NederlandsMade-up company data with a known answer teaches AI agents to analyse better
Researchers from Meituan and Chinese universities build fake worlds full of messy files, and they know exactly what lies behind them. That makes it possible to check whether an AI agent really draws the right conclusion. A model that practised with them became clearly better at data analysis.
More and more companies let AI agents analyse their data. Such an agent is an AI that goes to work by itself: it opens spreadsheets and other files, writes small programs to calculate with them and draws conclusions from them. For example: why have orders been arriving late more often recently?
To train such an agent well, you need a lot of practice material. Real company data is difficult for that. It is expensive to collect, often confidential, and for an open question nobody knows for sure what the right answer is. Then you also cannot check whether the AI is doing well, or only saying something that sounds plausible.
Researchers from the Chinese tech company Meituan and several Chinese universities have come up with a solution. Their system is called Data World. They published the work on 15 October 2026 on arXiv, a website where researchers share their papers before other experts have reviewed them.
First the world, then the files
The idea is to do it the other way round. Data World first invents the rules of a small world, for example how suppliers, warehouse stock and deliveries affect each other. Only then does it make the files that such a company would have: order lists, stock overviews and logs.
Those files are deliberately messy, just like in real life. Values are missing, there are duplicate lines, data arrives late and numbers do not always match up. A separate check then looks at whether the right answer can still be extracted from the files.
Compare it to a teacher who makes a puzzle. Because the teacher came up with the solution, they can check exactly whether a pupil has solved it correctly. In the same way, Data World can also make questions and a marking scheme that fit what really happens in the fake world.
Calculating goes well, explaining is harder
With the system, the researchers made a test of 100 tasks in 50 different worlds, spread across 16 fields. They used it to test ten of the best AI models of the moment. The best model, GPT-6-Astra from OpenAI, scored on average about 74 out of 100 points.
Strikingly, all models were better at calculating with the data than at finding the explanation behind it. So an AI can often neatly calculate that more orders are late, but has more trouble with the question why. That very question is the most interesting one for a company.
Practice helps
The researchers also made more than 2,100 practice tasks in more than 1,000 other worlds. They used them to train an open language model from the Chinese company Alibaba, by letting it practise and rewarding it for good answers. So the model was not given example solutions from stronger models.
According to the researchers, the trained model went from about 38 to 45 out of 100 points on their own test. That brings it closer to GPT-6-Astra. On three other existing tests for data analysis and work tasks, which also go up to 100 points, it improved by 12 to 21 points. On one of those tests, for example, the score rose from about 26 to 47 points. So what it learned in the fake worlds also helps with other tasks.
The researchers are mainly interested in the practice method, not in catching up with the largest models. The trained model is open, though: anyone can download it. Companies can run such a model on their own computers, so that confidential data stays inside.
These are first results, measured by the researchers themselves. The next step is to show how well it works with real company data. The code is publicly available on GitHub, so that others can use and test the method themselves.
What does this mean for the road to the singularity?
An AI only becomes really useful when you can trust that its conclusions are right. That applies to company figures, but also to science, where AI agents are already helping, for example in the search for new antibiotics. Data World shows how to build that trust: practise in a world whose answer you know, and only then let it loose on the real world.
The fact that the computer makes the practice material itself is an important step. The shortage of good, checkable training data is one of the limits of AI. If AI can build endless new practice worlds by itself, agents can keep learning without waiting for real data. In this way they grow step by step from a handy calculating aid into a colleague who also finds the explanation behind the numbers.
What is the technological singularity? Read the explanation →
