31 on JobBench and 44.70 against 37.64 on the held-out DataWorld-Bench.

By our arithmetic, the DABstep-Hard gain is about 81% over the base score, the figure Alex Veremeyenko gave in a post on X. The paper itself reports the gains in points.

The paper says JobBench covers 35 occupations, not just data analysis, and calls it “the least task-aligned of the four suites.” On general checks, it reports MMLU-Pro accuracy unchanged at 84.73%, LiveCodeBench-Hard down from 71.14% to 70.29% and SimpleQA-Verified up from 22.20 to 24.00.

Where Frontier Models Fall Short

The authors also ran ten frontier models through DataWorld-Bench. The paper reports what it calls “a consistent computation–insight gap”: each model scored higher on data computation than on insight discovery.

GPT-6-Astra had the top average, 74.43, with 87.80 on computation and 66.45 on insight, according to the paper’s table. Qwen-3.8-Max-0902 had the highest insight score, 67.95.

Connecting the Coverage

The paper speaks to a question we have raised before. In July, our piece on Cursor and SpaceX set out the thesis that harness traces plus frontier compute equals a credible path to model development, because real agent traces “do not exist on the public web.” Data World tests a different supply: invented environments in place of real ones.

It also touches the distillation debate. This morning we covered an IFP report on the US-China AI lead, published on 19 October, which treats distillation of US models as one of two levers in that gap. The Data World authors say their model was trained without trajectory distillation.

And it adds a training angle to our explainer this morning, Every AI Layer Is a Memory Store, which walks through how models store what they learn.

The Structural Read

As we read it, the training recipe is the notable part: the model learned only from its own attempts graded by rubrics, with no supervised fine-tuning on the tasks and no stronger model’s trajectories, according to the paper.

As we read it, the gains reach beyond data analysis: the paper reports a 12.13-point rise on JobBench, which it calls the least task-aligned of its four test suites.

As we read it, the frontier results point to where the hard part sits: every one of the ten models scored higher on computing numbers than on reaching the insight the task was built around.

Three Implications

INVENTED WORLDS, KNOWN ANSWERS The paper reports 2,135 training tasks from 1,037 worlds, each graded against a rubric built from that world’s own mechanisms.

GAINS MEASURED AGAINST THE BASE MODEL DABstep-Hard rose 20.90 points, DAComp 14.28, JobBench 12.13 and the held-out DataWorld-Bench 7.06, per the paper’s Table 3.

RELEASED, NOT YET REPLICATED The authors list the trained model, benchmark and training tasks on Hugging Face; the results are their own.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework describes “The Verifier Moat — built by RLVR on domain-specific reward signals, building into a new asset class for reasoning-heavy verticals.”

As we read it, Data World is a way to manufacture those reward signals for data analysis: its rubrics come from the world’s own mechanisms rather than from customer records. The framework was published in July 2026 and does not discuss this paper.

What Is Not Established

The paper lists its own limits. It says the cost of regenerating worlds limits exhaustive ablations, so “our results support the utility of the framework as a whole rather than isolate the contribution of each component.”

It says “fully synthetic tasks may differ from the questions human users pose to AI agents in practice”, that its rubrics “do not exhaust all valid analytical findings”, and that “LLM-based grading may introduce judge bias.” It adds that synthesized environments are meant “to complement, not replace, real-world data” and that it does not compare which source is more effective.

The results are the authors’ own and the paper is a preprint, and we found no independent replication. The authors list code on GitHub and, on Hugging Face, the trained model, the benchmark and the training tasks, the model and benchmark tagged with an Apache 2.0 license; we have not run them.

The Bottom Line

A Meituan-led team reports that an open 35B-parameter model trained only on tasks from invented worlds gained 20.90 points on DABstep-Hard and kept its general-knowledge score. The authors present invented worlds as a complement to scarce real data, and they flag that synthetic tasks may not match the questions real users ask.